Spaces:
Build error
Build error
Upload 146 files
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- Dockerfile +41 -0
- LICENSE +1347 -0
- README.md +231 -12
- assets/yolo_arch.png +0 -0
- assets/yolo_logo.png +0 -0
- configs/finetune_coco/README.md +26 -0
- configs/finetune_coco/yolo_world_l_dual_vlpan_2e-4_80e_8gpus_finetune_coco.py +179 -0
- configs/finetune_coco/yolo_world_l_dual_vlpan_2e-4_80e_8gpus_mask-refine_finetune_coco.py +181 -0
- configs/finetune_coco/yolo_world_l_efficient_neck_2e-4_80e_8gpus_mask-refine_finetune_coco.py +159 -0
- configs/finetune_coco/yolo_world_v2_l_efficient_neck_2e-4_80e_8gpus_mask-refine_finetune_coco.py +182 -0
- configs/finetune_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py +181 -0
- configs/finetune_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_finetune_coco.py +160 -0
- configs/finetune_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_80e_8gpus_mask-refine_finetune_coco.py +161 -0
- configs/finetune_coco/yolo_world_v2_m_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py +182 -0
- configs/finetune_coco/yolo_world_v2_s_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py +184 -0
- configs/finetune_coco/yolo_world_v2_x_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py +183 -0
- configs/finetune_coco/yolo_world_v2_xl_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py +173 -0
- configs/pretrain/yolo_world_v2_l_clip_large_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_800ft_lvis_minival.py +200 -0
- configs/pretrain/yolo_world_v2_l_clip_large_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +171 -0
- configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py +202 -0
- configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +171 -0
- configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_val.py +171 -0
- configs/pretrain/yolo_world_v2_m_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py +198 -0
- configs/pretrain/yolo_world_v2_m_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +171 -0
- configs/pretrain/yolo_world_v2_m_vlpan_bn_noeinsum_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +176 -0
- configs/pretrain/yolo_world_v2_s_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py +195 -0
- configs/pretrain/yolo_world_v2_s_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +170 -0
- configs/pretrain/yolo_world_v2_x_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +171 -0
- configs/pretrain/yolo_world_v2_xl_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +185 -0
- configs/pretrain_v1/README.md +21 -0
- configs/pretrain_v1/yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +172 -0
- configs/pretrain_v1/yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_val.py +172 -0
- configs/pretrain_v1/yolo_world_m_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +172 -0
- configs/pretrain_v1/yolo_world_s_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +172 -0
- configs/pretrain_v1/yolo_world_x_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py +172 -0
- configs/prompt_tuning_coco/READEME.md +12 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_all_fine_tuning_coco.py +118 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_all_fine_tuning_rmdecay_rmmosaic_coco.py +114 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_all_fine_tuning_coco.py +156 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_fine_prompt_tuning_coco.py +156 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_prompt_tuning_coco.py +161 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_20e_8gpus_all_fine_tuning_rmdecay_coco.py +113 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_all_fine_tuning_rmdecay_coco.py +111 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_80e_8gpus_all_fine_tuning_coco.py +109 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_80e_8gpus_all_fine_tuning_rmdecay_coco.py +113 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_80e_8gpus_all_fine_tuning_rmdecay_coco_fixed.py +111 -0
- configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-4_80e_8gpus_all_fine_tuning_coco.py +109 -0
- configs/segmentation/README.md +27 -0
- configs/segmentation/yolo_world_seg_l_dual_vlpan_2e-4_80e_8gpus_allmodules_finetune_lvis.py +227 -0
- configs/segmentation/yolo_world_seg_l_dual_vlpan_2e-4_80e_8gpus_seghead_finetune_lvis.py +237 -0
Dockerfile
ADDED
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
FROM nvidia/cuda:11.8.0-devel-ubuntu22.04
|
2 |
+
|
3 |
+
ARG MODEL="yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py"
|
4 |
+
ARG WEIGHT="yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth"
|
5 |
+
|
6 |
+
ENV FORCE_CUDA="1"
|
7 |
+
ENV MMCV_WITH_OPS=1
|
8 |
+
|
9 |
+
RUN apt-get update && apt-get install -y --no-install-recommends \
|
10 |
+
python3-pip \
|
11 |
+
libgl1-mesa-glx \
|
12 |
+
libsm6 \
|
13 |
+
libxext6 \
|
14 |
+
libxrender-dev \
|
15 |
+
libglib2.0-0 \
|
16 |
+
git \
|
17 |
+
python3-dev \
|
18 |
+
python3-wheel
|
19 |
+
|
20 |
+
RUN pip3 install --upgrade pip \
|
21 |
+
&& pip3 install \
|
22 |
+
gradio \
|
23 |
+
opencv-python \
|
24 |
+
supervision \
|
25 |
+
mmengine \
|
26 |
+
setuptools \
|
27 |
+
&& pip3 install --no-cache-dir --index-url https://download.pytorch.org/whl/cu118 \
|
28 |
+
wheel \
|
29 |
+
torch \
|
30 |
+
torchvision \
|
31 |
+
torchaudio
|
32 |
+
|
33 |
+
COPY . /yolo
|
34 |
+
WORKDIR /yolo
|
35 |
+
|
36 |
+
RUN pip3 install -e .
|
37 |
+
|
38 |
+
RUN curl -o weights/$WEIGHT -L https://huggingface.co/wondervictor/YOLO-World/resolve/main/$WEIGHT
|
39 |
+
|
40 |
+
ENTRYPOINT [ "python3", "demo.py" ]
|
41 |
+
CMD ["configs/pretrain/$MODEL", "weights/$WEIGHT"]
|
LICENSE
ADDED
@@ -0,0 +1,1347 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
GNU GENERAL PUBLIC LICENSE
|
2 |
+
Version 3, 29 June 2007
|
3 |
+
|
4 |
+
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
|
5 |
+
Everyone is permitted to copy and distribute verbatim copies
|
6 |
+
of this license document, but changing it is not allowed.
|
7 |
+
|
8 |
+
Preamble
|
9 |
+
|
10 |
+
The GNU General Public License is a free, copyleft license for
|
11 |
+
software and other kinds of works.
|
12 |
+
|
13 |
+
The licenses for most software and other practical works are designed
|
14 |
+
to take away your freedom to share and change the works. By contrast,
|
15 |
+
the GNU General Public License is intended to guarantee your freedom to
|
16 |
+
share and change all versions of a program--to make sure it remains free
|
17 |
+
software for all its users. We, the Free Software Foundation, use the
|
18 |
+
GNU General Public License for most of our software; it applies also to
|
19 |
+
any other work released this way by its authors. You can apply it to
|
20 |
+
your programs, too.
|
21 |
+
|
22 |
+
When we speak of free software, we are referring to freedom, not
|
23 |
+
price. Our General Public Licenses are designed to make sure that you
|
24 |
+
have the freedom to distribute copies of free software (and charge for
|
25 |
+
them if you wish), that you receive source code or can get it if you
|
26 |
+
want it, that you can change the software or use pieces of it in new
|
27 |
+
free programs, and that you know you can do these things.
|
28 |
+
|
29 |
+
To protect your rights, we need to prevent others from denying you
|
30 |
+
these rights or asking you to surrender the rights. Therefore, you have
|
31 |
+
certain responsibilities if you distribute copies of the software, or if
|
32 |
+
you modify it: responsibilities to respect the freedom of others.
|
33 |
+
|
34 |
+
For example, if you distribute copies of such a program, whether
|
35 |
+
gratis or for a fee, you must pass on to the recipients the same
|
36 |
+
freedoms that you received. You must make sure that they, too, receive
|
37 |
+
or can get the source code. And you must show them these terms so they
|
38 |
+
know their rights.
|
39 |
+
|
40 |
+
Developers that use the GNU GPL protect your rights with two steps:
|
41 |
+
(1) assert copyright on the software, and (2) offer you this License
|
42 |
+
giving you legal permission to copy, distribute and/or modify it.
|
43 |
+
|
44 |
+
For the developers' and authors' protection, the GPL clearly explains
|
45 |
+
that there is no warranty for this free software. For both users' and
|
46 |
+
authors' sake, the GPL requires that modified versions be marked as
|
47 |
+
changed, so that their problems will not be attributed erroneously to
|
48 |
+
authors of previous versions.
|
49 |
+
|
50 |
+
Some devices are designed to deny users access to install or run
|
51 |
+
modified versions of the software inside them, although the manufacturer
|
52 |
+
can do so. This is fundamentally incompatible with the aim of
|
53 |
+
protecting users' freedom to change the software. The systematic
|
54 |
+
pattern of such abuse occurs in the area of products for individuals to
|
55 |
+
use, which is precisely where it is most unacceptable. Therefore, we
|
56 |
+
have designed this version of the GPL to prohibit the practice for those
|
57 |
+
products. If such problems arise substantially in other domains, we
|
58 |
+
stand ready to extend this provision to those domains in future versions
|
59 |
+
of the GPL, as needed to protect the freedom of users.
|
60 |
+
|
61 |
+
Finally, every program is threatened constantly by software patents.
|
62 |
+
States should not allow patents to restrict development and use of
|
63 |
+
software on general-purpose computers, but in those that do, we wish to
|
64 |
+
avoid the special danger that patents applied to a free program could
|
65 |
+
make it effectively proprietary. To prevent this, the GPL assures that
|
66 |
+
patents cannot be used to render the program non-free.
|
67 |
+
|
68 |
+
The precise terms and conditions for copying, distribution and
|
69 |
+
modification follow.
|
70 |
+
|
71 |
+
TERMS AND CONDITIONS
|
72 |
+
|
73 |
+
0. Definitions.
|
74 |
+
|
75 |
+
"This License" refers to version 3 of the GNU General Public License.
|
76 |
+
|
77 |
+
"Copyright" also means copyright-like laws that apply to other kinds of
|
78 |
+
works, such as semiconductor masks.
|
79 |
+
|
80 |
+
"The Program" refers to any copyrightable work licensed under this
|
81 |
+
License. Each licensee is addressed as "you". "Licensees" and
|
82 |
+
"recipients" may be individuals or organizations.
|
83 |
+
|
84 |
+
To "modify" a work means to copy from or adapt all or part of the work
|
85 |
+
in a fashion requiring copyright permission, other than the making of an
|
86 |
+
exact copy. The resulting work is called a "modified version" of the
|
87 |
+
earlier work or a work "based on" the earlier work.
|
88 |
+
|
89 |
+
A "covered work" means either the unmodified Program or a work based
|
90 |
+
on the Program.
|
91 |
+
|
92 |
+
To "propagate" a work means to do anything with it that, without
|
93 |
+
permission, would make you directly or secondarily liable for
|
94 |
+
infringement under applicable copyright law, except executing it on a
|
95 |
+
computer or modifying a private copy. Propagation includes copying,
|
96 |
+
distribution (with or without modification), making available to the
|
97 |
+
public, and in some countries other activities as well.
|
98 |
+
|
99 |
+
To "convey" a work means any kind of propagation that enables other
|
100 |
+
parties to make or receive copies. Mere interaction with a user through
|
101 |
+
a computer network, with no transfer of a copy, is not conveying.
|
102 |
+
|
103 |
+
An interactive user interface displays "Appropriate Legal Notices"
|
104 |
+
to the extent that it includes a convenient and prominently visible
|
105 |
+
feature that (1) displays an appropriate copyright notice, and (2)
|
106 |
+
tells the user that there is no warranty for the work (except to the
|
107 |
+
extent that warranties are provided), that licensees may convey the
|
108 |
+
work under this License, and how to view a copy of this License. If
|
109 |
+
the interface presents a list of user commands or options, such as a
|
110 |
+
menu, a prominent item in the list meets this criterion.
|
111 |
+
|
112 |
+
1. Source Code.
|
113 |
+
|
114 |
+
The "source code" for a work means the preferred form of the work
|
115 |
+
for making modifications to it. "Object code" means any non-source
|
116 |
+
form of a work.
|
117 |
+
|
118 |
+
A "Standard Interface" means an interface that either is an official
|
119 |
+
standard defined by a recognized standards body, or, in the case of
|
120 |
+
interfaces specified for a particular programming language, one that
|
121 |
+
is widely used among developers working in that language.
|
122 |
+
|
123 |
+
The "System Libraries" of an executable work include anything, other
|
124 |
+
than the work as a whole, that (a) is included in the normal form of
|
125 |
+
packaging a Major Component, but which is not part of that Major
|
126 |
+
Component, and (b) serves only to enable use of the work with that
|
127 |
+
Major Component, or to implement a Standard Interface for which an
|
128 |
+
implementation is available to the public in source code form. A
|
129 |
+
"Major Component", in this context, means a major essential component
|
130 |
+
(kernel, window system, and so on) of the specific operating system
|
131 |
+
(if any) on which the executable work runs, or a compiler used to
|
132 |
+
produce the work, or an object code interpreter used to run it.
|
133 |
+
|
134 |
+
The "Corresponding Source" for a work in object code form means all
|
135 |
+
the source code needed to generate, install, and (for an executable
|
136 |
+
work) run the object code and to modify the work, including scripts to
|
137 |
+
control those activities. However, it does not include the work's
|
138 |
+
System Libraries, or general-purpose tools or generally available free
|
139 |
+
programs which are used unmodified in performing those activities but
|
140 |
+
which are not part of the work. For example, Corresponding Source
|
141 |
+
includes interface definition files associated with source files for
|
142 |
+
the work, and the source code for shared libraries and dynamically
|
143 |
+
linked subprograms that the work is specifically designed to require,
|
144 |
+
such as by intimate data communication or control flow between those
|
145 |
+
subprograms and other parts of the work.
|
146 |
+
|
147 |
+
The Corresponding Source need not include anything that users
|
148 |
+
can regenerate automatically from other parts of the Corresponding
|
149 |
+
Source.
|
150 |
+
|
151 |
+
The Corresponding Source for a work in source code form is that
|
152 |
+
same work.
|
153 |
+
|
154 |
+
2. Basic Permissions.
|
155 |
+
|
156 |
+
All rights granted under this License are granted for the term of
|
157 |
+
copyright on the Program, and are irrevocable provided the stated
|
158 |
+
conditions are met. This License explicitly affirms your unlimited
|
159 |
+
permission to run the unmodified Program. The output from running a
|
160 |
+
covered work is covered by this License only if the output, given its
|
161 |
+
content, constitutes a covered work. This License acknowledges your
|
162 |
+
rights of fair use or other equivalent, as provided by copyright law.
|
163 |
+
|
164 |
+
You may make, run and propagate covered works that you do not
|
165 |
+
convey, without conditions so long as your license otherwise remains
|
166 |
+
in force. You may convey covered works to others for the sole purpose
|
167 |
+
of having them make modifications exclusively for you, or provide you
|
168 |
+
with facilities for running those works, provided that you comply with
|
169 |
+
the terms of this License in conveying all material for which you do
|
170 |
+
not control copyright. Those thus making or running the covered works
|
171 |
+
for you must do so exclusively on your behalf, under your direction
|
172 |
+
and control, on terms that prohibit them from making any copies of
|
173 |
+
your copyrighted material outside their relationship with you.
|
174 |
+
|
175 |
+
Conveying under any other circumstances is permitted solely under
|
176 |
+
the conditions stated below. Sublicensing is not allowed; section 10
|
177 |
+
makes it unnecessary.
|
178 |
+
|
179 |
+
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
|
180 |
+
|
181 |
+
No covered work shall be deemed part of an effective technological
|
182 |
+
measure under any applicable law fulfilling obligations under article
|
183 |
+
11 of the WIPO copyright treaty adopted on 20 December 1996, or
|
184 |
+
similar laws prohibiting or restricting circumvention of such
|
185 |
+
measures.
|
186 |
+
|
187 |
+
When you convey a covered work, you waive any legal power to forbid
|
188 |
+
circumvention of technological measures to the extent such circumvention
|
189 |
+
is effected by exercising rights under this License with respect to
|
190 |
+
the covered work, and you disclaim any intention to limit operation or
|
191 |
+
modification of the work as a means of enforcing, against the work's
|
192 |
+
users, your or third parties' legal rights to forbid circumvention of
|
193 |
+
technological measures.
|
194 |
+
|
195 |
+
4. Conveying Verbatim Copies.
|
196 |
+
|
197 |
+
You may convey verbatim copies of the Program's source code as you
|
198 |
+
receive it, in any medium, provided that you conspicuously and
|
199 |
+
appropriately publish on each copy an appropriate copyright notice;
|
200 |
+
keep intact all notices stating that this License and any
|
201 |
+
non-permissive terms added in accord with section 7 apply to the code;
|
202 |
+
keep intact all notices of the absence of any warranty; and give all
|
203 |
+
recipients a copy of this License along with the Program.
|
204 |
+
|
205 |
+
You may charge any price or no price for each copy that you convey,
|
206 |
+
and you may offer support or warranty protection for a fee.
|
207 |
+
|
208 |
+
5. Conveying Modified Source Versions.
|
209 |
+
|
210 |
+
You may convey a work based on the Program, or the modifications to
|
211 |
+
produce it from the Program, in the form of source code under the
|
212 |
+
terms of section 4, provided that you also meet all of these conditions:
|
213 |
+
|
214 |
+
a) The work must carry prominent notices stating that you modified
|
215 |
+
it, and giving a relevant date.
|
216 |
+
|
217 |
+
b) The work must carry prominent notices stating that it is
|
218 |
+
released under this License and any conditions added under section
|
219 |
+
7. This requirement modifies the requirement in section 4 to
|
220 |
+
"keep intact all notices".
|
221 |
+
|
222 |
+
c) You must license the entire work, as a whole, under this
|
223 |
+
License to anyone who comes into possession of a copy. This
|
224 |
+
License will therefore apply, along with any applicable section 7
|
225 |
+
additional terms, to the whole of the work, and all its parts,
|
226 |
+
regardless of how they are packaged. This License gives no
|
227 |
+
permission to license the work in any other way, but it does not
|
228 |
+
invalidate such permission if you have separately received it.
|
229 |
+
|
230 |
+
d) If the work has interactive user interfaces, each must display
|
231 |
+
Appropriate Legal Notices; however, if the Program has interactive
|
232 |
+
interfaces that do not display Appropriate Legal Notices, your
|
233 |
+
work need not make them do so.
|
234 |
+
|
235 |
+
A compilation of a covered work with other separate and independent
|
236 |
+
works, which are not by their nature extensions of the covered work,
|
237 |
+
and which are not combined with it such as to form a larger program,
|
238 |
+
in or on a volume of a storage or distribution medium, is called an
|
239 |
+
"aggregate" if the compilation and its resulting copyright are not
|
240 |
+
used to limit the access or legal rights of the compilation's users
|
241 |
+
beyond what the individual works permit. Inclusion of a covered work
|
242 |
+
in an aggregate does not cause this License to apply to the other
|
243 |
+
parts of the aggregate.
|
244 |
+
|
245 |
+
6. Conveying Non-Source Forms.
|
246 |
+
|
247 |
+
You may convey a covered work in object code form under the terms
|
248 |
+
of sections 4 and 5, provided that you also convey the
|
249 |
+
machine-readable Corresponding Source under the terms of this License,
|
250 |
+
in one of these ways:
|
251 |
+
|
252 |
+
a) Convey the object code in, or embodied in, a physical product
|
253 |
+
(including a physical distribution medium), accompanied by the
|
254 |
+
Corresponding Source fixed on a durable physical medium
|
255 |
+
customarily used for software interchange.
|
256 |
+
|
257 |
+
b) Convey the object code in, or embodied in, a physical product
|
258 |
+
(including a physical distribution medium), accompanied by a
|
259 |
+
written offer, valid for at least three years and valid for as
|
260 |
+
long as you offer spare parts or customer support for that product
|
261 |
+
model, to give anyone who possesses the object code either (1) a
|
262 |
+
copy of the Corresponding Source for all the software in the
|
263 |
+
product that is covered by this License, on a durable physical
|
264 |
+
medium customarily used for software interchange, for a price no
|
265 |
+
more than your reasonable cost of physically performing this
|
266 |
+
conveying of source, or (2) access to copy the
|
267 |
+
Corresponding Source from a network server at no charge.
|
268 |
+
|
269 |
+
c) Convey individual copies of the object code with a copy of the
|
270 |
+
written offer to provide the Corresponding Source. This
|
271 |
+
alternative is allowed only occasionally and noncommercially, and
|
272 |
+
only if you received the object code with such an offer, in accord
|
273 |
+
with subsection 6b.
|
274 |
+
|
275 |
+
d) Convey the object code by offering access from a designated
|
276 |
+
place (gratis or for a charge), and offer equivalent access to the
|
277 |
+
Corresponding Source in the same way through the same place at no
|
278 |
+
further charge. You need not require recipients to copy the
|
279 |
+
Corresponding Source along with the object code. If the place to
|
280 |
+
copy the object code is a network server, the Corresponding Source
|
281 |
+
may be on a different server (operated by you or a third party)
|
282 |
+
that supports equivalent copying facilities, provided you maintain
|
283 |
+
clear directions next to the object code saying where to find the
|
284 |
+
Corresponding Source. Regardless of what server hosts the
|
285 |
+
Corresponding Source, you remain obligated to ensure that it is
|
286 |
+
available for as long as needed to satisfy these requirements.
|
287 |
+
|
288 |
+
e) Convey the object code using peer-to-peer transmission, provided
|
289 |
+
you inform other peers where the object code and Corresponding
|
290 |
+
Source of the work are being offered to the general public at no
|
291 |
+
charge under subsection 6d.
|
292 |
+
|
293 |
+
A separable portion of the object code, whose source code is excluded
|
294 |
+
from the Corresponding Source as a System Library, need not be
|
295 |
+
included in conveying the object code work.
|
296 |
+
|
297 |
+
A "User Product" is either (1) a "consumer product", which means any
|
298 |
+
tangible personal property which is normally used for personal, family,
|
299 |
+
or household purposes, or (2) anything designed or sold for incorporation
|
300 |
+
into a dwelling. In determining whether a product is a consumer product,
|
301 |
+
doubtful cases shall be resolved in favor of coverage. For a particular
|
302 |
+
product received by a particular user, "normally used" refers to a
|
303 |
+
typical or common use of that class of product, regardless of the status
|
304 |
+
of the particular user or of the way in which the particular user
|
305 |
+
actually uses, or expects or is expected to use, the product. A product
|
306 |
+
is a consumer product regardless of whether the product has substantial
|
307 |
+
commercial, industrial or non-consumer uses, unless such uses represent
|
308 |
+
the only significant mode of use of the product.
|
309 |
+
|
310 |
+
"Installation Information" for a User Product means any methods,
|
311 |
+
procedures, authorization keys, or other information required to install
|
312 |
+
and execute modified versions of a covered work in that User Product from
|
313 |
+
a modified version of its Corresponding Source. The information must
|
314 |
+
suffice to ensure that the continued functioning of the modified object
|
315 |
+
code is in no case prevented or interfered with solely because
|
316 |
+
modification has been made.
|
317 |
+
|
318 |
+
If you convey an object code work under this section in, or with, or
|
319 |
+
specifically for use in, a User Product, and the conveying occurs as
|
320 |
+
part of a transaction in which the right of possession and use of the
|
321 |
+
User Product is transferred to the recipient in perpetuity or for a
|
322 |
+
fixed term (regardless of how the transaction is characterized), the
|
323 |
+
Corresponding Source conveyed under this section must be accompanied
|
324 |
+
by the Installation Information. But this requirement does not apply
|
325 |
+
if neither you nor any third party retains the ability to install
|
326 |
+
modified object code on the User Product (for example, the work has
|
327 |
+
been installed in ROM).
|
328 |
+
|
329 |
+
The requirement to provide Installation Information does not include a
|
330 |
+
requirement to continue to provide support service, warranty, or updates
|
331 |
+
for a work that has been modified or installed by the recipient, or for
|
332 |
+
the User Product in which it has been modified or installed. Access to a
|
333 |
+
network may be denied when the modification itself materially and
|
334 |
+
adversely affects the operation of the network or violates the rules and
|
335 |
+
protocols for communication across the network.
|
336 |
+
|
337 |
+
Corresponding Source conveyed, and Installation Information provided,
|
338 |
+
in accord with this section must be in a format that is publicly
|
339 |
+
documented (and with an implementation available to the public in
|
340 |
+
source code form), and must require no special password or key for
|
341 |
+
unpacking, reading or copying.
|
342 |
+
|
343 |
+
7. Additional Terms.
|
344 |
+
|
345 |
+
"Additional permissions" are terms that supplement the terms of this
|
346 |
+
License by making exceptions from one or more of its conditions.
|
347 |
+
Additional permissions that are applicable to the entire Program shall
|
348 |
+
be treated as though they were included in this License, to the extent
|
349 |
+
that they are valid under applicable law. If additional permissions
|
350 |
+
apply only to part of the Program, that part may be used separately
|
351 |
+
under those permissions, but the entire Program remains governed by
|
352 |
+
this License without regard to the additional permissions.
|
353 |
+
|
354 |
+
When you convey a copy of a covered work, you may at your option
|
355 |
+
remove any additional permissions from that copy, or from any part of
|
356 |
+
it. (Additional permissions may be written to require their own
|
357 |
+
removal in certain cases when you modify the work.) You may place
|
358 |
+
additional permissions on material, added by you to a covered work,
|
359 |
+
for which you have or can give appropriate copyright permission.
|
360 |
+
|
361 |
+
Notwithstanding any other provision of this License, for material you
|
362 |
+
add to a covered work, you may (if authorized by the copyright holders of
|
363 |
+
that material) supplement the terms of this License with terms:
|
364 |
+
|
365 |
+
a) Disclaiming warranty or limiting liability differently from the
|
366 |
+
terms of sections 15 and 16 of this License; or
|
367 |
+
|
368 |
+
b) Requiring preservation of specified reasonable legal notices or
|
369 |
+
author attributions in that material or in the Appropriate Legal
|
370 |
+
Notices displayed by works containing it; or
|
371 |
+
|
372 |
+
c) Prohibiting misrepresentation of the origin of that material, or
|
373 |
+
requiring that modified versions of such material be marked in
|
374 |
+
reasonable ways as different from the original version; or
|
375 |
+
|
376 |
+
d) Limiting the use for publicity purposes of names of licensors or
|
377 |
+
authors of the material; or
|
378 |
+
|
379 |
+
e) Declining to grant rights under trademark law for use of some
|
380 |
+
trade names, trademarks, or service marks; or
|
381 |
+
|
382 |
+
f) Requiring indemnification of licensors and authors of that
|
383 |
+
material by anyone who conveys the material (or modified versions of
|
384 |
+
it) with contractual assumptions of liability to the recipient, for
|
385 |
+
any liability that these contractual assumptions directly impose on
|
386 |
+
those licensors and authors.
|
387 |
+
|
388 |
+
All other non-permissive additional terms are considered "further
|
389 |
+
restrictions" within the meaning of section 10. If the Program as you
|
390 |
+
received it, or any part of it, contains a notice stating that it is
|
391 |
+
governed by this License along with a term that is a further
|
392 |
+
restriction, you may remove that term. If a license document contains
|
393 |
+
a further restriction but permits relicensing or conveying under this
|
394 |
+
License, you may add to a covered work material governed by the terms
|
395 |
+
of that license document, provided that the further restriction does
|
396 |
+
not survive such relicensing or conveying.
|
397 |
+
|
398 |
+
If you add terms to a covered work in accord with this section, you
|
399 |
+
must place, in the relevant source files, a statement of the
|
400 |
+
additional terms that apply to those files, or a notice indicating
|
401 |
+
where to find the applicable terms.
|
402 |
+
|
403 |
+
Additional terms, permissive or non-permissive, may be stated in the
|
404 |
+
form of a separately written license, or stated as exceptions;
|
405 |
+
the above requirements apply either way.
|
406 |
+
|
407 |
+
8. Termination.
|
408 |
+
|
409 |
+
You may not propagate or modify a covered work except as expressly
|
410 |
+
provided under this License. Any attempt otherwise to propagate or
|
411 |
+
modify it is void, and will automatically terminate your rights under
|
412 |
+
this License (including any patent licenses granted under the third
|
413 |
+
paragraph of section 11).
|
414 |
+
|
415 |
+
However, if you cease all violation of this License, then your
|
416 |
+
license from a particular copyright holder is reinstated (a)
|
417 |
+
provisionally, unless and until the copyright holder explicitly and
|
418 |
+
finally terminates your license, and (b) permanently, if the copyright
|
419 |
+
holder fails to notify you of the violation by some reasonable means
|
420 |
+
prior to 60 days after the cessation.
|
421 |
+
|
422 |
+
Moreover, your license from a particular copyright holder is
|
423 |
+
reinstated permanently if the copyright holder notifies you of the
|
424 |
+
violation by some reasonable means, this is the first time you have
|
425 |
+
received notice of violation of this License (for any work) from that
|
426 |
+
copyright holder, and you cure the violation prior to 30 days after
|
427 |
+
your receipt of the notice.
|
428 |
+
|
429 |
+
Termination of your rights under this section does not terminate the
|
430 |
+
licenses of parties who have received copies or rights from you under
|
431 |
+
this License. If your rights have been terminated and not permanently
|
432 |
+
reinstated, you do not qualify to receive new licenses for the same
|
433 |
+
material under section 10.
|
434 |
+
|
435 |
+
9. Acceptance Not Required for Having Copies.
|
436 |
+
|
437 |
+
You are not required to accept this License in order to receive or
|
438 |
+
run a copy of the Program. Ancillary propagation of a covered work
|
439 |
+
occurring solely as a consequence of using peer-to-peer transmission
|
440 |
+
to receive a copy likewise does not require acceptance. However,
|
441 |
+
nothing other than this License grants you permission to propagate or
|
442 |
+
modify any covered work. These actions infringe copyright if you do
|
443 |
+
not accept this License. Therefore, by modifying or propagating a
|
444 |
+
covered work, you indicate your acceptance of this License to do so.
|
445 |
+
|
446 |
+
10. Automatic Licensing of Downstream Recipients.
|
447 |
+
|
448 |
+
Each time you convey a covered work, the recipient automatically
|
449 |
+
receives a license from the original licensors, to run, modify and
|
450 |
+
propagate that work, subject to this License. You are not responsible
|
451 |
+
for enforcing compliance by third parties with this License.
|
452 |
+
|
453 |
+
An "entity transaction" is a transaction transferring control of an
|
454 |
+
organization, or substantially all assets of one, or subdividing an
|
455 |
+
organization, or merging organizations. If propagation of a covered
|
456 |
+
work results from an entity transaction, each party to that
|
457 |
+
transaction who receives a copy of the work also receives whatever
|
458 |
+
licenses to the work the party's predecessor in interest had or could
|
459 |
+
give under the previous paragraph, plus a right to possession of the
|
460 |
+
Corresponding Source of the work from the predecessor in interest, if
|
461 |
+
the predecessor has it or can get it with reasonable efforts.
|
462 |
+
|
463 |
+
You may not impose any further restrictions on the exercise of the
|
464 |
+
rights granted or affirmed under this License. For example, you may
|
465 |
+
not impose a license fee, royalty, or other charge for exercise of
|
466 |
+
rights granted under this License, and you may not initiate litigation
|
467 |
+
(including a cross-claim or counterclaim in a lawsuit) alleging that
|
468 |
+
any patent claim is infringed by making, using, selling, offering for
|
469 |
+
sale, or importing the Program or any portion of it.
|
470 |
+
|
471 |
+
11. Patents.
|
472 |
+
|
473 |
+
A "contributor" is a copyright holder who authorizes use under this
|
474 |
+
License of the Program or a work on which the Program is based. The
|
475 |
+
work thus licensed is called the contributor's "contributor version".
|
476 |
+
|
477 |
+
A contributor's "essential patent claims" are all patent claims
|
478 |
+
owned or controlled by the contributor, whether already acquired or
|
479 |
+
hereafter acquired, that would be infringed by some manner, permitted
|
480 |
+
by this License, of making, using, or selling its contributor version,
|
481 |
+
but do not include claims that would be infringed only as a
|
482 |
+
consequence of further modification of the contributor version. For
|
483 |
+
purposes of this definition, "control" includes the right to grant
|
484 |
+
patent sublicenses in a manner consistent with the requirements of
|
485 |
+
this License.
|
486 |
+
|
487 |
+
Each contributor grants you a non-exclusive, worldwide, royalty-free
|
488 |
+
patent license under the contributor's essential patent claims, to
|
489 |
+
make, use, sell, offer for sale, import and otherwise run, modify and
|
490 |
+
propagate the contents of its contributor version.
|
491 |
+
|
492 |
+
In the following three paragraphs, a "patent license" is any express
|
493 |
+
agreement or commitment, however denominated, not to enforce a patent
|
494 |
+
(such as an express permission to practice a patent or covenant not to
|
495 |
+
sue for patent infringement). To "grant" such a patent license to a
|
496 |
+
party means to make such an agreement or commitment not to enforce a
|
497 |
+
patent against the party.
|
498 |
+
|
499 |
+
If you convey a covered work, knowingly relying on a patent license,
|
500 |
+
and the Corresponding Source of the work is not available for anyone
|
501 |
+
to copy, free of charge and under the terms of this License, through a
|
502 |
+
publicly available network server or other readily accessible means,
|
503 |
+
then you must either (1) cause the Corresponding Source to be so
|
504 |
+
available, or (2) arrange to deprive yourself of the benefit of the
|
505 |
+
patent license for this particular work, or (3) arrange, in a manner
|
506 |
+
consistent with the requirements of this License, to extend the patent
|
507 |
+
license to downstream recipients. "Knowingly relying" means you have
|
508 |
+
actual knowledge that, but for the patent license, your conveying the
|
509 |
+
covered work in a country, or your recipient's use of the covered work
|
510 |
+
in a country, would infringe one or more identifiable patents in that
|
511 |
+
country that you have reason to believe are valid.
|
512 |
+
|
513 |
+
If, pursuant to or in connection with a single transaction or
|
514 |
+
arrangement, you convey, or propagate by procuring conveyance of, a
|
515 |
+
covered work, and grant a patent license to some of the parties
|
516 |
+
receiving the covered work authorizing them to use, propagate, modify
|
517 |
+
or convey a specific copy of the covered work, then the patent license
|
518 |
+
you grant is automatically extended to all recipients of the covered
|
519 |
+
work and works based on it.
|
520 |
+
|
521 |
+
A patent license is "discriminatory" if it does not include within
|
522 |
+
the scope of its coverage, prohibits the exercise of, or is
|
523 |
+
conditioned on the non-exercise of one or more of the rights that are
|
524 |
+
specifically granted under this License. You may not convey a covered
|
525 |
+
work if you are a party to an arrangement with a third party that is
|
526 |
+
in the business of distributing software, under which you make payment
|
527 |
+
to the third party based on the extent of your activity of conveying
|
528 |
+
the work, and under which the third party grants, to any of the
|
529 |
+
parties who would receive the covered work from you, a discriminatory
|
530 |
+
patent license (a) in connection with copies of the covered work
|
531 |
+
conveyed by you (or copies made from those copies), or (b) primarily
|
532 |
+
for and in connection with specific products or compilations that
|
533 |
+
contain the covered work, unless you entered into that arrangement,
|
534 |
+
or that patent license was granted, prior to 28 March 2007.
|
535 |
+
|
536 |
+
Nothing in this License shall be construed as excluding or limiting
|
537 |
+
any implied license or other defenses to infringement that may
|
538 |
+
otherwise be available to you under applicable patent law.
|
539 |
+
|
540 |
+
12. No Surrender of Others' Freedom.
|
541 |
+
|
542 |
+
If conditions are imposed on you (whether by court order, agreement or
|
543 |
+
otherwise) that contradict the conditions of this License, they do not
|
544 |
+
excuse you from the conditions of this License. If you cannot convey a
|
545 |
+
covered work so as to satisfy simultaneously your obligations under this
|
546 |
+
License and any other pertinent obligations, then as a consequence you may
|
547 |
+
not convey it at all. For example, if you agree to terms that obligate you
|
548 |
+
to collect a royalty for further conveying from those to whom you convey
|
549 |
+
the Program, the only way you could satisfy both those terms and this
|
550 |
+
License would be to refrain entirely from conveying the Program.
|
551 |
+
|
552 |
+
13. Use with the GNU Affero General Public License.
|
553 |
+
|
554 |
+
Notwithstanding any other provision of this License, you have
|
555 |
+
permission to link or combine any covered work with a work licensed
|
556 |
+
under version 3 of the GNU Affero General Public License into a single
|
557 |
+
combined work, and to convey the resulting work. The terms of this
|
558 |
+
License will continue to apply to the part which is the covered work,
|
559 |
+
but the special requirements of the GNU Affero General Public License,
|
560 |
+
section 13, concerning interaction through a network will apply to the
|
561 |
+
combination as such.
|
562 |
+
|
563 |
+
14. Revised Versions of this License.
|
564 |
+
|
565 |
+
The Free Software Foundation may publish revised and/or new versions of
|
566 |
+
the GNU General Public License from time to time. Such new versions will
|
567 |
+
be similar in spirit to the present version, but may differ in detail to
|
568 |
+
address new problems or concerns.
|
569 |
+
|
570 |
+
Each version is given a distinguishing version number. If the
|
571 |
+
Program specifies that a certain numbered version of the GNU General
|
572 |
+
Public License "or any later version" applies to it, you have the
|
573 |
+
option of following the terms and conditions either of that numbered
|
574 |
+
version or of any later version published by the Free Software
|
575 |
+
Foundation. If the Program does not specify a version number of the
|
576 |
+
GNU General Public License, you may choose any version ever published
|
577 |
+
by the Free Software Foundation.
|
578 |
+
|
579 |
+
If the Program specifies that a proxy can decide which future
|
580 |
+
versions of the GNU General Public License can be used, that proxy's
|
581 |
+
public statement of acceptance of a version permanently authorizes you
|
582 |
+
to choose that version for the Program.
|
583 |
+
|
584 |
+
Later license versions may give you additional or different
|
585 |
+
permissions. However, no additional obligations are imposed on any
|
586 |
+
author or copyright holder as a result of your choosing to follow a
|
587 |
+
later version.
|
588 |
+
|
589 |
+
15. Disclaimer of Warranty.
|
590 |
+
|
591 |
+
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
|
592 |
+
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
|
593 |
+
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
|
594 |
+
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
|
595 |
+
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
596 |
+
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
|
597 |
+
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
|
598 |
+
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
|
599 |
+
|
600 |
+
16. Limitation of Liability.
|
601 |
+
|
602 |
+
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
|
603 |
+
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
|
604 |
+
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
|
605 |
+
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
|
606 |
+
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
|
607 |
+
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
|
608 |
+
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
|
609 |
+
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
|
610 |
+
SUCH DAMAGES.
|
611 |
+
|
612 |
+
17. Interpretation of Sections 15 and 16.
|
613 |
+
|
614 |
+
If the disclaimer of warranty and limitation of liability provided
|
615 |
+
above cannot be given local legal effect according to their terms,
|
616 |
+
reviewing courts shall apply local law that most closely approximates
|
617 |
+
an absolute waiver of all civil liability in connection with the
|
618 |
+
Program, unless a warranty or assumption of liability accompanies a
|
619 |
+
copy of the Program in return for a fee.
|
620 |
+
|
621 |
+
END OF TERMS AND CONDITIONS
|
622 |
+
|
623 |
+
How to Apply These Terms to Your New Programs
|
624 |
+
|
625 |
+
If you develop a new program, and you want it to be of the greatest
|
626 |
+
possible use to the public, the best way to achieve this is to make it
|
627 |
+
free software which everyone can redistribute and change under these terms.
|
628 |
+
|
629 |
+
To do so, attach the following notices to the program. It is safest
|
630 |
+
to attach them to the start of each source file to most effectively
|
631 |
+
state the exclusion of warranty; and each file should have at least
|
632 |
+
the "copyright" line and a pointer to where the full notice is found.
|
633 |
+
|
634 |
+
<one line to give the program's name and a brief idea of what it does.>
|
635 |
+
Copyright (C) <year> <name of author>
|
636 |
+
|
637 |
+
This program is free software: you can redistribute it and/or modify
|
638 |
+
it under the terms of the GNU General Public License as published by
|
639 |
+
the Free Software Foundation, either version 3 of the License, or
|
640 |
+
(at your option) any later version.
|
641 |
+
|
642 |
+
This program is distributed in the hope that it will be useful,
|
643 |
+
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
644 |
+
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
645 |
+
GNU General Public License for more details.
|
646 |
+
|
647 |
+
You should have received a copy of the GNU General Public License
|
648 |
+
along with this program. If not, see <https://www.gnu.org/licenses/>.
|
649 |
+
|
650 |
+
Also add information on how to contact you by electronic and paper mail.
|
651 |
+
|
652 |
+
If the program does terminal interaction, make it output a short
|
653 |
+
notice like this when it starts in an interactive mode:
|
654 |
+
|
655 |
+
<program> Copyright (C) <year> <name of author>
|
656 |
+
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
|
657 |
+
This is free software, and you are welcome to redistribute it
|
658 |
+
under certain conditions; type `show c' for details.
|
659 |
+
|
660 |
+
The hypothetical commands `show w' and `show c' should show the appropriate
|
661 |
+
parts of the General Public License. Of course, your program's commands
|
662 |
+
might be different; for a GUI interface, you would use an "about box".
|
663 |
+
|
664 |
+
You should also get your employer (if you work as a programmer) or school,
|
665 |
+
if any, to sign a "copyright disclaimer" for the program, if necessary.
|
666 |
+
For more information on this, and how to apply and follow the GNU GPL, see
|
667 |
+
<https://www.gnu.org/licenses/>.
|
668 |
+
|
669 |
+
The GNU General Public License does not permit incorporating your program
|
670 |
+
into proprietary programs. If your program is a subroutine library, you
|
671 |
+
may consider it more useful to permit linking proprietary applications with
|
672 |
+
the library. If this is what you want to do, use the GNU Lesser General
|
673 |
+
Public License instead of this License. But first, please read
|
674 |
+
<https://www.gnu.org/licenses/why-not-lgpl.html>. GNU GENERAL PUBLIC LICENSE
|
675 |
+
Version 3, 29 June 2007
|
676 |
+
|
677 |
+
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
|
678 |
+
Everyone is permitted to copy and distribute verbatim copies
|
679 |
+
of this license document, but changing it is not allowed.
|
680 |
+
|
681 |
+
Preamble
|
682 |
+
|
683 |
+
The GNU General Public License is a free, copyleft license for
|
684 |
+
software and other kinds of works.
|
685 |
+
|
686 |
+
The licenses for most software and other practical works are designed
|
687 |
+
to take away your freedom to share and change the works. By contrast,
|
688 |
+
the GNU General Public License is intended to guarantee your freedom to
|
689 |
+
share and change all versions of a program--to make sure it remains free
|
690 |
+
software for all its users. We, the Free Software Foundation, use the
|
691 |
+
GNU General Public License for most of our software; it applies also to
|
692 |
+
any other work released this way by its authors. You can apply it to
|
693 |
+
your programs, too.
|
694 |
+
|
695 |
+
When we speak of free software, we are referring to freedom, not
|
696 |
+
price. Our General Public Licenses are designed to make sure that you
|
697 |
+
have the freedom to distribute copies of free software (and charge for
|
698 |
+
them if you wish), that you receive source code or can get it if you
|
699 |
+
want it, that you can change the software or use pieces of it in new
|
700 |
+
free programs, and that you know you can do these things.
|
701 |
+
|
702 |
+
To protect your rights, we need to prevent others from denying you
|
703 |
+
these rights or asking you to surrender the rights. Therefore, you have
|
704 |
+
certain responsibilities if you distribute copies of the software, or if
|
705 |
+
you modify it: responsibilities to respect the freedom of others.
|
706 |
+
|
707 |
+
For example, if you distribute copies of such a program, whether
|
708 |
+
gratis or for a fee, you must pass on to the recipients the same
|
709 |
+
freedoms that you received. You must make sure that they, too, receive
|
710 |
+
or can get the source code. And you must show them these terms so they
|
711 |
+
know their rights.
|
712 |
+
|
713 |
+
Developers that use the GNU GPL protect your rights with two steps:
|
714 |
+
(1) assert copyright on the software, and (2) offer you this License
|
715 |
+
giving you legal permission to copy, distribute and/or modify it.
|
716 |
+
|
717 |
+
For the developers' and authors' protection, the GPL clearly explains
|
718 |
+
that there is no warranty for this free software. For both users' and
|
719 |
+
authors' sake, the GPL requires that modified versions be marked as
|
720 |
+
changed, so that their problems will not be attributed erroneously to
|
721 |
+
authors of previous versions.
|
722 |
+
|
723 |
+
Some devices are designed to deny users access to install or run
|
724 |
+
modified versions of the software inside them, although the manufacturer
|
725 |
+
can do so. This is fundamentally incompatible with the aim of
|
726 |
+
protecting users' freedom to change the software. The systematic
|
727 |
+
pattern of such abuse occurs in the area of products for individuals to
|
728 |
+
use, which is precisely where it is most unacceptable. Therefore, we
|
729 |
+
have designed this version of the GPL to prohibit the practice for those
|
730 |
+
products. If such problems arise substantially in other domains, we
|
731 |
+
stand ready to extend this provision to those domains in future versions
|
732 |
+
of the GPL, as needed to protect the freedom of users.
|
733 |
+
|
734 |
+
Finally, every program is threatened constantly by software patents.
|
735 |
+
States should not allow patents to restrict development and use of
|
736 |
+
software on general-purpose computers, but in those that do, we wish to
|
737 |
+
avoid the special danger that patents applied to a free program could
|
738 |
+
make it effectively proprietary. To prevent this, the GPL assures that
|
739 |
+
patents cannot be used to render the program non-free.
|
740 |
+
|
741 |
+
The precise terms and conditions for copying, distribution and
|
742 |
+
modification follow.
|
743 |
+
|
744 |
+
TERMS AND CONDITIONS
|
745 |
+
|
746 |
+
0. Definitions.
|
747 |
+
|
748 |
+
"This License" refers to version 3 of the GNU General Public License.
|
749 |
+
|
750 |
+
"Copyright" also means copyright-like laws that apply to other kinds of
|
751 |
+
works, such as semiconductor masks.
|
752 |
+
|
753 |
+
"The Program" refers to any copyrightable work licensed under this
|
754 |
+
License. Each licensee is addressed as "you". "Licensees" and
|
755 |
+
"recipients" may be individuals or organizations.
|
756 |
+
|
757 |
+
To "modify" a work means to copy from or adapt all or part of the work
|
758 |
+
in a fashion requiring copyright permission, other than the making of an
|
759 |
+
exact copy. The resulting work is called a "modified version" of the
|
760 |
+
earlier work or a work "based on" the earlier work.
|
761 |
+
|
762 |
+
A "covered work" means either the unmodified Program or a work based
|
763 |
+
on the Program.
|
764 |
+
|
765 |
+
To "propagate" a work means to do anything with it that, without
|
766 |
+
permission, would make you directly or secondarily liable for
|
767 |
+
infringement under applicable copyright law, except executing it on a
|
768 |
+
computer or modifying a private copy. Propagation includes copying,
|
769 |
+
distribution (with or without modification), making available to the
|
770 |
+
public, and in some countries other activities as well.
|
771 |
+
|
772 |
+
To "convey" a work means any kind of propagation that enables other
|
773 |
+
parties to make or receive copies. Mere interaction with a user through
|
774 |
+
a computer network, with no transfer of a copy, is not conveying.
|
775 |
+
|
776 |
+
An interactive user interface displays "Appropriate Legal Notices"
|
777 |
+
to the extent that it includes a convenient and prominently visible
|
778 |
+
feature that (1) displays an appropriate copyright notice, and (2)
|
779 |
+
tells the user that there is no warranty for the work (except to the
|
780 |
+
extent that warranties are provided), that licensees may convey the
|
781 |
+
work under this License, and how to view a copy of this License. If
|
782 |
+
the interface presents a list of user commands or options, such as a
|
783 |
+
menu, a prominent item in the list meets this criterion.
|
784 |
+
|
785 |
+
1. Source Code.
|
786 |
+
|
787 |
+
The "source code" for a work means the preferred form of the work
|
788 |
+
for making modifications to it. "Object code" means any non-source
|
789 |
+
form of a work.
|
790 |
+
|
791 |
+
A "Standard Interface" means an interface that either is an official
|
792 |
+
standard defined by a recognized standards body, or, in the case of
|
793 |
+
interfaces specified for a particular programming language, one that
|
794 |
+
is widely used among developers working in that language.
|
795 |
+
|
796 |
+
The "System Libraries" of an executable work include anything, other
|
797 |
+
than the work as a whole, that (a) is included in the normal form of
|
798 |
+
packaging a Major Component, but which is not part of that Major
|
799 |
+
Component, and (b) serves only to enable use of the work with that
|
800 |
+
Major Component, or to implement a Standard Interface for which an
|
801 |
+
implementation is available to the public in source code form. A
|
802 |
+
"Major Component", in this context, means a major essential component
|
803 |
+
(kernel, window system, and so on) of the specific operating system
|
804 |
+
(if any) on which the executable work runs, or a compiler used to
|
805 |
+
produce the work, or an object code interpreter used to run it.
|
806 |
+
|
807 |
+
The "Corresponding Source" for a work in object code form means all
|
808 |
+
the source code needed to generate, install, and (for an executable
|
809 |
+
work) run the object code and to modify the work, including scripts to
|
810 |
+
control those activities. However, it does not include the work's
|
811 |
+
System Libraries, or general-purpose tools or generally available free
|
812 |
+
programs which are used unmodified in performing those activities but
|
813 |
+
which are not part of the work. For example, Corresponding Source
|
814 |
+
includes interface definition files associated with source files for
|
815 |
+
the work, and the source code for shared libraries and dynamically
|
816 |
+
linked subprograms that the work is specifically designed to require,
|
817 |
+
such as by intimate data communication or control flow between those
|
818 |
+
subprograms and other parts of the work.
|
819 |
+
|
820 |
+
The Corresponding Source need not include anything that users
|
821 |
+
can regenerate automatically from other parts of the Corresponding
|
822 |
+
Source.
|
823 |
+
|
824 |
+
The Corresponding Source for a work in source code form is that
|
825 |
+
same work.
|
826 |
+
|
827 |
+
2. Basic Permissions.
|
828 |
+
|
829 |
+
All rights granted under this License are granted for the term of
|
830 |
+
copyright on the Program, and are irrevocable provided the stated
|
831 |
+
conditions are met. This License explicitly affirms your unlimited
|
832 |
+
permission to run the unmodified Program. The output from running a
|
833 |
+
covered work is covered by this License only if the output, given its
|
834 |
+
content, constitutes a covered work. This License acknowledges your
|
835 |
+
rights of fair use or other equivalent, as provided by copyright law.
|
836 |
+
|
837 |
+
You may make, run and propagate covered works that you do not
|
838 |
+
convey, without conditions so long as your license otherwise remains
|
839 |
+
in force. You may convey covered works to others for the sole purpose
|
840 |
+
of having them make modifications exclusively for you, or provide you
|
841 |
+
with facilities for running those works, provided that you comply with
|
842 |
+
the terms of this License in conveying all material for which you do
|
843 |
+
not control copyright. Those thus making or running the covered works
|
844 |
+
for you must do so exclusively on your behalf, under your direction
|
845 |
+
and control, on terms that prohibit them from making any copies of
|
846 |
+
your copyrighted material outside their relationship with you.
|
847 |
+
|
848 |
+
Conveying under any other circumstances is permitted solely under
|
849 |
+
the conditions stated below. Sublicensing is not allowed; section 10
|
850 |
+
makes it unnecessary.
|
851 |
+
|
852 |
+
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
|
853 |
+
|
854 |
+
No covered work shall be deemed part of an effective technological
|
855 |
+
measure under any applicable law fulfilling obligations under article
|
856 |
+
11 of the WIPO copyright treaty adopted on 20 December 1996, or
|
857 |
+
similar laws prohibiting or restricting circumvention of such
|
858 |
+
measures.
|
859 |
+
|
860 |
+
When you convey a covered work, you waive any legal power to forbid
|
861 |
+
circumvention of technological measures to the extent such circumvention
|
862 |
+
is effected by exercising rights under this License with respect to
|
863 |
+
the covered work, and you disclaim any intention to limit operation or
|
864 |
+
modification of the work as a means of enforcing, against the work's
|
865 |
+
users, your or third parties' legal rights to forbid circumvention of
|
866 |
+
technological measures.
|
867 |
+
|
868 |
+
4. Conveying Verbatim Copies.
|
869 |
+
|
870 |
+
You may convey verbatim copies of the Program's source code as you
|
871 |
+
receive it, in any medium, provided that you conspicuously and
|
872 |
+
appropriately publish on each copy an appropriate copyright notice;
|
873 |
+
keep intact all notices stating that this License and any
|
874 |
+
non-permissive terms added in accord with section 7 apply to the code;
|
875 |
+
keep intact all notices of the absence of any warranty; and give all
|
876 |
+
recipients a copy of this License along with the Program.
|
877 |
+
|
878 |
+
You may charge any price or no price for each copy that you convey,
|
879 |
+
and you may offer support or warranty protection for a fee.
|
880 |
+
|
881 |
+
5. Conveying Modified Source Versions.
|
882 |
+
|
883 |
+
You may convey a work based on the Program, or the modifications to
|
884 |
+
produce it from the Program, in the form of source code under the
|
885 |
+
terms of section 4, provided that you also meet all of these conditions:
|
886 |
+
|
887 |
+
a) The work must carry prominent notices stating that you modified
|
888 |
+
it, and giving a relevant date.
|
889 |
+
|
890 |
+
b) The work must carry prominent notices stating that it is
|
891 |
+
released under this License and any conditions added under section
|
892 |
+
7. This requirement modifies the requirement in section 4 to
|
893 |
+
"keep intact all notices".
|
894 |
+
|
895 |
+
c) You must license the entire work, as a whole, under this
|
896 |
+
License to anyone who comes into possession of a copy. This
|
897 |
+
License will therefore apply, along with any applicable section 7
|
898 |
+
additional terms, to the whole of the work, and all its parts,
|
899 |
+
regardless of how they are packaged. This License gives no
|
900 |
+
permission to license the work in any other way, but it does not
|
901 |
+
invalidate such permission if you have separately received it.
|
902 |
+
|
903 |
+
d) If the work has interactive user interfaces, each must display
|
904 |
+
Appropriate Legal Notices; however, if the Program has interactive
|
905 |
+
interfaces that do not display Appropriate Legal Notices, your
|
906 |
+
work need not make them do so.
|
907 |
+
|
908 |
+
A compilation of a covered work with other separate and independent
|
909 |
+
works, which are not by their nature extensions of the covered work,
|
910 |
+
and which are not combined with it such as to form a larger program,
|
911 |
+
in or on a volume of a storage or distribution medium, is called an
|
912 |
+
"aggregate" if the compilation and its resulting copyright are not
|
913 |
+
used to limit the access or legal rights of the compilation's users
|
914 |
+
beyond what the individual works permit. Inclusion of a covered work
|
915 |
+
in an aggregate does not cause this License to apply to the other
|
916 |
+
parts of the aggregate.
|
917 |
+
|
918 |
+
6. Conveying Non-Source Forms.
|
919 |
+
|
920 |
+
You may convey a covered work in object code form under the terms
|
921 |
+
of sections 4 and 5, provided that you also convey the
|
922 |
+
machine-readable Corresponding Source under the terms of this License,
|
923 |
+
in one of these ways:
|
924 |
+
|
925 |
+
a) Convey the object code in, or embodied in, a physical product
|
926 |
+
(including a physical distribution medium), accompanied by the
|
927 |
+
Corresponding Source fixed on a durable physical medium
|
928 |
+
customarily used for software interchange.
|
929 |
+
|
930 |
+
b) Convey the object code in, or embodied in, a physical product
|
931 |
+
(including a physical distribution medium), accompanied by a
|
932 |
+
written offer, valid for at least three years and valid for as
|
933 |
+
long as you offer spare parts or customer support for that product
|
934 |
+
model, to give anyone who possesses the object code either (1) a
|
935 |
+
copy of the Corresponding Source for all the software in the
|
936 |
+
product that is covered by this License, on a durable physical
|
937 |
+
medium customarily used for software interchange, for a price no
|
938 |
+
more than your reasonable cost of physically performing this
|
939 |
+
conveying of source, or (2) access to copy the
|
940 |
+
Corresponding Source from a network server at no charge.
|
941 |
+
|
942 |
+
c) Convey individual copies of the object code with a copy of the
|
943 |
+
written offer to provide the Corresponding Source. This
|
944 |
+
alternative is allowed only occasionally and noncommercially, and
|
945 |
+
only if you received the object code with such an offer, in accord
|
946 |
+
with subsection 6b.
|
947 |
+
|
948 |
+
d) Convey the object code by offering access from a designated
|
949 |
+
place (gratis or for a charge), and offer equivalent access to the
|
950 |
+
Corresponding Source in the same way through the same place at no
|
951 |
+
further charge. You need not require recipients to copy the
|
952 |
+
Corresponding Source along with the object code. If the place to
|
953 |
+
copy the object code is a network server, the Corresponding Source
|
954 |
+
may be on a different server (operated by you or a third party)
|
955 |
+
that supports equivalent copying facilities, provided you maintain
|
956 |
+
clear directions next to the object code saying where to find the
|
957 |
+
Corresponding Source. Regardless of what server hosts the
|
958 |
+
Corresponding Source, you remain obligated to ensure that it is
|
959 |
+
available for as long as needed to satisfy these requirements.
|
960 |
+
|
961 |
+
e) Convey the object code using peer-to-peer transmission, provided
|
962 |
+
you inform other peers where the object code and Corresponding
|
963 |
+
Source of the work are being offered to the general public at no
|
964 |
+
charge under subsection 6d.
|
965 |
+
|
966 |
+
A separable portion of the object code, whose source code is excluded
|
967 |
+
from the Corresponding Source as a System Library, need not be
|
968 |
+
included in conveying the object code work.
|
969 |
+
|
970 |
+
A "User Product" is either (1) a "consumer product", which means any
|
971 |
+
tangible personal property which is normally used for personal, family,
|
972 |
+
or household purposes, or (2) anything designed or sold for incorporation
|
973 |
+
into a dwelling. In determining whether a product is a consumer product,
|
974 |
+
doubtful cases shall be resolved in favor of coverage. For a particular
|
975 |
+
product received by a particular user, "normally used" refers to a
|
976 |
+
typical or common use of that class of product, regardless of the status
|
977 |
+
of the particular user or of the way in which the particular user
|
978 |
+
actually uses, or expects or is expected to use, the product. A product
|
979 |
+
is a consumer product regardless of whether the product has substantial
|
980 |
+
commercial, industrial or non-consumer uses, unless such uses represent
|
981 |
+
the only significant mode of use of the product.
|
982 |
+
|
983 |
+
"Installation Information" for a User Product means any methods,
|
984 |
+
procedures, authorization keys, or other information required to install
|
985 |
+
and execute modified versions of a covered work in that User Product from
|
986 |
+
a modified version of its Corresponding Source. The information must
|
987 |
+
suffice to ensure that the continued functioning of the modified object
|
988 |
+
code is in no case prevented or interfered with solely because
|
989 |
+
modification has been made.
|
990 |
+
|
991 |
+
If you convey an object code work under this section in, or with, or
|
992 |
+
specifically for use in, a User Product, and the conveying occurs as
|
993 |
+
part of a transaction in which the right of possession and use of the
|
994 |
+
User Product is transferred to the recipient in perpetuity or for a
|
995 |
+
fixed term (regardless of how the transaction is characterized), the
|
996 |
+
Corresponding Source conveyed under this section must be accompanied
|
997 |
+
by the Installation Information. But this requirement does not apply
|
998 |
+
if neither you nor any third party retains the ability to install
|
999 |
+
modified object code on the User Product (for example, the work has
|
1000 |
+
been installed in ROM).
|
1001 |
+
|
1002 |
+
The requirement to provide Installation Information does not include a
|
1003 |
+
requirement to continue to provide support service, warranty, or updates
|
1004 |
+
for a work that has been modified or installed by the recipient, or for
|
1005 |
+
the User Product in which it has been modified or installed. Access to a
|
1006 |
+
network may be denied when the modification itself materially and
|
1007 |
+
adversely affects the operation of the network or violates the rules and
|
1008 |
+
protocols for communication across the network.
|
1009 |
+
|
1010 |
+
Corresponding Source conveyed, and Installation Information provided,
|
1011 |
+
in accord with this section must be in a format that is publicly
|
1012 |
+
documented (and with an implementation available to the public in
|
1013 |
+
source code form), and must require no special password or key for
|
1014 |
+
unpacking, reading or copying.
|
1015 |
+
|
1016 |
+
7. Additional Terms.
|
1017 |
+
|
1018 |
+
"Additional permissions" are terms that supplement the terms of this
|
1019 |
+
License by making exceptions from one or more of its conditions.
|
1020 |
+
Additional permissions that are applicable to the entire Program shall
|
1021 |
+
be treated as though they were included in this License, to the extent
|
1022 |
+
that they are valid under applicable law. If additional permissions
|
1023 |
+
apply only to part of the Program, that part may be used separately
|
1024 |
+
under those permissions, but the entire Program remains governed by
|
1025 |
+
this License without regard to the additional permissions.
|
1026 |
+
|
1027 |
+
When you convey a copy of a covered work, you may at your option
|
1028 |
+
remove any additional permissions from that copy, or from any part of
|
1029 |
+
it. (Additional permissions may be written to require their own
|
1030 |
+
removal in certain cases when you modify the work.) You may place
|
1031 |
+
additional permissions on material, added by you to a covered work,
|
1032 |
+
for which you have or can give appropriate copyright permission.
|
1033 |
+
|
1034 |
+
Notwithstanding any other provision of this License, for material you
|
1035 |
+
add to a covered work, you may (if authorized by the copyright holders of
|
1036 |
+
that material) supplement the terms of this License with terms:
|
1037 |
+
|
1038 |
+
a) Disclaiming warranty or limiting liability differently from the
|
1039 |
+
terms of sections 15 and 16 of this License; or
|
1040 |
+
|
1041 |
+
b) Requiring preservation of specified reasonable legal notices or
|
1042 |
+
author attributions in that material or in the Appropriate Legal
|
1043 |
+
Notices displayed by works containing it; or
|
1044 |
+
|
1045 |
+
c) Prohibiting misrepresentation of the origin of that material, or
|
1046 |
+
requiring that modified versions of such material be marked in
|
1047 |
+
reasonable ways as different from the original version; or
|
1048 |
+
|
1049 |
+
d) Limiting the use for publicity purposes of names of licensors or
|
1050 |
+
authors of the material; or
|
1051 |
+
|
1052 |
+
e) Declining to grant rights under trademark law for use of some
|
1053 |
+
trade names, trademarks, or service marks; or
|
1054 |
+
|
1055 |
+
f) Requiring indemnification of licensors and authors of that
|
1056 |
+
material by anyone who conveys the material (or modified versions of
|
1057 |
+
it) with contractual assumptions of liability to the recipient, for
|
1058 |
+
any liability that these contractual assumptions directly impose on
|
1059 |
+
those licensors and authors.
|
1060 |
+
|
1061 |
+
All other non-permissive additional terms are considered "further
|
1062 |
+
restrictions" within the meaning of section 10. If the Program as you
|
1063 |
+
received it, or any part of it, contains a notice stating that it is
|
1064 |
+
governed by this License along with a term that is a further
|
1065 |
+
restriction, you may remove that term. If a license document contains
|
1066 |
+
a further restriction but permits relicensing or conveying under this
|
1067 |
+
License, you may add to a covered work material governed by the terms
|
1068 |
+
of that license document, provided that the further restriction does
|
1069 |
+
not survive such relicensing or conveying.
|
1070 |
+
|
1071 |
+
If you add terms to a covered work in accord with this section, you
|
1072 |
+
must place, in the relevant source files, a statement of the
|
1073 |
+
additional terms that apply to those files, or a notice indicating
|
1074 |
+
where to find the applicable terms.
|
1075 |
+
|
1076 |
+
Additional terms, permissive or non-permissive, may be stated in the
|
1077 |
+
form of a separately written license, or stated as exceptions;
|
1078 |
+
the above requirements apply either way.
|
1079 |
+
|
1080 |
+
8. Termination.
|
1081 |
+
|
1082 |
+
You may not propagate or modify a covered work except as expressly
|
1083 |
+
provided under this License. Any attempt otherwise to propagate or
|
1084 |
+
modify it is void, and will automatically terminate your rights under
|
1085 |
+
this License (including any patent licenses granted under the third
|
1086 |
+
paragraph of section 11).
|
1087 |
+
|
1088 |
+
However, if you cease all violation of this License, then your
|
1089 |
+
license from a particular copyright holder is reinstated (a)
|
1090 |
+
provisionally, unless and until the copyright holder explicitly and
|
1091 |
+
finally terminates your license, and (b) permanently, if the copyright
|
1092 |
+
holder fails to notify you of the violation by some reasonable means
|
1093 |
+
prior to 60 days after the cessation.
|
1094 |
+
|
1095 |
+
Moreover, your license from a particular copyright holder is
|
1096 |
+
reinstated permanently if the copyright holder notifies you of the
|
1097 |
+
violation by some reasonable means, this is the first time you have
|
1098 |
+
received notice of violation of this License (for any work) from that
|
1099 |
+
copyright holder, and you cure the violation prior to 30 days after
|
1100 |
+
your receipt of the notice.
|
1101 |
+
|
1102 |
+
Termination of your rights under this section does not terminate the
|
1103 |
+
licenses of parties who have received copies or rights from you under
|
1104 |
+
this License. If your rights have been terminated and not permanently
|
1105 |
+
reinstated, you do not qualify to receive new licenses for the same
|
1106 |
+
material under section 10.
|
1107 |
+
|
1108 |
+
9. Acceptance Not Required for Having Copies.
|
1109 |
+
|
1110 |
+
You are not required to accept this License in order to receive or
|
1111 |
+
run a copy of the Program. Ancillary propagation of a covered work
|
1112 |
+
occurring solely as a consequence of using peer-to-peer transmission
|
1113 |
+
to receive a copy likewise does not require acceptance. However,
|
1114 |
+
nothing other than this License grants you permission to propagate or
|
1115 |
+
modify any covered work. These actions infringe copyright if you do
|
1116 |
+
not accept this License. Therefore, by modifying or propagating a
|
1117 |
+
covered work, you indicate your acceptance of this License to do so.
|
1118 |
+
|
1119 |
+
10. Automatic Licensing of Downstream Recipients.
|
1120 |
+
|
1121 |
+
Each time you convey a covered work, the recipient automatically
|
1122 |
+
receives a license from the original licensors, to run, modify and
|
1123 |
+
propagate that work, subject to this License. You are not responsible
|
1124 |
+
for enforcing compliance by third parties with this License.
|
1125 |
+
|
1126 |
+
An "entity transaction" is a transaction transferring control of an
|
1127 |
+
organization, or substantially all assets of one, or subdividing an
|
1128 |
+
organization, or merging organizations. If propagation of a covered
|
1129 |
+
work results from an entity transaction, each party to that
|
1130 |
+
transaction who receives a copy of the work also receives whatever
|
1131 |
+
licenses to the work the party's predecessor in interest had or could
|
1132 |
+
give under the previous paragraph, plus a right to possession of the
|
1133 |
+
Corresponding Source of the work from the predecessor in interest, if
|
1134 |
+
the predecessor has it or can get it with reasonable efforts.
|
1135 |
+
|
1136 |
+
You may not impose any further restrictions on the exercise of the
|
1137 |
+
rights granted or affirmed under this License. For example, you may
|
1138 |
+
not impose a license fee, royalty, or other charge for exercise of
|
1139 |
+
rights granted under this License, and you may not initiate litigation
|
1140 |
+
(including a cross-claim or counterclaim in a lawsuit) alleging that
|
1141 |
+
any patent claim is infringed by making, using, selling, offering for
|
1142 |
+
sale, or importing the Program or any portion of it.
|
1143 |
+
|
1144 |
+
11. Patents.
|
1145 |
+
|
1146 |
+
A "contributor" is a copyright holder who authorizes use under this
|
1147 |
+
License of the Program or a work on which the Program is based. The
|
1148 |
+
work thus licensed is called the contributor's "contributor version".
|
1149 |
+
|
1150 |
+
A contributor's "essential patent claims" are all patent claims
|
1151 |
+
owned or controlled by the contributor, whether already acquired or
|
1152 |
+
hereafter acquired, that would be infringed by some manner, permitted
|
1153 |
+
by this License, of making, using, or selling its contributor version,
|
1154 |
+
but do not include claims that would be infringed only as a
|
1155 |
+
consequence of further modification of the contributor version. For
|
1156 |
+
purposes of this definition, "control" includes the right to grant
|
1157 |
+
patent sublicenses in a manner consistent with the requirements of
|
1158 |
+
this License.
|
1159 |
+
|
1160 |
+
Each contributor grants you a non-exclusive, worldwide, royalty-free
|
1161 |
+
patent license under the contributor's essential patent claims, to
|
1162 |
+
make, use, sell, offer for sale, import and otherwise run, modify and
|
1163 |
+
propagate the contents of its contributor version.
|
1164 |
+
|
1165 |
+
In the following three paragraphs, a "patent license" is any express
|
1166 |
+
agreement or commitment, however denominated, not to enforce a patent
|
1167 |
+
(such as an express permission to practice a patent or covenant not to
|
1168 |
+
sue for patent infringement). To "grant" such a patent license to a
|
1169 |
+
party means to make such an agreement or commitment not to enforce a
|
1170 |
+
patent against the party.
|
1171 |
+
|
1172 |
+
If you convey a covered work, knowingly relying on a patent license,
|
1173 |
+
and the Corresponding Source of the work is not available for anyone
|
1174 |
+
to copy, free of charge and under the terms of this License, through a
|
1175 |
+
publicly available network server or other readily accessible means,
|
1176 |
+
then you must either (1) cause the Corresponding Source to be so
|
1177 |
+
available, or (2) arrange to deprive yourself of the benefit of the
|
1178 |
+
patent license for this particular work, or (3) arrange, in a manner
|
1179 |
+
consistent with the requirements of this License, to extend the patent
|
1180 |
+
license to downstream recipients. "Knowingly relying" means you have
|
1181 |
+
actual knowledge that, but for the patent license, your conveying the
|
1182 |
+
covered work in a country, or your recipient's use of the covered work
|
1183 |
+
in a country, would infringe one or more identifiable patents in that
|
1184 |
+
country that you have reason to believe are valid.
|
1185 |
+
|
1186 |
+
If, pursuant to or in connection with a single transaction or
|
1187 |
+
arrangement, you convey, or propagate by procuring conveyance of, a
|
1188 |
+
covered work, and grant a patent license to some of the parties
|
1189 |
+
receiving the covered work authorizing them to use, propagate, modify
|
1190 |
+
or convey a specific copy of the covered work, then the patent license
|
1191 |
+
you grant is automatically extended to all recipients of the covered
|
1192 |
+
work and works based on it.
|
1193 |
+
|
1194 |
+
A patent license is "discriminatory" if it does not include within
|
1195 |
+
the scope of its coverage, prohibits the exercise of, or is
|
1196 |
+
conditioned on the non-exercise of one or more of the rights that are
|
1197 |
+
specifically granted under this License. You may not convey a covered
|
1198 |
+
work if you are a party to an arrangement with a third party that is
|
1199 |
+
in the business of distributing software, under which you make payment
|
1200 |
+
to the third party based on the extent of your activity of conveying
|
1201 |
+
the work, and under which the third party grants, to any of the
|
1202 |
+
parties who would receive the covered work from you, a discriminatory
|
1203 |
+
patent license (a) in connection with copies of the covered work
|
1204 |
+
conveyed by you (or copies made from those copies), or (b) primarily
|
1205 |
+
for and in connection with specific products or compilations that
|
1206 |
+
contain the covered work, unless you entered into that arrangement,
|
1207 |
+
or that patent license was granted, prior to 28 March 2007.
|
1208 |
+
|
1209 |
+
Nothing in this License shall be construed as excluding or limiting
|
1210 |
+
any implied license or other defenses to infringement that may
|
1211 |
+
otherwise be available to you under applicable patent law.
|
1212 |
+
|
1213 |
+
12. No Surrender of Others' Freedom.
|
1214 |
+
|
1215 |
+
If conditions are imposed on you (whether by court order, agreement or
|
1216 |
+
otherwise) that contradict the conditions of this License, they do not
|
1217 |
+
excuse you from the conditions of this License. If you cannot convey a
|
1218 |
+
covered work so as to satisfy simultaneously your obligations under this
|
1219 |
+
License and any other pertinent obligations, then as a consequence you may
|
1220 |
+
not convey it at all. For example, if you agree to terms that obligate you
|
1221 |
+
to collect a royalty for further conveying from those to whom you convey
|
1222 |
+
the Program, the only way you could satisfy both those terms and this
|
1223 |
+
License would be to refrain entirely from conveying the Program.
|
1224 |
+
|
1225 |
+
13. Use with the GNU Affero General Public License.
|
1226 |
+
|
1227 |
+
Notwithstanding any other provision of this License, you have
|
1228 |
+
permission to link or combine any covered work with a work licensed
|
1229 |
+
under version 3 of the GNU Affero General Public License into a single
|
1230 |
+
combined work, and to convey the resulting work. The terms of this
|
1231 |
+
License will continue to apply to the part which is the covered work,
|
1232 |
+
but the special requirements of the GNU Affero General Public License,
|
1233 |
+
section 13, concerning interaction through a network will apply to the
|
1234 |
+
combination as such.
|
1235 |
+
|
1236 |
+
14. Revised Versions of this License.
|
1237 |
+
|
1238 |
+
The Free Software Foundation may publish revised and/or new versions of
|
1239 |
+
the GNU General Public License from time to time. Such new versions will
|
1240 |
+
be similar in spirit to the present version, but may differ in detail to
|
1241 |
+
address new problems or concerns.
|
1242 |
+
|
1243 |
+
Each version is given a distinguishing version number. If the
|
1244 |
+
Program specifies that a certain numbered version of the GNU General
|
1245 |
+
Public License "or any later version" applies to it, you have the
|
1246 |
+
option of following the terms and conditions either of that numbered
|
1247 |
+
version or of any later version published by the Free Software
|
1248 |
+
Foundation. If the Program does not specify a version number of the
|
1249 |
+
GNU General Public License, you may choose any version ever published
|
1250 |
+
by the Free Software Foundation.
|
1251 |
+
|
1252 |
+
If the Program specifies that a proxy can decide which future
|
1253 |
+
versions of the GNU General Public License can be used, that proxy's
|
1254 |
+
public statement of acceptance of a version permanently authorizes you
|
1255 |
+
to choose that version for the Program.
|
1256 |
+
|
1257 |
+
Later license versions may give you additional or different
|
1258 |
+
permissions. However, no additional obligations are imposed on any
|
1259 |
+
author or copyright holder as a result of your choosing to follow a
|
1260 |
+
later version.
|
1261 |
+
|
1262 |
+
15. Disclaimer of Warranty.
|
1263 |
+
|
1264 |
+
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
|
1265 |
+
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
|
1266 |
+
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
|
1267 |
+
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
|
1268 |
+
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
1269 |
+
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
|
1270 |
+
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
|
1271 |
+
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
|
1272 |
+
|
1273 |
+
16. Limitation of Liability.
|
1274 |
+
|
1275 |
+
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
|
1276 |
+
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
|
1277 |
+
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
|
1278 |
+
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
|
1279 |
+
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
|
1280 |
+
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
|
1281 |
+
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
|
1282 |
+
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
|
1283 |
+
SUCH DAMAGES.
|
1284 |
+
|
1285 |
+
17. Interpretation of Sections 15 and 16.
|
1286 |
+
|
1287 |
+
If the disclaimer of warranty and limitation of liability provided
|
1288 |
+
above cannot be given local legal effect according to their terms,
|
1289 |
+
reviewing courts shall apply local law that most closely approximates
|
1290 |
+
an absolute waiver of all civil liability in connection with the
|
1291 |
+
Program, unless a warranty or assumption of liability accompanies a
|
1292 |
+
copy of the Program in return for a fee.
|
1293 |
+
|
1294 |
+
END OF TERMS AND CONDITIONS
|
1295 |
+
|
1296 |
+
How to Apply These Terms to Your New Programs
|
1297 |
+
|
1298 |
+
If you develop a new program, and you want it to be of the greatest
|
1299 |
+
possible use to the public, the best way to achieve this is to make it
|
1300 |
+
free software which everyone can redistribute and change under these terms.
|
1301 |
+
|
1302 |
+
To do so, attach the following notices to the program. It is safest
|
1303 |
+
to attach them to the start of each source file to most effectively
|
1304 |
+
state the exclusion of warranty; and each file should have at least
|
1305 |
+
the "copyright" line and a pointer to where the full notice is found.
|
1306 |
+
|
1307 |
+
<one line to give the program's name and a brief idea of what it does.>
|
1308 |
+
Copyright (C) <year> <name of author>
|
1309 |
+
|
1310 |
+
This program is free software: you can redistribute it and/or modify
|
1311 |
+
it under the terms of the GNU General Public License as published by
|
1312 |
+
the Free Software Foundation, either version 3 of the License, or
|
1313 |
+
(at your option) any later version.
|
1314 |
+
|
1315 |
+
This program is distributed in the hope that it will be useful,
|
1316 |
+
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
1317 |
+
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
1318 |
+
GNU General Public License for more details.
|
1319 |
+
|
1320 |
+
You should have received a copy of the GNU General Public License
|
1321 |
+
along with this program. If not, see <https://www.gnu.org/licenses/>.
|
1322 |
+
|
1323 |
+
Also add information on how to contact you by electronic and paper mail.
|
1324 |
+
|
1325 |
+
If the program does terminal interaction, make it output a short
|
1326 |
+
notice like this when it starts in an interactive mode:
|
1327 |
+
|
1328 |
+
<program> Copyright (C) <year> <name of author>
|
1329 |
+
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
|
1330 |
+
This is free software, and you are welcome to redistribute it
|
1331 |
+
under certain conditions; type `show c' for details.
|
1332 |
+
|
1333 |
+
The hypothetical commands `show w' and `show c' should show the appropriate
|
1334 |
+
parts of the General Public License. Of course, your program's commands
|
1335 |
+
might be different; for a GUI interface, you would use an "about box".
|
1336 |
+
|
1337 |
+
You should also get your employer (if you work as a programmer) or school,
|
1338 |
+
if any, to sign a "copyright disclaimer" for the program, if necessary.
|
1339 |
+
For more information on this, and how to apply and follow the GNU GPL, see
|
1340 |
+
<https://www.gnu.org/licenses/>.
|
1341 |
+
|
1342 |
+
The GNU General Public License does not permit incorporating your program
|
1343 |
+
into proprietary programs. If your program is a subroutine library, you
|
1344 |
+
may consider it more useful to permit linking proprietary applications with
|
1345 |
+
the library. If this is what you want to do, use the GNU Lesser General
|
1346 |
+
Public License instead of this License. But first, please read
|
1347 |
+
<https://www.gnu.org/licenses/why-not-lgpl.html>.
|
README.md
CHANGED
@@ -1,12 +1,231 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<div align="center">
|
2 |
+
<img src="./assets/yolo_logo.png" width=60%>
|
3 |
+
<br>
|
4 |
+
<a href="https://scholar.google.com/citations?hl=zh-CN&user=PH8rJHYAAAAJ">Tianheng Cheng</a><sup><span>2,3,*</span></sup>,
|
5 |
+
<a href="https://linsong.info/">Lin Song</a><sup><span>1,📧,*</span></sup>,
|
6 |
+
<a href="https://yxgeee.github.io/">Yixiao Ge</a><sup><span>1,🌟,2</span></sup>,
|
7 |
+
<a href="http://eic.hust.edu.cn/professor/liuwenyu/"> Wenyu Liu</a><sup><span>3</span></sup>,
|
8 |
+
<a href="https://xwcv.github.io/">Xinggang Wang</a><sup><span>3,📧</span></sup>,
|
9 |
+
<a href="https://scholar.google.com/citations?user=4oXBp9UAAAAJ&hl=en">Ying Shan</a><sup><span>1,2</span></sup>
|
10 |
+
</br>
|
11 |
+
|
12 |
+
\* Equal contribution 🌟 Project lead 📧 Corresponding author
|
13 |
+
|
14 |
+
<sup>1</sup> Tencent AI Lab, <sup>2</sup> ARC Lab, Tencent PCG
|
15 |
+
<sup>3</sup> Huazhong University of Science and Technology
|
16 |
+
<br>
|
17 |
+
<div>
|
18 |
+
|
19 |
+
[![arxiv paper](https://img.shields.io/badge/Project-Page-green)](https://wondervictor.github.io/)
|
20 |
+
[![arxiv paper](https://img.shields.io/badge/arXiv-Paper-red)](https://arxiv.org/abs/2401.17270)
|
21 |
+
<a href="https://colab.research.google.com/github/AILab-CVC/YOLO-World/blob/master/inference.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a>
|
22 |
+
[![demo](https://img.shields.io/badge/🤗HugginngFace-Spaces-orange)](https://huggingface.co/spaces/stevengrove/YOLO-World)
|
23 |
+
[![Replicate](https://replicate.com/zsxkib/yolo-world/badge)](https://replicate.com/zsxkib/yolo-world)
|
24 |
+
[![hfpaper](https://img.shields.io/badge/🤗HugginngFace-Paper-yellow)](https://huggingface.co/papers/2401.17270)
|
25 |
+
[![license](https://img.shields.io/badge/License-GPLv3.0-blue)](LICENSE)
|
26 |
+
[![yoloworldseg](https://img.shields.io/badge/YOLOWorldxEfficientSAM-🤗Spaces-orange)](https://huggingface.co/spaces/SkalskiP/YOLO-World)
|
27 |
+
[![yologuide](https://img.shields.io/badge/📖Notebook-roboflow-purple)](https://supervision.roboflow.com/develop/notebooks/zero-shot-object-detection-with-yolo-world)
|
28 |
+
[![deploy](https://media.roboflow.com/deploy.svg)](https://inference.roboflow.com/foundation/yolo_world/)
|
29 |
+
|
30 |
+
</div>
|
31 |
+
</div>
|
32 |
+
|
33 |
+
## Notice
|
34 |
+
|
35 |
+
We recommend that everyone **use English to communicate on issues**, as this helps developers from around the world discuss, share experiences, and answer questions together.
|
36 |
+
|
37 |
+
## 🔥 Updates
|
38 |
+
`[2024-3-28]:` We provide: (1) more high-resolution pre-trained models (e.g., S, M, X) ([#142](https://github.com/AILab-CVC/YOLO-World/issues/142)); (2) pre-trained models with CLIP-Large text encoders. Most importantly, we preliminarily fix the **fine-tuning without `mask-refine`** and explore a new fine-tuning setting ([#160](https://github.com/AILab-CVC/YOLO-World/issues/160),[#76](https://github.com/AILab-CVC/YOLO-World/issues/76)). In addition, fine-tuning YOLO-World with `mask-refine` also obtains significant improvements, check more details in [configs/finetune_coco](./configs/finetune_coco/).
|
39 |
+
`[2024-3-16]:` We fix the bugs about the demo ([#110](https://github.com/AILab-CVC/YOLO-World/issues/110),[#94](https://github.com/AILab-CVC/YOLO-World/issues/94),[#129](https://github.com/AILab-CVC/YOLO-World/issues/129), [#125](https://github.com/AILab-CVC/YOLO-World/issues/125)) with visualizations of segmentation masks, and release [**YOLO-World with Embeddings**](./docs/prompt_yolo_world.md), which supports prompt tuning, text prompts and image prompts.
|
40 |
+
`[2024-3-3]:` We add the **high-resolution YOLO-World**, which supports `1280x1280` resolution with higher accuracy and better performance for small objects!
|
41 |
+
`[2024-2-29]:` We release the newest version of [ **YOLO-World-v2**](./docs/updates.md) with higher accuracy and faster speed! We hope the community can join us to improve YOLO-World!
|
42 |
+
`[2024-2-28]:` Excited to announce that YOLO-World has been accepted by **CVPR 2024**! We're continuing to make YOLO-World faster and stronger, as well as making it better to use for all.
|
43 |
+
`[2024-2-22]:` We sincerely thank [RoboFlow](https://roboflow.com/) and [@Skalskip92](https://twitter.com/skalskip92) for the [**Video Guide**](https://www.youtube.com/watch?v=X7gKBGVz4vs) about YOLO-World, nice work!
|
44 |
+
`[2024-2-18]:` We thank [@Skalskip92](https://twitter.com/skalskip92) for developing the wonderful segmentation demo via connecting YOLO-World and EfficientSAM. You can try it now at the [🤗 HuggingFace Spaces](https://huggingface.co/spaces/SkalskiP/YOLO-World).
|
45 |
+
`[2024-2-17]:` The largest model **X** of YOLO-World is released, which achieves better zero-shot performance!
|
46 |
+
`[2024-2-17]:` We release the code & models for **YOLO-World-Seg** now! YOLO-World now supports open-vocabulary / zero-shot object segmentation!
|
47 |
+
`[2024-2-15]:` The pre-traind YOLO-World-L with CC3M-Lite is released!
|
48 |
+
`[2024-2-14]:` We provide the [`image_demo`](demo.py) for inference on images or directories.
|
49 |
+
`[2024-2-10]:` We provide the [fine-tuning](./docs/finetuning.md) and [data](./docs/data.md) details for fine-tuning YOLO-World on the COCO dataset or the custom datasets!
|
50 |
+
`[2024-2-3]:` We support the `Gradio` demo now in the repo and you can build the YOLO-World demo on your own device!
|
51 |
+
`[2024-2-1]:` We've released the code and weights of YOLO-World now!
|
52 |
+
`[2024-2-1]:` We deploy the YOLO-World demo on [HuggingFace 🤗](https://huggingface.co/spaces/stevengrove/YOLO-World), you can try it now!
|
53 |
+
`[2024-1-31]:` We are excited to launch **YOLO-World**, a cutting-edge real-time open-vocabulary object detector.
|
54 |
+
|
55 |
+
|
56 |
+
## TODO
|
57 |
+
|
58 |
+
YOLO-World is under active development and please stay tuned ☕️!
|
59 |
+
If you have suggestions📃 or ideas💡,**we would love for you to bring them up in the [Roadmap](https://github.com/AILab-CVC/YOLO-World/issues/109)** ❤️!
|
60 |
+
> YOLO-World 目前正在积极开发中📃,如果你有建议或者想法💡,**我们非常希望您在 [Roadmap](https://github.com/AILab-CVC/YOLO-World/issues/109) 中提出来** ❤️!
|
61 |
+
|
62 |
+
## [FAQ (Frequently Asked Questions)](https://github.com/AILab-CVC/YOLO-World/discussions/149)
|
63 |
+
|
64 |
+
We have set up an FAQ about YOLO-World in the discussion on GitHub. We hope everyone can raise issues or solutions during use here, and we also hope that everyone can quickly find solutions from it.
|
65 |
+
|
66 |
+
> 我们在GitHub的discussion中建立了关于YOLO-World的常见问答,这里将收集一些常见问题,同时大家可以在此提出使用中的问题或者解决方案,也希望大家能够从中快速寻找到解决方案
|
67 |
+
|
68 |
+
|
69 |
+
## Highlights & Introduction
|
70 |
+
|
71 |
+
This repo contains the PyTorch implementation, pre-trained weights, and pre-training/fine-tuning code for YOLO-World.
|
72 |
+
|
73 |
+
* YOLO-World is pre-trained on large-scale datasets, including detection, grounding, and image-text datasets.
|
74 |
+
|
75 |
+
* YOLO-World is the next-generation YOLO detector, with a strong open-vocabulary detection capability and grounding ability.
|
76 |
+
|
77 |
+
* YOLO-World presents a *prompt-then-detect* paradigm for efficient user-vocabulary inference, which re-parameterizes vocabulary embeddings as parameters into the model and achieve superior inference speed. You can try to export your own detection model without extra training or fine-tuning in our [online demo](https://huggingface.co/spaces/stevengrove/YOLO-World)!
|
78 |
+
|
79 |
+
|
80 |
+
<center>
|
81 |
+
<img width=800px src="./assets/yolo_arch.png">
|
82 |
+
</center>
|
83 |
+
|
84 |
+
## Model Zoo
|
85 |
+
|
86 |
+
We've pre-trained YOLO-World-S/M/L from scratch and evaluate on the `LVIS val-1.0` and `LVIS minival`. We provide the pre-trained model weights and training logs for applications/research or re-producing the results.
|
87 |
+
|
88 |
+
### Zero-shot Inference on LVIS dataset
|
89 |
+
|
90 |
+
<div><font size=2>
|
91 |
+
|
92 |
+
| model | Pre-train Data | Size | AP<sup>mini</su> | AP<sub>r</sub> | AP<sub>c</sub> | AP<sub>f</sub> | AP<sup>val</su> | AP<sub>r</sub> | AP<sub>c</sub> | AP<sub>f</sub> | weights |
|
93 |
+
| :------------------------------------------------------------------------------------------------------------------- | :------------------- | :----------------- | :--------------: | :------------: | :------------: | :------------: | :-------------: | :------------: | :------------: | :------------: | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
|
94 |
+
| [YOLO-Worldv2-S](./configs/pretrain/yolo_world_v2_s_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG | 640 | 22.7 | 16.3 | 20.8 | 25.5 | 17.3 | 11.3 | 14.9 | 22.7 |[HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_s_obj365v1_goldg_pretrain-55b943ea.pth)|
|
95 |
+
| [YOLO-Worldv2-S](./configs/pretrain/yolo_world_v2_s_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py) | O365+GoldG | 1280🔸 | 24.1 | 18.7 | 22.0 | 26.9 | 18.8 | 14.1 | 16.3 | 23.8 |[HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_s_obj365v1_goldg_pretrain_1280ft-fc4ff4f7.pth)|
|
96 |
+
| [YOLO-Worldv2-M](./configs/pretrain/yolo_world_v2_m_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG | 640 | 30.0 | 25.0 | 27.2 | 33.4 | 23.5 | 17.1 | 20.0 | 30.1 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_m_obj365v1_goldg_pretrain-c6237d5b.pth)|
|
97 |
+
| [YOLO-Worldv2-M](./configs/pretrain/yolo_world_v2_m_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py) | O365+GoldG | 1280🔸 | 31.6 | 24.5 | 29.0 | 35.1 | 25.3 | 19.3 | 22.0 | 31.7 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_m_obj365v1_goldg_pretrain_1280ft-77d0346d.pth)|
|
98 |
+
| [YOLO-Worldv2-L](./configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG | 640 | 33.0 | 22.6 | 32.0 | 35.8 | 26.0 | 18.6 | 23.0 | 32.6 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_obj365v1_goldg_pretrain-a82b1fe3.pth)|
|
99 |
+
| [YOLO-Worldv2-L](./configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py) | O365+GoldG | 1280🔸 | 34.6 | 29.2 | 32.8 | 37.2 | 27.6 | 21.9 | 24.2 | 34.0 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_obj365v1_goldg_pretrain_1280ft-9babe3f6.pth)|
|
100 |
+
| [YOLO-Worldv2-L (CLIP-Large)](./configs/pretrain/yolo_world_v2_l_clip_large_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) 🔥 | O365+GoldG | 640 | 34.0 | 22.0 | 32.6 | 37.4 | 27.1 | 19.9 | 23.9 | 33.9 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_clip_large_o365v1_goldg_pretrain-8ff2e744.pth)|
|
101 |
+
| [YOLO-Worldv2-L (CLIP-Large)](./configs/pretrain/yolo_world_v2_l_clip_large_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_800ft_lvis_minival.py) 🔥 | O365+GoldG | 800🔸 | 35.5 | 28.3 | 33.2 | 38.8 | 28.6 | 22.0 | 25.1 | 35.4 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_clip_large_o365v1_goldg_pretrain_800ft-9df82e55.pth)|
|
102 |
+
| [YOLO-Worldv2-L](./configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG+CC3M-Lite | 640 | 32.9 | 25.3 | 31.1 | 35.8 | 26.1 | 20.6 | 22.6 | 32.3 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_obj365v1_goldg_cc3mlite_pretrain-ca93cd1f.pth)|
|
103 |
+
| [YOLO-Worldv2-X](./configs/pretrain/yolo_world_v2_x_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG+CC3M-Lite | 640 | 35.4 | 28.7 | 32.9 | 38.7 | 28.4 | 20.6 | 25.6 | 35.0 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_x_obj365v1_goldg_cc3mlite_pretrain-8698fbfa.pth) |
|
104 |
+
| [YOLO-Worldv2-XL](./configs/pretrain/yolo_world_v2_xl_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG+CC3M-Lite | 640 | 36.0 | 25.8 | 34.1 | 39.5 | 29.1 | 21.1 | 26.3 | 35.8 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_x_obj365v1_goldg_cc3mlite_pretrain-8698fbfa.pth) |
|
105 |
+
|
106 |
+
</font>
|
107 |
+
</div>
|
108 |
+
|
109 |
+
**NOTE:**
|
110 |
+
1. AP<sup>mini</sup>: evaluated on LVIS `minival`.
|
111 |
+
3. AP<sup>val</sup>: evaluated on LVIS `val 1.0`.
|
112 |
+
4. [HuggingFace Mirror](https://hf-mirror.com/) provides the mirror of HuggingFace, which is a choice for users who are unable to reach.
|
113 |
+
5. 🔸: fine-tuning models with the pre-trained data.
|
114 |
+
|
115 |
+
**Pre-training Logs:**
|
116 |
+
|
117 |
+
We provide the pre-training logs of `YOLO-World-v2`. Due to the unexpected errors of the local machines, the training might be interrupted several times.
|
118 |
+
|
119 |
+
| Model | YOLO-World-v2-S | YOLO-World-v2-M | YOLO-World-v2-L | YOLO-World-v2-X |
|
120 |
+
| :--- | :-------------: | :--------------: | :-------------: | :-------------: |
|
121 |
+
|Pre-training Log | [Part-1](https://drive.google.com/file/d/1oib7pKfA2h1U_5-85H_s0Nz8jWd0R-WP/view?usp=drive_link), [Part-2](https://drive.google.com/file/d/11cZ6OZy80VTvBlZy3kzLAHCxx5Iix5-n/view?usp=drive_link) | [Part-1](https://drive.google.com/file/d/1E6vYSS8kBipGc8oQnsjAfeUAx8I9yOX7/view?usp=drive_link), [Part-2](https://drive.google.com/file/d/1fbM7vt2tgSeB8o_7tUDofWvpPNSViNj5/view?usp=drive_link) | [Part-1](https://drive.google.com/file/d/1Tola1QGJZTL6nGy3SBxKuknfNfREDm8J/view?usp=drive_link), [Part-2](https://drive.google.com/file/d/1mTBXniioUb0CdctCG4ckIU6idGo0NnH8/view?usp=drive_link) | [Final part](https://drive.google.com/file/d/1aEUA_EPQbXOrpxHTQYB6ieGXudb1PLpd/view?usp=drive_link)|
|
122 |
+
|
123 |
+
|
124 |
+
## Getting started
|
125 |
+
|
126 |
+
### 1. Installation
|
127 |
+
|
128 |
+
YOLO-World is developed based on `torch==1.11.0` `mmyolo==0.6.0` and `mmdetection==3.0.0`.
|
129 |
+
|
130 |
+
#### Clone Project
|
131 |
+
|
132 |
+
```bash
|
133 |
+
git clone --recursive https://github.com/AILab-CVC/YOLO-World.git
|
134 |
+
```
|
135 |
+
#### Install
|
136 |
+
|
137 |
+
```bash
|
138 |
+
pip install torch wheel -q
|
139 |
+
pip install -e .
|
140 |
+
```
|
141 |
+
|
142 |
+
### 2. Preparing Data
|
143 |
+
|
144 |
+
We provide the details about the pre-training data in [docs/data](./docs/data.md).
|
145 |
+
|
146 |
+
|
147 |
+
## Training & Evaluation
|
148 |
+
|
149 |
+
We adopt the default [training](./tools/train.py) or [evaluation](./tools/test.py) scripts of [mmyolo](https://github.com/open-mmlab/mmyolo).
|
150 |
+
We provide the configs for pre-training and fine-tuning in `configs/pretrain` and `configs/finetune_coco`.
|
151 |
+
Training YOLO-World is easy:
|
152 |
+
|
153 |
+
```bash
|
154 |
+
chmod +x tools/dist_train.sh
|
155 |
+
# sample command for pre-training, use AMP for mixed-precision training
|
156 |
+
./tools/dist_train.sh configs/pretrain/yolo_world_l_t2i_bn_2e-4_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py 8 --amp
|
157 |
+
```
|
158 |
+
**NOTE:** YOLO-World is pre-trained on 4 nodes with 8 GPUs per node (32 GPUs in total). For pre-training, the `node_rank` and `nnodes` for multi-node training should be specified.
|
159 |
+
|
160 |
+
Evaluating YOLO-World is also easy:
|
161 |
+
|
162 |
+
```bash
|
163 |
+
chmod +x tools/dist_test.sh
|
164 |
+
./tools/dist_test.sh path/to/config path/to/weights 8
|
165 |
+
```
|
166 |
+
|
167 |
+
**NOTE:** We mainly evaluate the performance on LVIS-minival for pre-training.
|
168 |
+
|
169 |
+
## Fine-tuning YOLO-World
|
170 |
+
|
171 |
+
We provide the details about fine-tuning YOLO-World in [docs/fine-tuning](./docs/finetuning.md).
|
172 |
+
|
173 |
+
## Deployment
|
174 |
+
|
175 |
+
We provide the details about deployment for downstream applications in [docs/deployment](./docs/deploy.md).
|
176 |
+
You can directly download the ONNX model through the online [demo](https://huggingface.co/spaces/stevengrove/YOLO-World) in Huggingface Spaces 🤗.
|
177 |
+
|
178 |
+
## Demo
|
179 |
+
|
180 |
+
### Gradio Demo
|
181 |
+
|
182 |
+
We provide the [Gradio](https://www.gradio.app/) demo for local devices:
|
183 |
+
|
184 |
+
```bash
|
185 |
+
pip install gradio==4.16.0
|
186 |
+
python demo.py path/to/config path/to/weights
|
187 |
+
```
|
188 |
+
|
189 |
+
Additionaly, you can use a Dockerfile to build an image with gradio. As a prerequisite, make sure you have respective drivers installed alongside [nvidia-container-runtime](https://stackoverflow.com/questions/59691207/docker-build-with-nvidia-runtime). Replace MODEL_NAME and WEIGHT_NAME with the respective values or ommit this and use default values from the [Dockerfile](Dockerfile#3)
|
190 |
+
|
191 |
+
```bash
|
192 |
+
docker build --build-arg="MODEL=MODEL_NAME" --build-arg="WEIGHT=WEIGHT_NAME" -t yolo_demo .
|
193 |
+
docker run --runtime nvidia -p 8080:8080
|
194 |
+
```
|
195 |
+
|
196 |
+
### Image Demo
|
197 |
+
|
198 |
+
We provide a simple image demo for inference on images with visualization outputs.
|
199 |
+
|
200 |
+
```bash
|
201 |
+
python image_demo.py path/to/config path/to/weights image/path/directory 'person,dog,cat' --topk 100 --threshold 0.005 --output-dir demo_outputs
|
202 |
+
```
|
203 |
+
|
204 |
+
**Notes:**
|
205 |
+
* The `image` can be a directory or a single image.
|
206 |
+
* The `texts` can be a string of categories (noun phrases) which is separated by a comma. We also support `txt` file in which each line contains a category ( noun phrases).
|
207 |
+
* The `topk` and `threshold` control the number of predictions and the confidence threshold.
|
208 |
+
|
209 |
+
### Google Golab Notebook
|
210 |
+
|
211 |
+
We sincerely thank [Onuralp](https://github.com/onuralpszr) for sharing the [Colab Demo](https://colab.research.google.com/drive/1F_7S5lSaFM06irBCZqjhbN7MpUXo6WwO?usp=sharing), you can have a try 😊!
|
212 |
+
|
213 |
+
|
214 |
+
## Acknowledgement
|
215 |
+
|
216 |
+
We sincerely thank [mmyolo](https://github.com/open-mmlab/mmyolo), [mmdetection](https://github.com/open-mmlab/mmdetection), [GLIP](https://github.com/microsoft/GLIP), and [transformers](https://github.com/huggingface/transformers) for providing their wonderful code to the community!
|
217 |
+
|
218 |
+
## Citations
|
219 |
+
If you find YOLO-World is useful in your research or applications, please consider giving us a star 🌟 and citing it.
|
220 |
+
|
221 |
+
```bibtex
|
222 |
+
@inproceedings{Cheng2024YOLOWorld,
|
223 |
+
title={YOLO-World: Real-Time Open-Vocabulary Object Detection},
|
224 |
+
author={Cheng, Tianheng and Song, Lin and Ge, Yixiao and Liu, Wenyu and Wang, Xinggang and Shan, Ying},
|
225 |
+
booktitle={Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR)},
|
226 |
+
year={2024}
|
227 |
+
}
|
228 |
+
```
|
229 |
+
|
230 |
+
## Licence
|
231 |
+
YOLO-World is under the GPL-v3 Licence and is supported for comercial usage.
|
assets/yolo_arch.png
ADDED
![]() |
assets/yolo_logo.png
ADDED
![]() |
configs/finetune_coco/README.md
ADDED
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## Fine-tune YOLO-World on MS-COCO
|
2 |
+
|
3 |
+
|
4 |
+
### Updates
|
5 |
+
|
6 |
+
1. [2024-3-27]: Considering that fine-tuning YOLO-World on COCO **without `mask-refine`** obtains bad results, e.g., YOLO-World-L obtains 48.6 AP without `mask-refine` compared to 53.3 AP with `mask-refine`, we rethink the training process and explore new training schemes for fine-tuning without `mask-refine`.
|
7 |
+
BTW, the COCO fine-tuning results are updated with higher performance (with `mask-refine`)!
|
8 |
+
|
9 |
+
|
10 |
+
### COCO Results and Checkpoints
|
11 |
+
|
12 |
+
**NOTE:**
|
13 |
+
1. AP<sup>ZS</sup>: AP evaluated in the zero-shot setting (w/o fine-tuning on COCO dataset).
|
14 |
+
2. `mask-refine`: refine the box annotations with masks, and add `CopyPaste` augmentation during training.
|
15 |
+
|
16 |
+
| model | Schedule | `mask-refine` | efficient neck | AP<sup>ZS</sup>| AP | AP<sub>50</sub> | AP<sub>75</sub> | weights | log |
|
17 |
+
| :---- | :-------: | :----------: |:-------------: | :------------: | :-: | :--------------:| :-------------: |:------: | :-: |
|
18 |
+
| [YOLO-World-v2-S](./yolo_world_v2_s_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py) | AdamW, 2e-4, 80e | ✔️ | ✖️ | 37.5 | 46.1 | 62.0 | 49.9 | [HF Checkpoints](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_s_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_ep80-492dc329.pth) | [log](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_s_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_20240327_110411.log) |
|
19 |
+
| [YOLO-World-v2-M](./yolo_world_v2_m_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py) | AdamW, 2e-4, 80e | ✔️ | ✖️ | 42.8 | 51.0 | 67.5 | 55.2 | [HF Checkpoints](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_m_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_ep80-69c27ac7.pth) | [log](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_m_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_20240327_110411.log) |
|
20 |
+
| [YOLO-World-v2-L](./yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py) | AdamW, 2e-4, 80e | ✔️ | ✖️ | 45.1 | 53.9 | 70.9 | 58.8 | [HF Checkpoints](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_ep80-81c701ee.pth) | [log](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_20240326_160313.log) |
|
21 |
+
| [YOLO-World-v2-L](./yolo_world_v2_l_efficient_neck_2e-4_80e_8gpus_mask-refine_finetune_coco.py) | AdamW, 2e-4, 80e | ✔️ | ✔️ | 45.1 | | | | [HF Checkpoints]() | [log]() |
|
22 |
+
| [YOLO-World-v2-X](./yolo_world_v2_x_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py) | AdamW, 2e-4, 80e | ✔️ | ✖️ | 46.8 | 54.7 | 71.6 | 59.6 | [HF Checkpoints](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_x_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_ep80-76bc0cbd.pth) | [log](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_x_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_20240322_181232.log) |
|
23 |
+
| [YOLO-World-v2-L](./yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_finetune_coco.py) 🔥 | SGD, 1e-3, 40e | ✖️ | ✖️ | 45.1 | 52.8 | 69.5 | 57.8 | [HF Checkpoints](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_finetune_coco_ep80-e1288152.pth) | [log](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_finetuning_coco_20240327_014902.log) |
|
24 |
+
|
25 |
+
|
26 |
+
|
configs/finetune_coco/yolo_world_l_dual_vlpan_2e-4_80e_8gpus_finetune_coco.py
ADDED
@@ -0,0 +1,179 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = (
|
2 |
+
'../../third_party/mmyolo/configs/yolov8/'
|
3 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
4 |
+
custom_imports = dict(
|
5 |
+
imports=['yolo_world'],
|
6 |
+
allow_failed_imports=False)
|
7 |
+
|
8 |
+
# hyper-parameters
|
9 |
+
num_classes = 80
|
10 |
+
num_training_classes = 80
|
11 |
+
max_epochs = 80 # Maximum training epochs
|
12 |
+
close_mosaic_epochs = 10
|
13 |
+
save_epoch_intervals = 5
|
14 |
+
text_channels = 512
|
15 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
16 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
17 |
+
base_lr = 2e-4
|
18 |
+
weight_decay = 0.05
|
19 |
+
train_batch_size_per_gpu = 16
|
20 |
+
load_from='pretrained_models/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth'
|
21 |
+
persistent_workers = False
|
22 |
+
|
23 |
+
# model settings
|
24 |
+
model = dict(
|
25 |
+
type='YOLOWorldDetector',
|
26 |
+
mm_neck=True,
|
27 |
+
num_train_classes=num_training_classes,
|
28 |
+
num_test_classes=num_classes,
|
29 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
30 |
+
backbone=dict(
|
31 |
+
_delete_=True,
|
32 |
+
type='MultiModalYOLOBackbone',
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
text_model=dict(
|
35 |
+
type='HuggingCLIPLanguageBackbone',
|
36 |
+
model_name='openai/clip-vit-base-patch32',
|
37 |
+
frozen_modules=['all'])),
|
38 |
+
neck=dict(type='YOLOWorldDualPAFPN',
|
39 |
+
guide_channels=text_channels,
|
40 |
+
embed_channels=neck_embed_channels,
|
41 |
+
num_heads=neck_num_heads,
|
42 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
|
43 |
+
text_enhancder=dict(type='ImagePoolingAttentionModule',
|
44 |
+
embed_channels=256,
|
45 |
+
num_heads=8)),
|
46 |
+
bbox_head=dict(type='YOLOWorldHead',
|
47 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
48 |
+
embed_dims=text_channels,
|
49 |
+
num_classes=num_training_classes)),
|
50 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
51 |
+
|
52 |
+
# dataset settings
|
53 |
+
text_transform = [
|
54 |
+
dict(type='RandomLoadText',
|
55 |
+
num_neg_samples=(num_classes, num_classes),
|
56 |
+
max_num_samples=num_training_classes,
|
57 |
+
padding_to_max=True,
|
58 |
+
padding_value=''),
|
59 |
+
dict(type='mmdet.PackDetInputs',
|
60 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
61 |
+
'flip_direction', 'texts'))
|
62 |
+
]
|
63 |
+
mosaic_affine_transform = [
|
64 |
+
dict(
|
65 |
+
type='MultiModalMosaic',
|
66 |
+
img_scale=_base_.img_scale,
|
67 |
+
pad_val=114.0,
|
68 |
+
pre_transform=_base_.pre_transform),
|
69 |
+
dict(
|
70 |
+
type='YOLOv5RandomAffine',
|
71 |
+
max_rotate_degree=0.0,
|
72 |
+
max_shear_degree=0.0,
|
73 |
+
max_aspect_ratio=100.,
|
74 |
+
scaling_ratio_range=(1 - _base_.affine_scale,
|
75 |
+
1 + _base_.affine_scale),
|
76 |
+
# img_scale is (width, height)
|
77 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
78 |
+
border_val=(114, 114, 114))
|
79 |
+
]
|
80 |
+
train_pipeline = [
|
81 |
+
*_base_.pre_transform,
|
82 |
+
*mosaic_affine_transform,
|
83 |
+
dict(
|
84 |
+
type='YOLOv5MultiModalMixUp',
|
85 |
+
prob=_base_.mixup_prob,
|
86 |
+
pre_transform=[*_base_.pre_transform,
|
87 |
+
*mosaic_affine_transform]),
|
88 |
+
*_base_.last_transform[:-1],
|
89 |
+
*text_transform
|
90 |
+
]
|
91 |
+
train_pipeline_stage2 = [
|
92 |
+
*_base_.train_pipeline_stage2[:-1],
|
93 |
+
*text_transform
|
94 |
+
]
|
95 |
+
coco_train_dataset = dict(
|
96 |
+
_delete_=True,
|
97 |
+
type='MultiModalDataset',
|
98 |
+
dataset=dict(
|
99 |
+
type='YOLOv5CocoDataset',
|
100 |
+
data_root='data/coco',
|
101 |
+
ann_file='annotations/instances_train2017.json',
|
102 |
+
data_prefix=dict(img='train2017/'),
|
103 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
104 |
+
class_text_path='data/texts/coco_class_texts.json',
|
105 |
+
pipeline=train_pipeline)
|
106 |
+
|
107 |
+
train_dataloader = dict(
|
108 |
+
persistent_workers=persistent_workers,
|
109 |
+
batch_size=train_batch_size_per_gpu,
|
110 |
+
collate_fn=dict(type='yolow_collate'),
|
111 |
+
dataset=coco_train_dataset)
|
112 |
+
test_pipeline = [
|
113 |
+
*_base_.test_pipeline[:-1],
|
114 |
+
dict(type='LoadText'),
|
115 |
+
dict(
|
116 |
+
type='mmdet.PackDetInputs',
|
117 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
118 |
+
'scale_factor', 'pad_param', 'texts'))
|
119 |
+
]
|
120 |
+
coco_val_dataset = dict(
|
121 |
+
_delete_=True,
|
122 |
+
type='MultiModalDataset',
|
123 |
+
dataset=dict(
|
124 |
+
type='YOLOv5CocoDataset',
|
125 |
+
data_root='data/coco',
|
126 |
+
ann_file='annotations/instances_val2017.json',
|
127 |
+
data_prefix=dict(img='val2017/'),
|
128 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
129 |
+
class_text_path='data/texts/coco_class_texts.json',
|
130 |
+
pipeline=test_pipeline)
|
131 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
132 |
+
test_dataloader = val_dataloader
|
133 |
+
# training settings
|
134 |
+
default_hooks = dict(
|
135 |
+
param_scheduler=dict(
|
136 |
+
scheduler_type='linear',
|
137 |
+
lr_factor=0.01,
|
138 |
+
max_epochs=max_epochs),
|
139 |
+
checkpoint=dict(
|
140 |
+
max_keep_ckpts=-1,
|
141 |
+
save_best=None,
|
142 |
+
interval=save_epoch_intervals))
|
143 |
+
custom_hooks = [
|
144 |
+
dict(
|
145 |
+
type='EMAHook',
|
146 |
+
ema_type='ExpMomentumEMA',
|
147 |
+
momentum=0.0001,
|
148 |
+
update_buffers=True,
|
149 |
+
strict_load=False,
|
150 |
+
priority=49),
|
151 |
+
dict(
|
152 |
+
type='mmdet.PipelineSwitchHook',
|
153 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
154 |
+
switch_pipeline=train_pipeline_stage2)
|
155 |
+
]
|
156 |
+
train_cfg = dict(
|
157 |
+
max_epochs=max_epochs,
|
158 |
+
val_interval=5,
|
159 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
160 |
+
_base_.val_interval_stage2)])
|
161 |
+
optim_wrapper = dict(
|
162 |
+
optimizer=dict(
|
163 |
+
_delete_=True,
|
164 |
+
type='AdamW',
|
165 |
+
lr=base_lr,
|
166 |
+
weight_decay=weight_decay,
|
167 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
168 |
+
paramwise_cfg=dict(
|
169 |
+
custom_keys={'backbone.text_model': dict(lr_mult=0.01),
|
170 |
+
'logit_scale': dict(weight_decay=0.0)}),
|
171 |
+
constructor='YOLOWv5OptimizerConstructor')
|
172 |
+
|
173 |
+
# evaluation settings
|
174 |
+
val_evaluator = dict(
|
175 |
+
_delete_=True,
|
176 |
+
type='mmdet.CocoMetric',
|
177 |
+
proposal_nums=(100, 1, 10),
|
178 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
179 |
+
metric='bbox')
|
configs/finetune_coco/yolo_world_l_dual_vlpan_2e-4_80e_8gpus_mask-refine_finetune_coco.py
ADDED
@@ -0,0 +1,181 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = (
|
2 |
+
'../../third_party/mmyolo/configs/yolov8/'
|
3 |
+
'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
4 |
+
custom_imports = dict(
|
5 |
+
imports=['yolo_world'],
|
6 |
+
allow_failed_imports=False)
|
7 |
+
|
8 |
+
# hyper-parameters
|
9 |
+
num_classes = 80
|
10 |
+
num_training_classes = 80
|
11 |
+
max_epochs = 80 # Maximum training epochs
|
12 |
+
close_mosaic_epochs = 10
|
13 |
+
save_epoch_intervals = 5
|
14 |
+
text_channels = 512
|
15 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
16 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
17 |
+
base_lr = 2e-4
|
18 |
+
weight_decay = 0.05
|
19 |
+
train_batch_size_per_gpu = 16
|
20 |
+
load_from='pretrained_models/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth'
|
21 |
+
persistent_workers = False
|
22 |
+
|
23 |
+
# model settings
|
24 |
+
model = dict(
|
25 |
+
type='YOLOWorldDetector',
|
26 |
+
mm_neck=True,
|
27 |
+
num_train_classes=num_training_classes,
|
28 |
+
num_test_classes=num_classes,
|
29 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
30 |
+
backbone=dict(
|
31 |
+
_delete_=True,
|
32 |
+
type='MultiModalYOLOBackbone',
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
text_model=dict(
|
35 |
+
type='HuggingCLIPLanguageBackbone',
|
36 |
+
model_name='openai/clip-vit-base-patch32',
|
37 |
+
frozen_modules=['all'])),
|
38 |
+
neck=dict(type='YOLOWorldDualPAFPN',
|
39 |
+
guide_channels=text_channels,
|
40 |
+
embed_channels=neck_embed_channels,
|
41 |
+
num_heads=neck_num_heads,
|
42 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
|
43 |
+
text_enhancder=dict(type='ImagePoolingAttentionModule',
|
44 |
+
embed_channels=256,
|
45 |
+
num_heads=8)),
|
46 |
+
bbox_head=dict(type='YOLOWorldHead',
|
47 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
48 |
+
embed_dims=text_channels,
|
49 |
+
num_classes=num_training_classes)),
|
50 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
51 |
+
|
52 |
+
# dataset settings
|
53 |
+
text_transform = [
|
54 |
+
dict(type='RandomLoadText',
|
55 |
+
num_neg_samples=(num_classes, num_classes),
|
56 |
+
max_num_samples=num_training_classes,
|
57 |
+
padding_to_max=True,
|
58 |
+
padding_value=''),
|
59 |
+
dict(type='mmdet.PackDetInputs',
|
60 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
61 |
+
'flip_direction', 'texts'))
|
62 |
+
]
|
63 |
+
mosaic_affine_transform = [
|
64 |
+
dict(
|
65 |
+
type='MultiModalMosaic',
|
66 |
+
img_scale=_base_.img_scale,
|
67 |
+
pad_val=114.0,
|
68 |
+
pre_transform=_base_.pre_transform),
|
69 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
70 |
+
dict(
|
71 |
+
type='YOLOv5RandomAffine',
|
72 |
+
max_rotate_degree=0.0,
|
73 |
+
max_shear_degree=0.0,
|
74 |
+
max_aspect_ratio=100.,
|
75 |
+
scaling_ratio_range=(1 - _base_.affine_scale,
|
76 |
+
1 + _base_.affine_scale),
|
77 |
+
# img_scale is (width, height)
|
78 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
79 |
+
border_val=(114, 114, 114),
|
80 |
+
min_area_ratio=_base_.min_area_ratio,
|
81 |
+
use_mask_refine=_base_.use_mask2refine)
|
82 |
+
]
|
83 |
+
train_pipeline = [
|
84 |
+
*_base_.pre_transform,
|
85 |
+
*mosaic_affine_transform,
|
86 |
+
dict(
|
87 |
+
type='YOLOv5MultiModalMixUp',
|
88 |
+
prob=_base_.mixup_prob,
|
89 |
+
pre_transform=[*_base_.pre_transform,
|
90 |
+
*mosaic_affine_transform]),
|
91 |
+
*_base_.last_transform[:-1],
|
92 |
+
*text_transform
|
93 |
+
]
|
94 |
+
train_pipeline_stage2 = [
|
95 |
+
*_base_.train_pipeline_stage2[:-1],
|
96 |
+
*text_transform
|
97 |
+
]
|
98 |
+
coco_train_dataset = dict(
|
99 |
+
_delete_=True,
|
100 |
+
type='MultiModalDataset',
|
101 |
+
dataset=dict(
|
102 |
+
type='YOLOv5CocoDataset',
|
103 |
+
data_root='data/coco',
|
104 |
+
ann_file='annotations/instances_train2017.json',
|
105 |
+
data_prefix=dict(img='train2017/'),
|
106 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
107 |
+
class_text_path='data/texts/coco_class_texts.json',
|
108 |
+
pipeline=train_pipeline)
|
109 |
+
|
110 |
+
train_dataloader = dict(
|
111 |
+
persistent_workers=persistent_workers,
|
112 |
+
batch_size=train_batch_size_per_gpu,
|
113 |
+
collate_fn=dict(type='yolow_collate'),
|
114 |
+
dataset=coco_train_dataset)
|
115 |
+
test_pipeline = [
|
116 |
+
*_base_.test_pipeline[:-1],
|
117 |
+
dict(type='LoadText'),
|
118 |
+
dict(
|
119 |
+
type='mmdet.PackDetInputs',
|
120 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
121 |
+
'scale_factor', 'pad_param', 'texts'))
|
122 |
+
]
|
123 |
+
coco_val_dataset = dict(
|
124 |
+
_delete_=True,
|
125 |
+
type='MultiModalDataset',
|
126 |
+
dataset=dict(
|
127 |
+
type='YOLOv5CocoDataset',
|
128 |
+
data_root='data/coco',
|
129 |
+
ann_file='annotations/instances_val2017.json',
|
130 |
+
data_prefix=dict(img='val2017/'),
|
131 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
132 |
+
class_text_path='data/texts/coco_class_texts.json',
|
133 |
+
pipeline=test_pipeline)
|
134 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
135 |
+
test_dataloader = val_dataloader
|
136 |
+
# training settings
|
137 |
+
default_hooks = dict(
|
138 |
+
param_scheduler=dict(
|
139 |
+
scheduler_type='linear',
|
140 |
+
lr_factor=0.01,
|
141 |
+
max_epochs=max_epochs),
|
142 |
+
checkpoint=dict(
|
143 |
+
max_keep_ckpts=-1,
|
144 |
+
save_best=None,
|
145 |
+
interval=save_epoch_intervals))
|
146 |
+
custom_hooks = [
|
147 |
+
dict(
|
148 |
+
type='EMAHook',
|
149 |
+
ema_type='ExpMomentumEMA',
|
150 |
+
momentum=0.0001,
|
151 |
+
update_buffers=True,
|
152 |
+
strict_load=False,
|
153 |
+
priority=49),
|
154 |
+
dict(
|
155 |
+
type='mmdet.PipelineSwitchHook',
|
156 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
157 |
+
switch_pipeline=train_pipeline_stage2)
|
158 |
+
]
|
159 |
+
train_cfg = dict(
|
160 |
+
max_epochs=max_epochs,
|
161 |
+
val_interval=5,
|
162 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
163 |
+
_base_.val_interval_stage2)])
|
164 |
+
optim_wrapper = dict(
|
165 |
+
optimizer=dict(
|
166 |
+
_delete_=True,
|
167 |
+
type='AdamW',
|
168 |
+
lr=base_lr,
|
169 |
+
weight_decay=weight_decay,
|
170 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
171 |
+
paramwise_cfg=dict(
|
172 |
+
custom_keys={'backbone.text_model': dict(lr_mult=0.01),
|
173 |
+
'logit_scale': dict(weight_decay=0.0)}),
|
174 |
+
constructor='YOLOWv5OptimizerConstructor')
|
175 |
+
# evaluation settings
|
176 |
+
val_evaluator = dict(
|
177 |
+
_delete_=True,
|
178 |
+
type='mmdet.CocoMetric',
|
179 |
+
proposal_nums=(100, 1, 10),
|
180 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
181 |
+
metric='bbox')
|
configs/finetune_coco/yolo_world_l_efficient_neck_2e-4_80e_8gpus_mask-refine_finetune_coco.py
ADDED
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 2e-4
|
15 |
+
weight_decay = 0.05
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth'
|
18 |
+
# huggingface text model
|
19 |
+
text_model_name = 'openai/clip-vit-base-patch32'
|
20 |
+
persistent_workers = False
|
21 |
+
|
22 |
+
# model settings
|
23 |
+
model = dict(
|
24 |
+
type='YOLOWorldDetector',
|
25 |
+
mm_neck=True,
|
26 |
+
num_train_classes=num_training_classes,
|
27 |
+
num_test_classes=num_classes,
|
28 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
29 |
+
backbone=dict(
|
30 |
+
_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
image_model={{_base_.model.backbone}},
|
33 |
+
text_model=dict(
|
34 |
+
type='HuggingCLIPLanguageBackbone',
|
35 |
+
model_name=text_model_name,
|
36 |
+
frozen_modules=['all'])),
|
37 |
+
neck=dict(type='YOLOWorldPAFPN',
|
38 |
+
guide_channels=text_channels,
|
39 |
+
embed_channels=neck_embed_channels,
|
40 |
+
num_heads=neck_num_heads,
|
41 |
+
block_cfg=dict(type='EfficientCSPLayerWithTwoConv')),
|
42 |
+
bbox_head=dict(type='YOLOWorldHead',
|
43 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
44 |
+
embed_dims=text_channels,
|
45 |
+
num_classes=num_training_classes)),
|
46 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
47 |
+
|
48 |
+
# dataset settings
|
49 |
+
text_transform = [
|
50 |
+
dict(type='RandomLoadText',
|
51 |
+
num_neg_samples=(num_classes, num_classes),
|
52 |
+
max_num_samples=num_training_classes,
|
53 |
+
padding_to_max=True,
|
54 |
+
padding_value=''),
|
55 |
+
dict(type='mmdet.PackDetInputs',
|
56 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
57 |
+
'flip_direction', 'texts'))
|
58 |
+
]
|
59 |
+
mosaic_affine_transform = [
|
60 |
+
dict(type='MultiModalMosaic',
|
61 |
+
img_scale=_base_.img_scale,
|
62 |
+
pad_val=114.0,
|
63 |
+
pre_transform=_base_.pre_transform),
|
64 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
65 |
+
dict(
|
66 |
+
type='YOLOv5RandomAffine',
|
67 |
+
max_rotate_degree=0.0,
|
68 |
+
max_shear_degree=0.0,
|
69 |
+
max_aspect_ratio=100.,
|
70 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
71 |
+
# img_scale is (width, height)
|
72 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
73 |
+
border_val=(114, 114, 114),
|
74 |
+
min_area_ratio=_base_.min_area_ratio,
|
75 |
+
use_mask_refine=_base_.use_mask2refine)
|
76 |
+
]
|
77 |
+
train_pipeline = [
|
78 |
+
*_base_.pre_transform, *mosaic_affine_transform,
|
79 |
+
dict(type='YOLOv5MultiModalMixUp',
|
80 |
+
prob=_base_.mixup_prob,
|
81 |
+
pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
|
82 |
+
*_base_.last_transform[:-1], *text_transform
|
83 |
+
]
|
84 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
85 |
+
coco_train_dataset = dict(_delete_=True,
|
86 |
+
type='MultiModalDataset',
|
87 |
+
dataset=dict(
|
88 |
+
type='YOLOv5CocoDataset',
|
89 |
+
data_root='data/coco',
|
90 |
+
ann_file='annotations/instances_train2017.json',
|
91 |
+
data_prefix=dict(img='train2017/'),
|
92 |
+
filter_cfg=dict(filter_empty_gt=False,
|
93 |
+
min_size=32)),
|
94 |
+
class_text_path='data/texts/coco_class_texts.json',
|
95 |
+
pipeline=train_pipeline)
|
96 |
+
|
97 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
98 |
+
batch_size=train_batch_size_per_gpu,
|
99 |
+
collate_fn=dict(type='yolow_collate'),
|
100 |
+
dataset=coco_train_dataset)
|
101 |
+
test_pipeline = [
|
102 |
+
*_base_.test_pipeline[:-1],
|
103 |
+
dict(type='LoadText'),
|
104 |
+
dict(type='mmdet.PackDetInputs',
|
105 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
106 |
+
'scale_factor', 'pad_param', 'texts'))
|
107 |
+
]
|
108 |
+
coco_val_dataset = dict(
|
109 |
+
_delete_=True,
|
110 |
+
type='MultiModalDataset',
|
111 |
+
dataset=dict(type='YOLOv5CocoDataset',
|
112 |
+
data_root='data/coco',
|
113 |
+
ann_file='annotations/instances_val2017.json',
|
114 |
+
data_prefix=dict(img='val2017/'),
|
115 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
116 |
+
class_text_path='data/texts/coco_class_texts.json',
|
117 |
+
pipeline=test_pipeline)
|
118 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
119 |
+
test_dataloader = val_dataloader
|
120 |
+
# training settings
|
121 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
122 |
+
lr_factor=0.01,
|
123 |
+
max_epochs=max_epochs),
|
124 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
125 |
+
save_best=None,
|
126 |
+
interval=save_epoch_intervals))
|
127 |
+
custom_hooks = [
|
128 |
+
dict(type='EMAHook',
|
129 |
+
ema_type='ExpMomentumEMA',
|
130 |
+
momentum=0.0001,
|
131 |
+
update_buffers=True,
|
132 |
+
strict_load=False,
|
133 |
+
priority=49),
|
134 |
+
dict(type='mmdet.PipelineSwitchHook',
|
135 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
136 |
+
switch_pipeline=train_pipeline_stage2)
|
137 |
+
]
|
138 |
+
train_cfg = dict(max_epochs=max_epochs,
|
139 |
+
val_interval=5,
|
140 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
141 |
+
_base_.val_interval_stage2)])
|
142 |
+
optim_wrapper = dict(
|
143 |
+
optimizer=dict(
|
144 |
+
_delete_=True,
|
145 |
+
type='AdamW',
|
146 |
+
lr=base_lr,
|
147 |
+
weight_decay=weight_decay,
|
148 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
149 |
+
paramwise_cfg=dict(
|
150 |
+
custom_keys={'backbone.text_model': dict(lr_mult=0.01),
|
151 |
+
'logit_scale': dict(weight_decay=0.0)}),
|
152 |
+
constructor='YOLOWv5OptimizerConstructor')
|
153 |
+
|
154 |
+
# evaluation settings
|
155 |
+
val_evaluator = dict(_delete_=True,
|
156 |
+
type='mmdet.CocoMetric',
|
157 |
+
proposal_nums=(100, 1, 10),
|
158 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
159 |
+
metric='bbox')
|
configs/finetune_coco/yolo_world_v2_l_efficient_neck_2e-4_80e_8gpus_mask-refine_finetune_coco.py
ADDED
@@ -0,0 +1,182 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = (
|
2 |
+
'../../third_party/mmyolo/configs/yolov8/'
|
3 |
+
'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
4 |
+
custom_imports = dict(
|
5 |
+
imports=['yolo_world'],
|
6 |
+
allow_failed_imports=False)
|
7 |
+
|
8 |
+
# hyper-parameters
|
9 |
+
num_classes = 80
|
10 |
+
num_training_classes = 80
|
11 |
+
max_epochs = 80 # Maximum training epochs
|
12 |
+
close_mosaic_epochs = 10
|
13 |
+
save_epoch_intervals = 5
|
14 |
+
text_channels = 512
|
15 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
16 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
17 |
+
base_lr = 2e-4
|
18 |
+
weight_decay = 0.05
|
19 |
+
train_batch_size_per_gpu = 16
|
20 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
21 |
+
text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
22 |
+
# text_model_name = 'openai/clip-vit-base-patch32'
|
23 |
+
persistent_workers = False
|
24 |
+
|
25 |
+
# model settings
|
26 |
+
model = dict(
|
27 |
+
type='YOLOWorldDetector',
|
28 |
+
mm_neck=True,
|
29 |
+
num_train_classes=num_training_classes,
|
30 |
+
num_test_classes=num_classes,
|
31 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
32 |
+
backbone=dict(
|
33 |
+
_delete_=True,
|
34 |
+
type='MultiModalYOLOBackbone',
|
35 |
+
image_model={{_base_.model.backbone}},
|
36 |
+
text_model=dict(
|
37 |
+
type='HuggingCLIPLanguageBackbone',
|
38 |
+
model_name=text_model_name,
|
39 |
+
frozen_modules=['all'])),
|
40 |
+
neck=dict(type='YOLOWorldPAFPN',
|
41 |
+
guide_channels=text_channels,
|
42 |
+
embed_channels=neck_embed_channels,
|
43 |
+
num_heads=neck_num_heads,
|
44 |
+
block_cfg=dict(type='EfficientCSPLayerWithTwoConv')),
|
45 |
+
bbox_head=dict(type='YOLOWorldHead',
|
46 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
47 |
+
use_bn_head=True,
|
48 |
+
embed_dims=text_channels,
|
49 |
+
num_classes=num_training_classes)),
|
50 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
51 |
+
|
52 |
+
# dataset settings
|
53 |
+
text_transform = [
|
54 |
+
dict(type='RandomLoadText',
|
55 |
+
num_neg_samples=(num_classes, num_classes),
|
56 |
+
max_num_samples=num_training_classes,
|
57 |
+
padding_to_max=True,
|
58 |
+
padding_value=''),
|
59 |
+
dict(type='mmdet.PackDetInputs',
|
60 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
61 |
+
'flip_direction', 'texts'))
|
62 |
+
]
|
63 |
+
mosaic_affine_transform = [
|
64 |
+
dict(
|
65 |
+
type='MultiModalMosaic',
|
66 |
+
img_scale=_base_.img_scale,
|
67 |
+
pad_val=114.0,
|
68 |
+
pre_transform=_base_.pre_transform),
|
69 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
70 |
+
dict(
|
71 |
+
type='YOLOv5RandomAffine',
|
72 |
+
max_rotate_degree=0.0,
|
73 |
+
max_shear_degree=0.0,
|
74 |
+
max_aspect_ratio=100.,
|
75 |
+
scaling_ratio_range=(1 - _base_.affine_scale,
|
76 |
+
1 + _base_.affine_scale),
|
77 |
+
# img_scale is (width, height)
|
78 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
79 |
+
border_val=(114, 114, 114),
|
80 |
+
min_area_ratio=_base_.min_area_ratio,
|
81 |
+
use_mask_refine=_base_.use_mask2refine)
|
82 |
+
]
|
83 |
+
train_pipeline = [
|
84 |
+
*_base_.pre_transform,
|
85 |
+
*mosaic_affine_transform,
|
86 |
+
dict(
|
87 |
+
type='YOLOv5MultiModalMixUp',
|
88 |
+
prob=_base_.mixup_prob,
|
89 |
+
pre_transform=[*_base_.pre_transform,
|
90 |
+
*mosaic_affine_transform]),
|
91 |
+
*_base_.last_transform[:-1],
|
92 |
+
*text_transform
|
93 |
+
]
|
94 |
+
train_pipeline_stage2 = [
|
95 |
+
*_base_.train_pipeline_stage2[:-1],
|
96 |
+
*text_transform
|
97 |
+
]
|
98 |
+
coco_train_dataset = dict(
|
99 |
+
_delete_=True,
|
100 |
+
type='MultiModalDataset',
|
101 |
+
dataset=dict(
|
102 |
+
type='YOLOv5CocoDataset',
|
103 |
+
data_root='data/coco',
|
104 |
+
ann_file='annotations/instances_train2017.json',
|
105 |
+
data_prefix=dict(img='train2017/'),
|
106 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
107 |
+
class_text_path='data/texts/coco_class_texts.json',
|
108 |
+
pipeline=train_pipeline)
|
109 |
+
|
110 |
+
train_dataloader = dict(
|
111 |
+
persistent_workers=persistent_workers,
|
112 |
+
batch_size=train_batch_size_per_gpu,
|
113 |
+
collate_fn=dict(type='yolow_collate'),
|
114 |
+
dataset=coco_train_dataset)
|
115 |
+
test_pipeline = [
|
116 |
+
*_base_.test_pipeline[:-1],
|
117 |
+
dict(type='LoadText'),
|
118 |
+
dict(
|
119 |
+
type='mmdet.PackDetInputs',
|
120 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
121 |
+
'scale_factor', 'pad_param', 'texts'))
|
122 |
+
]
|
123 |
+
coco_val_dataset = dict(
|
124 |
+
_delete_=True,
|
125 |
+
type='MultiModalDataset',
|
126 |
+
dataset=dict(
|
127 |
+
type='YOLOv5CocoDataset',
|
128 |
+
data_root='data/coco',
|
129 |
+
ann_file='annotations/instances_val2017.json',
|
130 |
+
data_prefix=dict(img='val2017/'),
|
131 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
132 |
+
class_text_path='data/texts/coco_class_texts.json',
|
133 |
+
pipeline=test_pipeline)
|
134 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
135 |
+
test_dataloader = val_dataloader
|
136 |
+
# training settings
|
137 |
+
default_hooks = dict(
|
138 |
+
param_scheduler=dict(
|
139 |
+
scheduler_type='linear',
|
140 |
+
lr_factor=0.01,
|
141 |
+
max_epochs=max_epochs),
|
142 |
+
checkpoint=dict(
|
143 |
+
max_keep_ckpts=-1,
|
144 |
+
save_best=None,
|
145 |
+
interval=save_epoch_intervals))
|
146 |
+
custom_hooks = [
|
147 |
+
dict(
|
148 |
+
type='EMAHook',
|
149 |
+
ema_type='ExpMomentumEMA',
|
150 |
+
momentum=0.0001,
|
151 |
+
update_buffers=True,
|
152 |
+
strict_load=False,
|
153 |
+
priority=49),
|
154 |
+
dict(
|
155 |
+
type='mmdet.PipelineSwitchHook',
|
156 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
157 |
+
switch_pipeline=train_pipeline_stage2)
|
158 |
+
]
|
159 |
+
train_cfg = dict(
|
160 |
+
max_epochs=max_epochs,
|
161 |
+
val_interval=5,
|
162 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
163 |
+
_base_.val_interval_stage2)])
|
164 |
+
optim_wrapper = dict(
|
165 |
+
optimizer=dict(
|
166 |
+
_delete_=True,
|
167 |
+
type='AdamW',
|
168 |
+
lr=base_lr,
|
169 |
+
weight_decay=weight_decay,
|
170 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
171 |
+
paramwise_cfg=dict(
|
172 |
+
custom_keys={'backbone.text_model': dict(lr_mult=0.01),
|
173 |
+
'logit_scale': dict(weight_decay=0.0)}),
|
174 |
+
constructor='YOLOWv5OptimizerConstructor')
|
175 |
+
|
176 |
+
# evaluation settings
|
177 |
+
val_evaluator = dict(
|
178 |
+
_delete_=True,
|
179 |
+
type='mmdet.CocoMetric',
|
180 |
+
proposal_nums=(100, 1, 10),
|
181 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
182 |
+
metric='bbox')
|
configs/finetune_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py
ADDED
@@ -0,0 +1,181 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = (
|
2 |
+
'../../third_party/mmyolo/configs/yolov8/'
|
3 |
+
'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
4 |
+
custom_imports = dict(
|
5 |
+
imports=['yolo_world'],
|
6 |
+
allow_failed_imports=False)
|
7 |
+
|
8 |
+
# hyper-parameters
|
9 |
+
num_classes = 80
|
10 |
+
num_training_classes = 80
|
11 |
+
max_epochs = 80 # Maximum training epochs
|
12 |
+
close_mosaic_epochs = 10
|
13 |
+
save_epoch_intervals = 5
|
14 |
+
text_channels = 512
|
15 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
16 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
17 |
+
base_lr = 2e-4
|
18 |
+
weight_decay = 0.05
|
19 |
+
train_batch_size_per_gpu = 16
|
20 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
21 |
+
# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
22 |
+
text_model_name = 'openai/clip-vit-base-patch32'
|
23 |
+
persistent_workers = False
|
24 |
+
|
25 |
+
# model settings
|
26 |
+
model = dict(
|
27 |
+
type='YOLOWorldDetector',
|
28 |
+
mm_neck=True,
|
29 |
+
num_train_classes=num_training_classes,
|
30 |
+
num_test_classes=num_classes,
|
31 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
32 |
+
backbone=dict(
|
33 |
+
_delete_=True,
|
34 |
+
type='MultiModalYOLOBackbone',
|
35 |
+
image_model={{_base_.model.backbone}},
|
36 |
+
text_model=dict(
|
37 |
+
type='HuggingCLIPLanguageBackbone',
|
38 |
+
model_name=text_model_name,
|
39 |
+
frozen_modules=['all'])),
|
40 |
+
neck=dict(type='YOLOWorldPAFPN',
|
41 |
+
guide_channels=text_channels,
|
42 |
+
embed_channels=neck_embed_channels,
|
43 |
+
num_heads=neck_num_heads,
|
44 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
45 |
+
bbox_head=dict(type='YOLOWorldHead',
|
46 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
47 |
+
use_bn_head=True,
|
48 |
+
embed_dims=text_channels,
|
49 |
+
num_classes=num_training_classes)),
|
50 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
51 |
+
|
52 |
+
# dataset settings
|
53 |
+
text_transform = [
|
54 |
+
dict(type='RandomLoadText',
|
55 |
+
num_neg_samples=(num_classes, num_classes),
|
56 |
+
max_num_samples=num_training_classes,
|
57 |
+
padding_to_max=True,
|
58 |
+
padding_value=''),
|
59 |
+
dict(type='mmdet.PackDetInputs',
|
60 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
61 |
+
'flip_direction', 'texts'))
|
62 |
+
]
|
63 |
+
mosaic_affine_transform = [
|
64 |
+
dict(
|
65 |
+
type='MultiModalMosaic',
|
66 |
+
img_scale=_base_.img_scale,
|
67 |
+
pad_val=114.0,
|
68 |
+
pre_transform=_base_.pre_transform),
|
69 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
70 |
+
dict(
|
71 |
+
type='YOLOv5RandomAffine',
|
72 |
+
max_rotate_degree=0.0,
|
73 |
+
max_shear_degree=0.0,
|
74 |
+
max_aspect_ratio=100.,
|
75 |
+
scaling_ratio_range=(1 - _base_.affine_scale,
|
76 |
+
1 + _base_.affine_scale),
|
77 |
+
# img_scale is (width, height)
|
78 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
79 |
+
border_val=(114, 114, 114),
|
80 |
+
min_area_ratio=_base_.min_area_ratio,
|
81 |
+
use_mask_refine=_base_.use_mask2refine)
|
82 |
+
]
|
83 |
+
train_pipeline = [
|
84 |
+
*_base_.pre_transform,
|
85 |
+
*mosaic_affine_transform,
|
86 |
+
dict(
|
87 |
+
type='YOLOv5MultiModalMixUp',
|
88 |
+
prob=_base_.mixup_prob,
|
89 |
+
pre_transform=[*_base_.pre_transform,
|
90 |
+
*mosaic_affine_transform]),
|
91 |
+
*_base_.last_transform[:-1],
|
92 |
+
*text_transform
|
93 |
+
]
|
94 |
+
train_pipeline_stage2 = [
|
95 |
+
*_base_.train_pipeline_stage2[:-1],
|
96 |
+
*text_transform
|
97 |
+
]
|
98 |
+
coco_train_dataset = dict(
|
99 |
+
_delete_=True,
|
100 |
+
type='MultiModalDataset',
|
101 |
+
dataset=dict(
|
102 |
+
type='YOLOv5CocoDataset',
|
103 |
+
data_root='data/coco',
|
104 |
+
ann_file='annotations/instances_train2017.json',
|
105 |
+
data_prefix=dict(img='train2017/'),
|
106 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
107 |
+
class_text_path='data/texts/coco_class_texts.json',
|
108 |
+
pipeline=train_pipeline)
|
109 |
+
|
110 |
+
train_dataloader = dict(
|
111 |
+
persistent_workers=persistent_workers,
|
112 |
+
batch_size=train_batch_size_per_gpu,
|
113 |
+
collate_fn=dict(type='yolow_collate'),
|
114 |
+
dataset=coco_train_dataset)
|
115 |
+
test_pipeline = [
|
116 |
+
*_base_.test_pipeline[:-1],
|
117 |
+
dict(type='LoadText'),
|
118 |
+
dict(
|
119 |
+
type='mmdet.PackDetInputs',
|
120 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
121 |
+
'scale_factor', 'pad_param', 'texts'))
|
122 |
+
]
|
123 |
+
coco_val_dataset = dict(
|
124 |
+
_delete_=True,
|
125 |
+
type='MultiModalDataset',
|
126 |
+
dataset=dict(
|
127 |
+
type='YOLOv5CocoDataset',
|
128 |
+
data_root='data/coco',
|
129 |
+
ann_file='annotations/instances_val2017.json',
|
130 |
+
data_prefix=dict(img='val2017/'),
|
131 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
132 |
+
class_text_path='data/texts/coco_class_texts.json',
|
133 |
+
pipeline=test_pipeline)
|
134 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
135 |
+
test_dataloader = val_dataloader
|
136 |
+
# training settings
|
137 |
+
default_hooks = dict(
|
138 |
+
param_scheduler=dict(
|
139 |
+
scheduler_type='linear',
|
140 |
+
lr_factor=0.01,
|
141 |
+
max_epochs=max_epochs),
|
142 |
+
checkpoint=dict(
|
143 |
+
max_keep_ckpts=-1,
|
144 |
+
save_best=None,
|
145 |
+
interval=save_epoch_intervals))
|
146 |
+
custom_hooks = [
|
147 |
+
dict(
|
148 |
+
type='EMAHook',
|
149 |
+
ema_type='ExpMomentumEMA',
|
150 |
+
momentum=0.0001,
|
151 |
+
update_buffers=True,
|
152 |
+
strict_load=False,
|
153 |
+
priority=49),
|
154 |
+
dict(
|
155 |
+
type='mmdet.PipelineSwitchHook',
|
156 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
157 |
+
switch_pipeline=train_pipeline_stage2)
|
158 |
+
]
|
159 |
+
train_cfg = dict(
|
160 |
+
max_epochs=max_epochs,
|
161 |
+
val_interval=5,
|
162 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
163 |
+
_base_.val_interval_stage2)])
|
164 |
+
optim_wrapper = dict(
|
165 |
+
optimizer=dict(
|
166 |
+
_delete_=True,
|
167 |
+
type='AdamW',
|
168 |
+
lr=base_lr,
|
169 |
+
weight_decay=weight_decay,
|
170 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
171 |
+
paramwise_cfg=dict(
|
172 |
+
custom_keys={'backbone.text_model': dict(lr_mult=0.01),
|
173 |
+
'logit_scale': dict(weight_decay=0.0)}),
|
174 |
+
constructor='YOLOWv5OptimizerConstructor')
|
175 |
+
# evaluation settings
|
176 |
+
val_evaluator = dict(
|
177 |
+
_delete_=True,
|
178 |
+
type='mmdet.CocoMetric',
|
179 |
+
proposal_nums=(100, 1, 10),
|
180 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
181 |
+
metric='bbox')
|
configs/finetune_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_finetune_coco.py
ADDED
@@ -0,0 +1,160 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 40 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 30
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 1e-3
|
15 |
+
weight_decay = 0.0005
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
19 |
+
# text_model_name = 'openai/clip-vit-base-patch32'
|
20 |
+
persistent_workers = False
|
21 |
+
|
22 |
+
# model settings
|
23 |
+
model = dict(type='YOLOWorldDetector',
|
24 |
+
mm_neck=True,
|
25 |
+
num_train_classes=num_training_classes,
|
26 |
+
num_test_classes=num_classes,
|
27 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
28 |
+
backbone=dict(_delete_=True,
|
29 |
+
type='MultiModalYOLOBackbone',
|
30 |
+
image_model={{_base_.model.backbone}},
|
31 |
+
text_model=dict(type='HuggingCLIPLanguageBackbone',
|
32 |
+
model_name=text_model_name,
|
33 |
+
frozen_modules=['all'])),
|
34 |
+
neck=dict(type='YOLOWorldPAFPN',
|
35 |
+
guide_channels=text_channels,
|
36 |
+
embed_channels=neck_embed_channels,
|
37 |
+
num_heads=neck_num_heads,
|
38 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
39 |
+
bbox_head=dict(type='YOLOWorldHead',
|
40 |
+
head_module=dict(
|
41 |
+
type='YOLOWorldHeadModule',
|
42 |
+
use_bn_head=True,
|
43 |
+
embed_dims=text_channels,
|
44 |
+
num_classes=num_training_classes)),
|
45 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
46 |
+
|
47 |
+
# dataset settings
|
48 |
+
text_transform = [
|
49 |
+
dict(type='RandomLoadText',
|
50 |
+
num_neg_samples=(num_classes, num_classes),
|
51 |
+
max_num_samples=num_training_classes,
|
52 |
+
padding_to_max=True,
|
53 |
+
padding_value=''),
|
54 |
+
dict(type='mmdet.PackDetInputs',
|
55 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
56 |
+
'flip_direction', 'texts'))
|
57 |
+
]
|
58 |
+
mosaic_affine_transform = [
|
59 |
+
dict(type='MultiModalMosaic',
|
60 |
+
img_scale=_base_.img_scale,
|
61 |
+
pad_val=114.0,
|
62 |
+
pre_transform=_base_.pre_transform),
|
63 |
+
dict(
|
64 |
+
type='YOLOv5RandomAffine',
|
65 |
+
max_rotate_degree=0.0,
|
66 |
+
max_shear_degree=0.0,
|
67 |
+
max_aspect_ratio=100.,
|
68 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
69 |
+
# img_scale is (width, height)
|
70 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
71 |
+
border_val=(114, 114, 114))
|
72 |
+
]
|
73 |
+
|
74 |
+
train_pipeline = [
|
75 |
+
*_base_.pre_transform, *mosaic_affine_transform,
|
76 |
+
dict(type='YOLOv5MultiModalMixUp',
|
77 |
+
prob=_base_.mixup_prob,
|
78 |
+
pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
|
79 |
+
*_base_.last_transform[:-1], *text_transform
|
80 |
+
]
|
81 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
82 |
+
|
83 |
+
coco_train_dataset = dict(_delete_=True,
|
84 |
+
type='MultiModalDataset',
|
85 |
+
dataset=dict(
|
86 |
+
type='YOLOv5CocoDataset',
|
87 |
+
data_root='data/coco',
|
88 |
+
ann_file='annotations/instances_train2017.json',
|
89 |
+
data_prefix=dict(img='train2017/'),
|
90 |
+
filter_cfg=dict(filter_empty_gt=False,
|
91 |
+
min_size=32)),
|
92 |
+
class_text_path='data/texts/coco_class_texts.json',
|
93 |
+
pipeline=train_pipeline)
|
94 |
+
|
95 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
96 |
+
batch_size=train_batch_size_per_gpu,
|
97 |
+
collate_fn=dict(type='yolow_collate'),
|
98 |
+
dataset=coco_train_dataset)
|
99 |
+
test_pipeline = [
|
100 |
+
*_base_.test_pipeline[:-1],
|
101 |
+
dict(type='LoadText'),
|
102 |
+
dict(type='mmdet.PackDetInputs',
|
103 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
104 |
+
'scale_factor', 'pad_param', 'texts'))
|
105 |
+
]
|
106 |
+
coco_val_dataset = dict(
|
107 |
+
_delete_=True,
|
108 |
+
type='MultiModalDataset',
|
109 |
+
dataset=dict(type='YOLOv5CocoDataset',
|
110 |
+
data_root='data/coco',
|
111 |
+
ann_file='annotations/instances_val2017.json',
|
112 |
+
data_prefix=dict(img='val2017/'),
|
113 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
114 |
+
class_text_path='data/texts/coco_class_texts.json',
|
115 |
+
pipeline=test_pipeline)
|
116 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
117 |
+
test_dataloader = val_dataloader
|
118 |
+
# training settings
|
119 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
120 |
+
lr_factor=0.01,
|
121 |
+
max_epochs=max_epochs),
|
122 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
123 |
+
save_best=None,
|
124 |
+
interval=save_epoch_intervals))
|
125 |
+
custom_hooks = [
|
126 |
+
dict(type='EMAHook',
|
127 |
+
ema_type='ExpMomentumEMA',
|
128 |
+
momentum=0.0001,
|
129 |
+
update_buffers=True,
|
130 |
+
strict_load=False,
|
131 |
+
priority=49),
|
132 |
+
dict(type='mmdet.PipelineSwitchHook',
|
133 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
134 |
+
switch_pipeline=train_pipeline_stage2)
|
135 |
+
]
|
136 |
+
train_cfg = dict(max_epochs=max_epochs,
|
137 |
+
val_interval=5,
|
138 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
139 |
+
_base_.val_interval_stage2)])
|
140 |
+
optim_wrapper = dict(optimizer=dict(
|
141 |
+
_delete_=True,
|
142 |
+
type='SGD',
|
143 |
+
lr=base_lr,
|
144 |
+
momentum=0.937,
|
145 |
+
nesterov=True,
|
146 |
+
weight_decay=weight_decay,
|
147 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
148 |
+
paramwise_cfg=dict(
|
149 |
+
custom_keys={
|
150 |
+
'backbone.text_model': dict(lr_mult=0.01),
|
151 |
+
'logit_scale': dict(weight_decay=0.0)
|
152 |
+
}),
|
153 |
+
constructor='YOLOWv5OptimizerConstructor')
|
154 |
+
|
155 |
+
# evaluation settings
|
156 |
+
val_evaluator = dict(_delete_=True,
|
157 |
+
type='mmdet.CocoMetric',
|
158 |
+
proposal_nums=(100, 1, 10),
|
159 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
160 |
+
metric='bbox')
|
configs/finetune_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_80e_8gpus_mask-refine_finetune_coco.py
ADDED
@@ -0,0 +1,161 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 1e-3
|
15 |
+
weight_decay = 0.0005
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
19 |
+
# text_model_name = 'openai/clip-vit-base-patch32'
|
20 |
+
persistent_workers = False
|
21 |
+
|
22 |
+
# model settings
|
23 |
+
model = dict(type='YOLOWorldDetector',
|
24 |
+
mm_neck=True,
|
25 |
+
num_train_classes=num_training_classes,
|
26 |
+
num_test_classes=num_classes,
|
27 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
28 |
+
backbone=dict(_delete_=True,
|
29 |
+
type='MultiModalYOLOBackbone',
|
30 |
+
image_model={{_base_.model.backbone}},
|
31 |
+
text_model=dict(type='HuggingCLIPLanguageBackbone',
|
32 |
+
model_name=text_model_name,
|
33 |
+
frozen_modules=['all'])),
|
34 |
+
neck=dict(type='YOLOWorldPAFPN',
|
35 |
+
guide_channels=text_channels,
|
36 |
+
embed_channels=neck_embed_channels,
|
37 |
+
num_heads=neck_num_heads,
|
38 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
39 |
+
bbox_head=dict(type='YOLOWorldHead',
|
40 |
+
head_module=dict(
|
41 |
+
type='YOLOWorldHeadModule',
|
42 |
+
use_bn_head=True,
|
43 |
+
embed_dims=text_channels,
|
44 |
+
num_classes=num_training_classes)),
|
45 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
46 |
+
|
47 |
+
# dataset settings
|
48 |
+
text_transform = [
|
49 |
+
dict(type='RandomLoadText',
|
50 |
+
num_neg_samples=(num_classes, num_classes),
|
51 |
+
max_num_samples=num_training_classes,
|
52 |
+
padding_to_max=True,
|
53 |
+
padding_value=''),
|
54 |
+
dict(type='mmdet.PackDetInputs',
|
55 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
56 |
+
'flip_direction', 'texts'))
|
57 |
+
]
|
58 |
+
mosaic_affine_transform = [
|
59 |
+
dict(type='MultiModalMosaic',
|
60 |
+
img_scale=_base_.img_scale,
|
61 |
+
pad_val=114.0,
|
62 |
+
pre_transform=_base_.pre_transform),
|
63 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
64 |
+
dict(
|
65 |
+
type='YOLOv5RandomAffine',
|
66 |
+
max_rotate_degree=0.0,
|
67 |
+
max_shear_degree=0.0,
|
68 |
+
max_aspect_ratio=100.,
|
69 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
70 |
+
# img_scale is (width, height)
|
71 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
72 |
+
border_val=(114, 114, 114),
|
73 |
+
min_area_ratio=_base_.min_area_ratio,
|
74 |
+
use_mask_refine=_base_.use_mask2refine)
|
75 |
+
]
|
76 |
+
train_pipeline = [
|
77 |
+
*_base_.pre_transform, *mosaic_affine_transform,
|
78 |
+
dict(type='YOLOv5MultiModalMixUp',
|
79 |
+
prob=_base_.mixup_prob,
|
80 |
+
pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
|
81 |
+
*_base_.last_transform[:-1], *text_transform
|
82 |
+
]
|
83 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
84 |
+
coco_train_dataset = dict(_delete_=True,
|
85 |
+
type='MultiModalDataset',
|
86 |
+
dataset=dict(
|
87 |
+
type='YOLOv5CocoDataset',
|
88 |
+
data_root='data/coco',
|
89 |
+
ann_file='annotations/instances_train2017.json',
|
90 |
+
data_prefix=dict(img='train2017/'),
|
91 |
+
filter_cfg=dict(filter_empty_gt=False,
|
92 |
+
min_size=32)),
|
93 |
+
class_text_path='data/texts/coco_class_texts.json',
|
94 |
+
pipeline=train_pipeline)
|
95 |
+
|
96 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
97 |
+
batch_size=train_batch_size_per_gpu,
|
98 |
+
collate_fn=dict(type='yolow_collate'),
|
99 |
+
dataset=coco_train_dataset)
|
100 |
+
test_pipeline = [
|
101 |
+
*_base_.test_pipeline[:-1],
|
102 |
+
dict(type='LoadText'),
|
103 |
+
dict(type='mmdet.PackDetInputs',
|
104 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
105 |
+
'scale_factor', 'pad_param', 'texts'))
|
106 |
+
]
|
107 |
+
coco_val_dataset = dict(
|
108 |
+
_delete_=True,
|
109 |
+
type='MultiModalDataset',
|
110 |
+
dataset=dict(type='YOLOv5CocoDataset',
|
111 |
+
data_root='data/coco',
|
112 |
+
ann_file='annotations/instances_val2017.json',
|
113 |
+
data_prefix=dict(img='val2017/'),
|
114 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
115 |
+
class_text_path='data/texts/coco_class_texts.json',
|
116 |
+
pipeline=test_pipeline)
|
117 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
118 |
+
test_dataloader = val_dataloader
|
119 |
+
# training settings
|
120 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
121 |
+
lr_factor=0.01,
|
122 |
+
max_epochs=max_epochs),
|
123 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
124 |
+
save_best=None,
|
125 |
+
interval=save_epoch_intervals))
|
126 |
+
custom_hooks = [
|
127 |
+
dict(type='EMAHook',
|
128 |
+
ema_type='ExpMomentumEMA',
|
129 |
+
momentum=0.0001,
|
130 |
+
update_buffers=True,
|
131 |
+
strict_load=False,
|
132 |
+
priority=49),
|
133 |
+
dict(type='mmdet.PipelineSwitchHook',
|
134 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
135 |
+
switch_pipeline=train_pipeline_stage2)
|
136 |
+
]
|
137 |
+
train_cfg = dict(max_epochs=max_epochs,
|
138 |
+
val_interval=5,
|
139 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
140 |
+
_base_.val_interval_stage2)])
|
141 |
+
optim_wrapper = dict(optimizer=dict(
|
142 |
+
_delete_=True,
|
143 |
+
type='SGD',
|
144 |
+
lr=base_lr,
|
145 |
+
momentum=0.937,
|
146 |
+
nesterov=True,
|
147 |
+
weight_decay=weight_decay,
|
148 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
149 |
+
paramwise_cfg=dict(
|
150 |
+
custom_keys={
|
151 |
+
'backbone.text_model': dict(lr_mult=0.01),
|
152 |
+
'logit_scale': dict(weight_decay=0.0)
|
153 |
+
}),
|
154 |
+
constructor='YOLOWv5OptimizerConstructor')
|
155 |
+
|
156 |
+
# evaluation settings
|
157 |
+
val_evaluator = dict(_delete_=True,
|
158 |
+
type='mmdet.CocoMetric',
|
159 |
+
proposal_nums=(100, 1, 10),
|
160 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
161 |
+
metric='bbox')
|
configs/finetune_coco/yolo_world_v2_m_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py
ADDED
@@ -0,0 +1,182 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = (
|
2 |
+
'../../third_party/mmyolo/configs/yolov8/'
|
3 |
+
'yolov8_m_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
4 |
+
custom_imports = dict(
|
5 |
+
imports=['yolo_world'],
|
6 |
+
allow_failed_imports=False)
|
7 |
+
|
8 |
+
# hyper-parameters
|
9 |
+
num_classes = 80
|
10 |
+
num_training_classes = 80
|
11 |
+
max_epochs = 80 # Maximum training epochs
|
12 |
+
close_mosaic_epochs = 10
|
13 |
+
save_epoch_intervals = 5
|
14 |
+
text_channels = 512
|
15 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
16 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
17 |
+
base_lr = 2e-4
|
18 |
+
weight_decay = 0.05
|
19 |
+
train_batch_size_per_gpu = 16
|
20 |
+
load_from = 'pretrained_models/yolo_world_m_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_train-c6237d5b.pth'
|
21 |
+
# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
22 |
+
text_model_name = 'openai/clip-vit-base-patch32'
|
23 |
+
persistent_workers = False
|
24 |
+
|
25 |
+
# model settings
|
26 |
+
model = dict(
|
27 |
+
type='YOLOWorldDetector',
|
28 |
+
mm_neck=True,
|
29 |
+
num_train_classes=num_training_classes,
|
30 |
+
num_test_classes=num_classes,
|
31 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
32 |
+
backbone=dict(
|
33 |
+
_delete_=True,
|
34 |
+
type='MultiModalYOLOBackbone',
|
35 |
+
image_model={{_base_.model.backbone}},
|
36 |
+
text_model=dict(
|
37 |
+
type='HuggingCLIPLanguageBackbone',
|
38 |
+
model_name=text_model_name,
|
39 |
+
frozen_modules=['all'])),
|
40 |
+
neck=dict(type='YOLOWorldPAFPN',
|
41 |
+
guide_channels=text_channels,
|
42 |
+
embed_channels=neck_embed_channels,
|
43 |
+
num_heads=neck_num_heads,
|
44 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
45 |
+
bbox_head=dict(type='YOLOWorldHead',
|
46 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
47 |
+
use_bn_head=True,
|
48 |
+
embed_dims=text_channels,
|
49 |
+
num_classes=num_training_classes)),
|
50 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
51 |
+
|
52 |
+
# dataset settings
|
53 |
+
text_transform = [
|
54 |
+
dict(type='RandomLoadText',
|
55 |
+
num_neg_samples=(num_classes, num_classes),
|
56 |
+
max_num_samples=num_training_classes,
|
57 |
+
padding_to_max=True,
|
58 |
+
padding_value=''),
|
59 |
+
dict(type='mmdet.PackDetInputs',
|
60 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
61 |
+
'flip_direction', 'texts'))
|
62 |
+
]
|
63 |
+
mosaic_affine_transform = [
|
64 |
+
dict(
|
65 |
+
type='MultiModalMosaic',
|
66 |
+
img_scale=_base_.img_scale,
|
67 |
+
pad_val=114.0,
|
68 |
+
pre_transform=_base_.pre_transform),
|
69 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
70 |
+
dict(
|
71 |
+
type='YOLOv5RandomAffine',
|
72 |
+
max_rotate_degree=0.0,
|
73 |
+
max_shear_degree=0.0,
|
74 |
+
max_aspect_ratio=100.,
|
75 |
+
scaling_ratio_range=(1 - _base_.affine_scale,
|
76 |
+
1 + _base_.affine_scale),
|
77 |
+
# img_scale is (width, height)
|
78 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
79 |
+
border_val=(114, 114, 114),
|
80 |
+
min_area_ratio=_base_.min_area_ratio,
|
81 |
+
use_mask_refine=_base_.use_mask2refine)
|
82 |
+
]
|
83 |
+
train_pipeline = [
|
84 |
+
*_base_.pre_transform,
|
85 |
+
*mosaic_affine_transform,
|
86 |
+
dict(
|
87 |
+
type='YOLOv5MultiModalMixUp',
|
88 |
+
prob=_base_.mixup_prob,
|
89 |
+
pre_transform=[*_base_.pre_transform,
|
90 |
+
*mosaic_affine_transform]),
|
91 |
+
*_base_.last_transform[:-1],
|
92 |
+
*text_transform
|
93 |
+
]
|
94 |
+
train_pipeline_stage2 = [
|
95 |
+
*_base_.train_pipeline_stage2[:-1],
|
96 |
+
*text_transform
|
97 |
+
]
|
98 |
+
coco_train_dataset = dict(
|
99 |
+
_delete_=True,
|
100 |
+
type='MultiModalDataset',
|
101 |
+
dataset=dict(
|
102 |
+
type='YOLOv5CocoDataset',
|
103 |
+
data_root='data/coco',
|
104 |
+
ann_file='annotations/instances_train2017.json',
|
105 |
+
data_prefix=dict(img='train2017/'),
|
106 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
107 |
+
class_text_path='data/texts/coco_class_texts.json',
|
108 |
+
pipeline=train_pipeline)
|
109 |
+
|
110 |
+
train_dataloader = dict(
|
111 |
+
persistent_workers=persistent_workers,
|
112 |
+
batch_size=train_batch_size_per_gpu,
|
113 |
+
collate_fn=dict(type='yolow_collate'),
|
114 |
+
dataset=coco_train_dataset)
|
115 |
+
test_pipeline = [
|
116 |
+
*_base_.test_pipeline[:-1],
|
117 |
+
dict(type='LoadText'),
|
118 |
+
dict(
|
119 |
+
type='mmdet.PackDetInputs',
|
120 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
121 |
+
'scale_factor', 'pad_param', 'texts'))
|
122 |
+
]
|
123 |
+
coco_val_dataset = dict(
|
124 |
+
_delete_=True,
|
125 |
+
type='MultiModalDataset',
|
126 |
+
dataset=dict(
|
127 |
+
type='YOLOv5CocoDataset',
|
128 |
+
data_root='data/coco',
|
129 |
+
ann_file='annotations/instances_val2017.json',
|
130 |
+
data_prefix=dict(img='val2017/'),
|
131 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
132 |
+
class_text_path='data/texts/coco_class_texts.json',
|
133 |
+
pipeline=test_pipeline)
|
134 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
135 |
+
test_dataloader = val_dataloader
|
136 |
+
# training settings
|
137 |
+
default_hooks = dict(
|
138 |
+
param_scheduler=dict(
|
139 |
+
scheduler_type='linear',
|
140 |
+
lr_factor=0.01,
|
141 |
+
max_epochs=max_epochs),
|
142 |
+
checkpoint=dict(
|
143 |
+
max_keep_ckpts=-1,
|
144 |
+
save_best=None,
|
145 |
+
interval=save_epoch_intervals))
|
146 |
+
custom_hooks = [
|
147 |
+
dict(
|
148 |
+
type='EMAHook',
|
149 |
+
ema_type='ExpMomentumEMA',
|
150 |
+
momentum=0.0001,
|
151 |
+
update_buffers=True,
|
152 |
+
strict_load=False,
|
153 |
+
priority=49),
|
154 |
+
dict(
|
155 |
+
type='mmdet.PipelineSwitchHook',
|
156 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
157 |
+
switch_pipeline=train_pipeline_stage2)
|
158 |
+
]
|
159 |
+
train_cfg = dict(
|
160 |
+
max_epochs=max_epochs,
|
161 |
+
val_interval=5,
|
162 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
163 |
+
_base_.val_interval_stage2)])
|
164 |
+
optim_wrapper = dict(
|
165 |
+
optimizer=dict(
|
166 |
+
_delete_=True,
|
167 |
+
type='AdamW',
|
168 |
+
lr=base_lr,
|
169 |
+
weight_decay=weight_decay,
|
170 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
171 |
+
paramwise_cfg=dict(
|
172 |
+
custom_keys={'backbone.text_model': dict(lr_mult=0.01),
|
173 |
+
'logit_scale': dict(weight_decay=0.0)}),
|
174 |
+
constructor='YOLOWv5OptimizerConstructor')
|
175 |
+
|
176 |
+
# evaluation settings
|
177 |
+
val_evaluator = dict(
|
178 |
+
_delete_=True,
|
179 |
+
type='mmdet.CocoMetric',
|
180 |
+
proposal_nums=(100, 1, 10),
|
181 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
182 |
+
metric='bbox')
|
configs/finetune_coco/yolo_world_v2_s_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py
ADDED
@@ -0,0 +1,184 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = (
|
2 |
+
'../../third_party/mmyolo/configs/yolov8/'
|
3 |
+
'yolov8_s_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
4 |
+
custom_imports = dict(
|
5 |
+
imports=['yolo_world'],
|
6 |
+
allow_failed_imports=False)
|
7 |
+
|
8 |
+
# hyper-parameters
|
9 |
+
num_classes = 80
|
10 |
+
num_training_classes = 80
|
11 |
+
max_epochs = 80 # Maximum training epochs
|
12 |
+
close_mosaic_epochs = 10
|
13 |
+
save_epoch_intervals = 5
|
14 |
+
text_channels = 512
|
15 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
16 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
17 |
+
base_lr = 2e-4
|
18 |
+
weight_decay = 0.05
|
19 |
+
train_batch_size_per_gpu = 16
|
20 |
+
load_from = 'pretrained_models/yolo_world_s_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_train-55b943ea.pth'
|
21 |
+
# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
22 |
+
text_model_name = 'openai/clip-vit-base-patch32'
|
23 |
+
persistent_workers = False
|
24 |
+
mixup_prob = 0.15
|
25 |
+
copypaste_prob = 0.3
|
26 |
+
|
27 |
+
# model settings
|
28 |
+
model = dict(
|
29 |
+
type='YOLOWorldDetector',
|
30 |
+
mm_neck=True,
|
31 |
+
num_train_classes=num_training_classes,
|
32 |
+
num_test_classes=num_classes,
|
33 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
34 |
+
backbone=dict(
|
35 |
+
_delete_=True,
|
36 |
+
type='MultiModalYOLOBackbone',
|
37 |
+
image_model={{_base_.model.backbone}},
|
38 |
+
text_model=dict(
|
39 |
+
type='HuggingCLIPLanguageBackbone',
|
40 |
+
model_name=text_model_name,
|
41 |
+
frozen_modules=['all'])),
|
42 |
+
neck=dict(type='YOLOWorldPAFPN',
|
43 |
+
guide_channels=text_channels,
|
44 |
+
embed_channels=neck_embed_channels,
|
45 |
+
num_heads=neck_num_heads,
|
46 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
47 |
+
bbox_head=dict(type='YOLOWorldHead',
|
48 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
49 |
+
use_bn_head=True,
|
50 |
+
embed_dims=text_channels,
|
51 |
+
num_classes=num_training_classes)),
|
52 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
53 |
+
|
54 |
+
# dataset settings
|
55 |
+
text_transform = [
|
56 |
+
dict(type='RandomLoadText',
|
57 |
+
num_neg_samples=(num_classes, num_classes),
|
58 |
+
max_num_samples=num_training_classes,
|
59 |
+
padding_to_max=True,
|
60 |
+
padding_value=''),
|
61 |
+
dict(type='mmdet.PackDetInputs',
|
62 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
63 |
+
'flip_direction', 'texts'))
|
64 |
+
]
|
65 |
+
mosaic_affine_transform = [
|
66 |
+
dict(
|
67 |
+
type='MultiModalMosaic',
|
68 |
+
img_scale=_base_.img_scale,
|
69 |
+
pad_val=114.0,
|
70 |
+
pre_transform=_base_.pre_transform),
|
71 |
+
dict(type='YOLOv5CopyPaste', prob=copypaste_prob),
|
72 |
+
dict(
|
73 |
+
type='YOLOv5RandomAffine',
|
74 |
+
max_rotate_degree=0.0,
|
75 |
+
max_shear_degree=0.0,
|
76 |
+
max_aspect_ratio=100.,
|
77 |
+
scaling_ratio_range=(1 - _base_.affine_scale,
|
78 |
+
1 + _base_.affine_scale),
|
79 |
+
# img_scale is (width, height)
|
80 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
81 |
+
border_val=(114, 114, 114),
|
82 |
+
min_area_ratio=_base_.min_area_ratio,
|
83 |
+
use_mask_refine=_base_.use_mask2refine)
|
84 |
+
]
|
85 |
+
train_pipeline = [
|
86 |
+
*_base_.pre_transform,
|
87 |
+
*mosaic_affine_transform,
|
88 |
+
dict(
|
89 |
+
type='YOLOv5MultiModalMixUp',
|
90 |
+
prob=mixup_prob,
|
91 |
+
pre_transform=[*_base_.pre_transform,
|
92 |
+
*mosaic_affine_transform]),
|
93 |
+
*_base_.last_transform[:-1],
|
94 |
+
*text_transform
|
95 |
+
]
|
96 |
+
train_pipeline_stage2 = [
|
97 |
+
*_base_.train_pipeline_stage2[:-1],
|
98 |
+
*text_transform
|
99 |
+
]
|
100 |
+
coco_train_dataset = dict(
|
101 |
+
_delete_=True,
|
102 |
+
type='MultiModalDataset',
|
103 |
+
dataset=dict(
|
104 |
+
type='YOLOv5CocoDataset',
|
105 |
+
data_root='data/coco',
|
106 |
+
ann_file='annotations/instances_train2017.json',
|
107 |
+
data_prefix=dict(img='train2017/'),
|
108 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
109 |
+
class_text_path='data/texts/coco_class_texts.json',
|
110 |
+
pipeline=train_pipeline)
|
111 |
+
|
112 |
+
train_dataloader = dict(
|
113 |
+
persistent_workers=persistent_workers,
|
114 |
+
batch_size=train_batch_size_per_gpu,
|
115 |
+
collate_fn=dict(type='yolow_collate'),
|
116 |
+
dataset=coco_train_dataset)
|
117 |
+
test_pipeline = [
|
118 |
+
*_base_.test_pipeline[:-1],
|
119 |
+
dict(type='LoadText'),
|
120 |
+
dict(
|
121 |
+
type='mmdet.PackDetInputs',
|
122 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
123 |
+
'scale_factor', 'pad_param', 'texts'))
|
124 |
+
]
|
125 |
+
coco_val_dataset = dict(
|
126 |
+
_delete_=True,
|
127 |
+
type='MultiModalDataset',
|
128 |
+
dataset=dict(
|
129 |
+
type='YOLOv5CocoDataset',
|
130 |
+
data_root='data/coco',
|
131 |
+
ann_file='annotations/instances_val2017.json',
|
132 |
+
data_prefix=dict(img='val2017/'),
|
133 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
134 |
+
class_text_path='data/texts/coco_class_texts.json',
|
135 |
+
pipeline=test_pipeline)
|
136 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
137 |
+
test_dataloader = val_dataloader
|
138 |
+
# training settings
|
139 |
+
default_hooks = dict(
|
140 |
+
param_scheduler=dict(
|
141 |
+
scheduler_type='linear',
|
142 |
+
lr_factor=0.01,
|
143 |
+
max_epochs=max_epochs),
|
144 |
+
checkpoint=dict(
|
145 |
+
max_keep_ckpts=-1,
|
146 |
+
save_best=None,
|
147 |
+
interval=save_epoch_intervals))
|
148 |
+
custom_hooks = [
|
149 |
+
dict(
|
150 |
+
type='EMAHook',
|
151 |
+
ema_type='ExpMomentumEMA',
|
152 |
+
momentum=0.0001,
|
153 |
+
update_buffers=True,
|
154 |
+
strict_load=False,
|
155 |
+
priority=49),
|
156 |
+
dict(
|
157 |
+
type='mmdet.PipelineSwitchHook',
|
158 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
159 |
+
switch_pipeline=train_pipeline_stage2)
|
160 |
+
]
|
161 |
+
train_cfg = dict(
|
162 |
+
max_epochs=max_epochs,
|
163 |
+
val_interval=5,
|
164 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
165 |
+
_base_.val_interval_stage2)])
|
166 |
+
optim_wrapper = dict(
|
167 |
+
optimizer=dict(
|
168 |
+
_delete_=True,
|
169 |
+
type='AdamW',
|
170 |
+
lr=base_lr,
|
171 |
+
weight_decay=weight_decay,
|
172 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
173 |
+
paramwise_cfg=dict(
|
174 |
+
custom_keys={'backbone.text_model': dict(lr_mult=0.01),
|
175 |
+
'logit_scale': dict(weight_decay=0.0)}),
|
176 |
+
constructor='YOLOWv5OptimizerConstructor')
|
177 |
+
|
178 |
+
# evaluation settings
|
179 |
+
val_evaluator = dict(
|
180 |
+
_delete_=True,
|
181 |
+
type='mmdet.CocoMetric',
|
182 |
+
proposal_nums=(100, 1, 10),
|
183 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
184 |
+
metric='bbox')
|
configs/finetune_coco/yolo_world_v2_x_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py
ADDED
@@ -0,0 +1,183 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = (
|
2 |
+
'../../third_party/mmyolo/configs/yolov8/'
|
3 |
+
'yolov8_x_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
4 |
+
custom_imports = dict(
|
5 |
+
imports=['yolo_world'],
|
6 |
+
allow_failed_imports=False)
|
7 |
+
|
8 |
+
# hyper-parameters
|
9 |
+
num_classes = 80
|
10 |
+
num_training_classes = 80
|
11 |
+
max_epochs = 80 # Maximum training epochs
|
12 |
+
close_mosaic_epochs = 10
|
13 |
+
save_epoch_intervals = 5
|
14 |
+
text_channels = 512
|
15 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
16 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
17 |
+
base_lr = 2e-4
|
18 |
+
weight_decay = 0.05
|
19 |
+
train_batch_size_per_gpu = 16
|
20 |
+
load_from = 'pretrained_models/yolo_world_x_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc250k_train_lviseval-8698fbfa.pth'
|
21 |
+
text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
22 |
+
# text_model_name = 'openai/clip-vit-base-patch32'
|
23 |
+
persistent_workers = False
|
24 |
+
|
25 |
+
# model settings
|
26 |
+
model = dict(
|
27 |
+
type='YOLOWorldDetector',
|
28 |
+
mm_neck=True,
|
29 |
+
num_train_classes=num_training_classes,
|
30 |
+
num_test_classes=num_classes,
|
31 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
32 |
+
backbone=dict(
|
33 |
+
_delete_=True,
|
34 |
+
type='MultiModalYOLOBackbone',
|
35 |
+
image_model={{_base_.model.backbone}},
|
36 |
+
text_model=dict(
|
37 |
+
type='HuggingCLIPLanguageBackbone',
|
38 |
+
model_name=text_model_name,
|
39 |
+
frozen_modules=['all'])),
|
40 |
+
neck=dict(type='YOLOWorldPAFPN',
|
41 |
+
guide_channels=text_channels,
|
42 |
+
embed_channels=neck_embed_channels,
|
43 |
+
num_heads=neck_num_heads,
|
44 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
45 |
+
bbox_head=dict(type='YOLOWorldHead',
|
46 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
47 |
+
use_bn_head=True,
|
48 |
+
embed_dims=text_channels,
|
49 |
+
num_classes=num_training_classes)),
|
50 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
51 |
+
|
52 |
+
# dataset settings
|
53 |
+
text_transform = [
|
54 |
+
dict(type='RandomLoadText',
|
55 |
+
num_neg_samples=(num_classes, num_classes),
|
56 |
+
max_num_samples=num_training_classes,
|
57 |
+
padding_to_max=True,
|
58 |
+
padding_value=''),
|
59 |
+
dict(type='mmdet.PackDetInputs',
|
60 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
61 |
+
'flip_direction', 'texts'))
|
62 |
+
]
|
63 |
+
mosaic_affine_transform = [
|
64 |
+
dict(
|
65 |
+
type='MultiModalMosaic',
|
66 |
+
img_scale=_base_.img_scale,
|
67 |
+
pad_val=114.0,
|
68 |
+
pre_transform=_base_.pre_transform),
|
69 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
70 |
+
dict(
|
71 |
+
type='YOLOv5RandomAffine',
|
72 |
+
max_rotate_degree=0.0,
|
73 |
+
max_shear_degree=0.0,
|
74 |
+
max_aspect_ratio=100.,
|
75 |
+
scaling_ratio_range=(1 - _base_.affine_scale,
|
76 |
+
1 + _base_.affine_scale),
|
77 |
+
# img_scale is (width, height)
|
78 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
79 |
+
border_val=(114, 114, 114),
|
80 |
+
min_area_ratio=_base_.min_area_ratio,
|
81 |
+
use_mask_refine=_base_.use_mask2refine)
|
82 |
+
]
|
83 |
+
train_pipeline = [
|
84 |
+
*_base_.pre_transform,
|
85 |
+
*mosaic_affine_transform,
|
86 |
+
dict(
|
87 |
+
type='YOLOv5MultiModalMixUp',
|
88 |
+
prob=_base_.mixup_prob,
|
89 |
+
pre_transform=[*_base_.pre_transform,
|
90 |
+
*mosaic_affine_transform]),
|
91 |
+
*_base_.last_transform[:-1],
|
92 |
+
*text_transform
|
93 |
+
]
|
94 |
+
train_pipeline_stage2 = [
|
95 |
+
*_base_.train_pipeline_stage2[:-1],
|
96 |
+
*text_transform
|
97 |
+
]
|
98 |
+
coco_train_dataset = dict(
|
99 |
+
_delete_=True,
|
100 |
+
type='MultiModalDataset',
|
101 |
+
dataset=dict(
|
102 |
+
type='YOLOv5CocoDataset',
|
103 |
+
data_root='data/coco',
|
104 |
+
ann_file='annotations/instances_train2017.json',
|
105 |
+
data_prefix=dict(img='train2017/'),
|
106 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
107 |
+
class_text_path='data/texts/coco_class_texts.json',
|
108 |
+
pipeline=train_pipeline)
|
109 |
+
|
110 |
+
train_dataloader = dict(
|
111 |
+
persistent_workers=persistent_workers,
|
112 |
+
batch_size=train_batch_size_per_gpu,
|
113 |
+
collate_fn=dict(type='yolow_collate'),
|
114 |
+
dataset=coco_train_dataset)
|
115 |
+
|
116 |
+
test_pipeline = [
|
117 |
+
*_base_.test_pipeline[:-1],
|
118 |
+
dict(type='LoadText'),
|
119 |
+
dict(
|
120 |
+
type='mmdet.PackDetInputs',
|
121 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
122 |
+
'scale_factor', 'pad_param', 'texts'))
|
123 |
+
]
|
124 |
+
|
125 |
+
coco_val_dataset = dict(
|
126 |
+
_delete_=True,
|
127 |
+
type='MultiModalDataset',
|
128 |
+
dataset=dict(
|
129 |
+
type='YOLOv5CocoDataset',
|
130 |
+
data_root='data/coco',
|
131 |
+
ann_file='annotations/instances_val2017.json',
|
132 |
+
data_prefix=dict(img='val2017/'),
|
133 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
134 |
+
class_text_path='data/texts/coco_class_texts.json',
|
135 |
+
pipeline=test_pipeline)
|
136 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
137 |
+
test_dataloader = val_dataloader
|
138 |
+
# training settings
|
139 |
+
default_hooks = dict(
|
140 |
+
param_scheduler=dict(
|
141 |
+
scheduler_type='linear',
|
142 |
+
lr_factor=0.01,
|
143 |
+
max_epochs=max_epochs),
|
144 |
+
checkpoint=dict(
|
145 |
+
max_keep_ckpts=-1,
|
146 |
+
save_best=None,
|
147 |
+
interval=save_epoch_intervals))
|
148 |
+
custom_hooks = [
|
149 |
+
dict(
|
150 |
+
type='EMAHook',
|
151 |
+
ema_type='ExpMomentumEMA',
|
152 |
+
momentum=0.0001,
|
153 |
+
update_buffers=True,
|
154 |
+
strict_load=False,
|
155 |
+
priority=49),
|
156 |
+
dict(
|
157 |
+
type='mmdet.PipelineSwitchHook',
|
158 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
159 |
+
switch_pipeline=train_pipeline_stage2)
|
160 |
+
]
|
161 |
+
train_cfg = dict(
|
162 |
+
max_epochs=max_epochs,
|
163 |
+
val_interval=5,
|
164 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
165 |
+
_base_.val_interval_stage2)])
|
166 |
+
optim_wrapper = dict(
|
167 |
+
optimizer=dict(
|
168 |
+
_delete_=True,
|
169 |
+
type='AdamW',
|
170 |
+
lr=base_lr,
|
171 |
+
weight_decay=weight_decay,
|
172 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
173 |
+
paramwise_cfg=dict(
|
174 |
+
custom_keys={'backbone.text_model': dict(lr_mult=0.01),
|
175 |
+
'logit_scale': dict(weight_decay=0.0)}),
|
176 |
+
constructor='YOLOWv5OptimizerConstructor')
|
177 |
+
# evaluation settings
|
178 |
+
val_evaluator = dict(
|
179 |
+
_delete_=True,
|
180 |
+
type='mmdet.CocoMetric',
|
181 |
+
proposal_nums=(100, 1, 10),
|
182 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
183 |
+
metric='bbox')
|
configs/finetune_coco/yolo_world_v2_xl_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py
ADDED
@@ -0,0 +1,173 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_x_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 2e-4
|
15 |
+
weight_decay = 0.05
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
18 |
+
# text_model_name = 'openai/clip-vit-base-patch32'
|
19 |
+
persistent_workers = False
|
20 |
+
|
21 |
+
# scaling model from X to XL
|
22 |
+
deepen_factor = 1.0
|
23 |
+
widen_factor = 1.5
|
24 |
+
|
25 |
+
backbone = _base_.model.backbone
|
26 |
+
backbone.update(deepen_factor=deepen_factor, widen_factor=widen_factor)
|
27 |
+
|
28 |
+
# model settings
|
29 |
+
model = dict(type='YOLOWorldDetector',
|
30 |
+
mm_neck=True,
|
31 |
+
num_train_classes=num_training_classes,
|
32 |
+
num_test_classes=num_classes,
|
33 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
34 |
+
backbone=dict(_delete_=True,
|
35 |
+
type='MultiModalYOLOBackbone',
|
36 |
+
image_model=backbone,
|
37 |
+
text_model=dict(type='HuggingCLIPLanguageBackbone',
|
38 |
+
model_name=text_model_name,
|
39 |
+
frozen_modules=['all'])),
|
40 |
+
neck=dict(type='YOLOWorldPAFPN',
|
41 |
+
deepen_factor=deepen_factor,
|
42 |
+
widen_factor=widen_factor,
|
43 |
+
guide_channels=text_channels,
|
44 |
+
embed_channels=neck_embed_channels,
|
45 |
+
num_heads=neck_num_heads,
|
46 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
47 |
+
bbox_head=dict(type='YOLOWorldHead',
|
48 |
+
head_module=dict(
|
49 |
+
type='YOLOWorldHeadModule',
|
50 |
+
widen_factor=widen_factor,
|
51 |
+
use_bn_head=True,
|
52 |
+
embed_dims=text_channels,
|
53 |
+
num_classes=num_training_classes)),
|
54 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
55 |
+
|
56 |
+
# dataset settings
|
57 |
+
text_transform = [
|
58 |
+
dict(type='RandomLoadText',
|
59 |
+
num_neg_samples=(num_classes, num_classes),
|
60 |
+
max_num_samples=num_training_classes,
|
61 |
+
padding_to_max=True,
|
62 |
+
padding_value=''),
|
63 |
+
dict(type='mmdet.PackDetInputs',
|
64 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
65 |
+
'flip_direction', 'texts'))
|
66 |
+
]
|
67 |
+
mosaic_affine_transform = [
|
68 |
+
dict(type='MultiModalMosaic',
|
69 |
+
img_scale=_base_.img_scale,
|
70 |
+
pad_val=114.0,
|
71 |
+
pre_transform=_base_.pre_transform),
|
72 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
73 |
+
dict(
|
74 |
+
type='YOLOv5RandomAffine',
|
75 |
+
max_rotate_degree=0.0,
|
76 |
+
max_shear_degree=0.0,
|
77 |
+
max_aspect_ratio=100.,
|
78 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
79 |
+
# img_scale is (width, height)
|
80 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
81 |
+
border_val=(114, 114, 114),
|
82 |
+
min_area_ratio=_base_.min_area_ratio,
|
83 |
+
use_mask_refine=_base_.use_mask2refine)
|
84 |
+
]
|
85 |
+
train_pipeline = [
|
86 |
+
*_base_.pre_transform, *mosaic_affine_transform,
|
87 |
+
dict(type='YOLOv5MultiModalMixUp',
|
88 |
+
prob=_base_.mixup_prob,
|
89 |
+
pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
|
90 |
+
*_base_.last_transform[:-1], *text_transform
|
91 |
+
]
|
92 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
93 |
+
coco_train_dataset = dict(_delete_=True,
|
94 |
+
type='MultiModalDataset',
|
95 |
+
dataset=dict(
|
96 |
+
type='YOLOv5CocoDataset',
|
97 |
+
data_root='data/coco',
|
98 |
+
ann_file='annotations/instances_train2017.json',
|
99 |
+
data_prefix=dict(img='train2017/'),
|
100 |
+
filter_cfg=dict(filter_empty_gt=False,
|
101 |
+
min_size=32)),
|
102 |
+
class_text_path='data/texts/coco_class_texts.json',
|
103 |
+
pipeline=train_pipeline)
|
104 |
+
|
105 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
106 |
+
batch_size=train_batch_size_per_gpu,
|
107 |
+
collate_fn=dict(type='yolow_collate'),
|
108 |
+
dataset=coco_train_dataset)
|
109 |
+
|
110 |
+
test_pipeline = [
|
111 |
+
*_base_.test_pipeline[:-1],
|
112 |
+
dict(type='LoadText'),
|
113 |
+
dict(type='mmdet.PackDetInputs',
|
114 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
115 |
+
'scale_factor', 'pad_param', 'texts'))
|
116 |
+
]
|
117 |
+
|
118 |
+
coco_val_dataset = dict(
|
119 |
+
_delete_=True,
|
120 |
+
type='MultiModalDataset',
|
121 |
+
dataset=dict(type='YOLOv5CocoDataset',
|
122 |
+
data_root='data/coco',
|
123 |
+
ann_file='annotations/instances_val2017.json',
|
124 |
+
data_prefix=dict(img='val2017/'),
|
125 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
126 |
+
class_text_path='data/texts/coco_class_texts.json',
|
127 |
+
pipeline=test_pipeline)
|
128 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
129 |
+
test_dataloader = val_dataloader
|
130 |
+
# training settings
|
131 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
132 |
+
lr_factor=0.01,
|
133 |
+
max_epochs=max_epochs),
|
134 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
135 |
+
save_best=None,
|
136 |
+
interval=save_epoch_intervals))
|
137 |
+
custom_hooks = [
|
138 |
+
dict(type='EMAHook',
|
139 |
+
ema_type='ExpMomentumEMA',
|
140 |
+
momentum=0.0001,
|
141 |
+
update_buffers=True,
|
142 |
+
strict_load=False,
|
143 |
+
priority=49),
|
144 |
+
dict(type='mmdet.PipelineSwitchHook',
|
145 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
146 |
+
switch_pipeline=train_pipeline_stage2)
|
147 |
+
]
|
148 |
+
train_cfg = dict(max_epochs=max_epochs,
|
149 |
+
val_interval=5,
|
150 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
151 |
+
_base_.val_interval_stage2)])
|
152 |
+
optim_wrapper = dict(optimizer=dict(
|
153 |
+
_delete_=True,
|
154 |
+
type='AdamW',
|
155 |
+
lr=base_lr,
|
156 |
+
weight_decay=weight_decay,
|
157 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
158 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
159 |
+
norm_decay_mult=0.0,
|
160 |
+
custom_keys={
|
161 |
+
'backbone.text_model':
|
162 |
+
dict(lr_mult=0.01),
|
163 |
+
'logit_scale':
|
164 |
+
dict(weight_decay=0.0)
|
165 |
+
}),
|
166 |
+
constructor='YOLOWv5OptimizerConstructor')
|
167 |
+
|
168 |
+
# evaluation settings
|
169 |
+
val_evaluator = dict(_delete_=True,
|
170 |
+
type='mmdet.CocoMetric',
|
171 |
+
proposal_nums=(100, 1, 10),
|
172 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
173 |
+
metric='bbox')
|
configs/pretrain/yolo_world_v2_l_clip_large_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_800ft_lvis_minival.py
ADDED
@@ -0,0 +1,200 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 768
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.0125
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
# text_model_name = '../pretrained_models/clip-vit-large-patch14-336'
|
19 |
+
text_model_name = 'openai/clip-vit-large-patch14-336'
|
20 |
+
img_scale = (800, 800)
|
21 |
+
|
22 |
+
# model settings
|
23 |
+
model = dict(
|
24 |
+
type='YOLOWorldDetector',
|
25 |
+
mm_neck=True,
|
26 |
+
num_train_classes=num_training_classes,
|
27 |
+
num_test_classes=num_classes,
|
28 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
29 |
+
backbone=dict(
|
30 |
+
_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
image_model={{_base_.model.backbone}},
|
33 |
+
text_model=dict(
|
34 |
+
type='HuggingCLIPLanguageBackbone',
|
35 |
+
model_name=text_model_name,
|
36 |
+
frozen_modules=['all'])),
|
37 |
+
neck=dict(type='YOLOWorldPAFPN',
|
38 |
+
guide_channels=text_channels,
|
39 |
+
embed_channels=neck_embed_channels,
|
40 |
+
num_heads=neck_num_heads,
|
41 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
42 |
+
bbox_head=dict(type='YOLOWorldHead',
|
43 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
44 |
+
use_bn_head=True,
|
45 |
+
embed_dims=text_channels,
|
46 |
+
num_classes=num_training_classes)),
|
47 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
48 |
+
|
49 |
+
# dataset settings
|
50 |
+
text_transform = [
|
51 |
+
dict(type='RandomLoadText',
|
52 |
+
num_neg_samples=(num_classes, num_classes),
|
53 |
+
max_num_samples=num_training_classes,
|
54 |
+
padding_to_max=True,
|
55 |
+
padding_value=''),
|
56 |
+
dict(type='mmdet.PackDetInputs',
|
57 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
58 |
+
'flip_direction', 'texts'))
|
59 |
+
]
|
60 |
+
train_pipeline = [
|
61 |
+
*_base_.pre_transform,
|
62 |
+
dict(type='MultiModalMosaic',
|
63 |
+
img_scale=img_scale,
|
64 |
+
pad_val=114.0,
|
65 |
+
pre_transform=_base_.pre_transform),
|
66 |
+
dict(
|
67 |
+
type='YOLOv5RandomAffine',
|
68 |
+
max_rotate_degree=0.0,
|
69 |
+
max_shear_degree=0.0,
|
70 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
71 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
72 |
+
border=(-img_scale[0] // 2, -img_scale[1] // 2),
|
73 |
+
border_val=(114, 114, 114)),
|
74 |
+
*_base_.last_transform[:-1],
|
75 |
+
*text_transform,
|
76 |
+
]
|
77 |
+
|
78 |
+
train_pipeline_stage2 = [
|
79 |
+
*_base_.pre_transform,
|
80 |
+
dict(type='YOLOv5KeepRatioResize', scale=img_scale),
|
81 |
+
dict(
|
82 |
+
type='LetterResize',
|
83 |
+
scale=img_scale,
|
84 |
+
allow_scale_up=True,
|
85 |
+
pad_val=dict(img=114.0)),
|
86 |
+
dict(
|
87 |
+
type='YOLOv5RandomAffine',
|
88 |
+
max_rotate_degree=0.0,
|
89 |
+
max_shear_degree=0.0,
|
90 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
91 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
92 |
+
border_val=(114, 114, 114)),
|
93 |
+
*_base_.last_transform[:-1],
|
94 |
+
*text_transform
|
95 |
+
]
|
96 |
+
|
97 |
+
obj365v1_train_dataset = dict(
|
98 |
+
type='MultiModalDataset',
|
99 |
+
dataset=dict(
|
100 |
+
type='YOLOv5Objects365V1Dataset',
|
101 |
+
data_root='data/objects365v1/',
|
102 |
+
ann_file='annotations/objects365_train.json',
|
103 |
+
data_prefix=dict(img='train/'),
|
104 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
105 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
106 |
+
pipeline=train_pipeline)
|
107 |
+
|
108 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
109 |
+
data_root='data/mixed_grounding/',
|
110 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
111 |
+
data_prefix=dict(img='gqa/images/'),
|
112 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
113 |
+
pipeline=train_pipeline)
|
114 |
+
|
115 |
+
flickr_train_dataset = dict(
|
116 |
+
type='YOLOv5MixedGroundingDataset',
|
117 |
+
data_root='data/flickr/',
|
118 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
119 |
+
data_prefix=dict(img='full_images/'),
|
120 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
121 |
+
pipeline=train_pipeline)
|
122 |
+
|
123 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
124 |
+
collate_fn=dict(type='yolow_collate'),
|
125 |
+
dataset=dict(_delete_=True,
|
126 |
+
type='ConcatDataset',
|
127 |
+
datasets=[
|
128 |
+
obj365v1_train_dataset,
|
129 |
+
flickr_train_dataset, mg_train_dataset
|
130 |
+
],
|
131 |
+
ignore_keys=['classes', 'palette']))
|
132 |
+
|
133 |
+
test_pipeline = [
|
134 |
+
dict(type='LoadImageFromFile'),
|
135 |
+
dict(type='YOLOv5KeepRatioResize', scale=img_scale),
|
136 |
+
dict(
|
137 |
+
type='LetterResize',
|
138 |
+
scale=img_scale,
|
139 |
+
allow_scale_up=False,
|
140 |
+
pad_val=dict(img=114)),
|
141 |
+
dict(type='LoadAnnotations', with_bbox=True, _scope_='mmdet'),
|
142 |
+
dict(type='LoadText'),
|
143 |
+
dict(type='mmdet.PackDetInputs',
|
144 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
145 |
+
'scale_factor', 'pad_param', 'texts'))
|
146 |
+
]
|
147 |
+
|
148 |
+
coco_val_dataset = dict(
|
149 |
+
_delete_=True,
|
150 |
+
type='MultiModalDataset',
|
151 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
152 |
+
data_root='data/coco/',
|
153 |
+
test_mode=True,
|
154 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
155 |
+
data_prefix=dict(img=''),
|
156 |
+
batch_shapes_cfg=None),
|
157 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
158 |
+
pipeline=test_pipeline)
|
159 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
160 |
+
test_dataloader = val_dataloader
|
161 |
+
|
162 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
163 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
164 |
+
metric='bbox')
|
165 |
+
test_evaluator = val_evaluator
|
166 |
+
|
167 |
+
# training settings
|
168 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
169 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
170 |
+
rule='greater'))
|
171 |
+
custom_hooks = [
|
172 |
+
dict(type='EMAHook',
|
173 |
+
ema_type='ExpMomentumEMA',
|
174 |
+
momentum=0.0001,
|
175 |
+
update_buffers=True,
|
176 |
+
strict_load=False,
|
177 |
+
priority=49),
|
178 |
+
dict(type='mmdet.PipelineSwitchHook',
|
179 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
180 |
+
switch_pipeline=train_pipeline_stage2)
|
181 |
+
]
|
182 |
+
train_cfg = dict(max_epochs=max_epochs,
|
183 |
+
val_interval=10,
|
184 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
185 |
+
_base_.val_interval_stage2)])
|
186 |
+
optim_wrapper = dict(optimizer=dict(
|
187 |
+
_delete_=True,
|
188 |
+
type='AdamW',
|
189 |
+
lr=base_lr,
|
190 |
+
weight_decay=weight_decay,
|
191 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
192 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
193 |
+
norm_decay_mult=0.0,
|
194 |
+
custom_keys={
|
195 |
+
'backbone.text_model':
|
196 |
+
dict(lr_mult=0.01),
|
197 |
+
'logit_scale':
|
198 |
+
dict(weight_decay=0.0)
|
199 |
+
}),
|
200 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_l_clip_large_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 768
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.0125
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
# text_model_name = '../pretrained_models/clip-vit-large-patch14-336'
|
19 |
+
text_model_name = 'openai/clip-vit-large-patch14-336'
|
20 |
+
# model settings
|
21 |
+
model = dict(
|
22 |
+
type='YOLOWorldDetector',
|
23 |
+
mm_neck=True,
|
24 |
+
num_train_classes=num_training_classes,
|
25 |
+
num_test_classes=num_classes,
|
26 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
27 |
+
backbone=dict(
|
28 |
+
_delete_=True,
|
29 |
+
type='MultiModalYOLOBackbone',
|
30 |
+
image_model={{_base_.model.backbone}},
|
31 |
+
text_model=dict(
|
32 |
+
type='HuggingCLIPLanguageBackbone',
|
33 |
+
model_name=text_model_name,
|
34 |
+
frozen_modules=['all'])),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
guide_channels=text_channels,
|
37 |
+
embed_channels=neck_embed_channels,
|
38 |
+
num_heads=neck_num_heads,
|
39 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
40 |
+
bbox_head=dict(type='YOLOWorldHead',
|
41 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
42 |
+
use_bn_head=True,
|
43 |
+
embed_dims=text_channels,
|
44 |
+
num_classes=num_training_classes)),
|
45 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
46 |
+
|
47 |
+
# dataset settings
|
48 |
+
text_transform = [
|
49 |
+
dict(type='RandomLoadText',
|
50 |
+
num_neg_samples=(num_classes, num_classes),
|
51 |
+
max_num_samples=num_training_classes,
|
52 |
+
padding_to_max=True,
|
53 |
+
padding_value=''),
|
54 |
+
dict(type='mmdet.PackDetInputs',
|
55 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
56 |
+
'flip_direction', 'texts'))
|
57 |
+
]
|
58 |
+
train_pipeline = [
|
59 |
+
*_base_.pre_transform,
|
60 |
+
dict(type='MultiModalMosaic',
|
61 |
+
img_scale=_base_.img_scale,
|
62 |
+
pad_val=114.0,
|
63 |
+
pre_transform=_base_.pre_transform),
|
64 |
+
dict(
|
65 |
+
type='YOLOv5RandomAffine',
|
66 |
+
max_rotate_degree=0.0,
|
67 |
+
max_shear_degree=0.0,
|
68 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
69 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
70 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
71 |
+
border_val=(114, 114, 114)),
|
72 |
+
*_base_.last_transform[:-1],
|
73 |
+
*text_transform,
|
74 |
+
]
|
75 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
76 |
+
obj365v1_train_dataset = dict(
|
77 |
+
type='MultiModalDataset',
|
78 |
+
dataset=dict(
|
79 |
+
type='YOLOv5Objects365V1Dataset',
|
80 |
+
data_root='data/objects365v1/',
|
81 |
+
ann_file='annotations/objects365_train.json',
|
82 |
+
data_prefix=dict(img='train/'),
|
83 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
84 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
85 |
+
pipeline=train_pipeline)
|
86 |
+
|
87 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
88 |
+
data_root='data/mixed_grounding/',
|
89 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
90 |
+
data_prefix=dict(img='gqa/images/'),
|
91 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
92 |
+
pipeline=train_pipeline)
|
93 |
+
|
94 |
+
flickr_train_dataset = dict(
|
95 |
+
type='YOLOv5MixedGroundingDataset',
|
96 |
+
data_root='data/flickr/',
|
97 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
98 |
+
data_prefix=dict(img='full_images/'),
|
99 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
100 |
+
pipeline=train_pipeline)
|
101 |
+
|
102 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
103 |
+
collate_fn=dict(type='yolow_collate'),
|
104 |
+
dataset=dict(_delete_=True,
|
105 |
+
type='ConcatDataset',
|
106 |
+
datasets=[
|
107 |
+
obj365v1_train_dataset,
|
108 |
+
flickr_train_dataset, mg_train_dataset
|
109 |
+
],
|
110 |
+
ignore_keys=['classes', 'palette']))
|
111 |
+
|
112 |
+
test_pipeline = [
|
113 |
+
*_base_.test_pipeline[:-1],
|
114 |
+
dict(type='LoadText'),
|
115 |
+
dict(type='mmdet.PackDetInputs',
|
116 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
117 |
+
'scale_factor', 'pad_param', 'texts'))
|
118 |
+
]
|
119 |
+
coco_val_dataset = dict(
|
120 |
+
_delete_=True,
|
121 |
+
type='MultiModalDataset',
|
122 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
123 |
+
data_root='data/coco/',
|
124 |
+
test_mode=True,
|
125 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
126 |
+
data_prefix=dict(img=''),
|
127 |
+
batch_shapes_cfg=None),
|
128 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
129 |
+
pipeline=test_pipeline)
|
130 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
131 |
+
test_dataloader = val_dataloader
|
132 |
+
|
133 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
134 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
135 |
+
metric='bbox')
|
136 |
+
test_evaluator = val_evaluator
|
137 |
+
|
138 |
+
# training settings
|
139 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
140 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
141 |
+
rule='greater'))
|
142 |
+
custom_hooks = [
|
143 |
+
dict(type='EMAHook',
|
144 |
+
ema_type='ExpMomentumEMA',
|
145 |
+
momentum=0.0001,
|
146 |
+
update_buffers=True,
|
147 |
+
strict_load=False,
|
148 |
+
priority=49),
|
149 |
+
dict(type='mmdet.PipelineSwitchHook',
|
150 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
151 |
+
switch_pipeline=train_pipeline_stage2)
|
152 |
+
]
|
153 |
+
train_cfg = dict(max_epochs=max_epochs,
|
154 |
+
val_interval=10,
|
155 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
156 |
+
_base_.val_interval_stage2)])
|
157 |
+
optim_wrapper = dict(optimizer=dict(
|
158 |
+
_delete_=True,
|
159 |
+
type='AdamW',
|
160 |
+
lr=base_lr,
|
161 |
+
weight_decay=weight_decay,
|
162 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
163 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
164 |
+
norm_decay_mult=0.0,
|
165 |
+
custom_keys={
|
166 |
+
'backbone.text_model':
|
167 |
+
dict(lr_mult=0.01),
|
168 |
+
'logit_scale':
|
169 |
+
dict(weight_decay=0.0)
|
170 |
+
}),
|
171 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py
ADDED
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 20 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-4
|
16 |
+
weight_decay = 0.025
|
17 |
+
train_batch_size_per_gpu = 4
|
18 |
+
load_from = "pretrained_models/yolo_world_v2_l_obj365v1_goldg_pretrain-a82b1fe3.pth"
|
19 |
+
# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
20 |
+
text_model_name = 'openai/clip-vit-base-patch32'
|
21 |
+
img_scale = (1280, 1280)
|
22 |
+
|
23 |
+
# model settings
|
24 |
+
model = dict(
|
25 |
+
type='YOLOWorldDetector',
|
26 |
+
mm_neck=True,
|
27 |
+
num_train_classes=num_training_classes,
|
28 |
+
num_test_classes=num_classes,
|
29 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
30 |
+
backbone=dict(
|
31 |
+
_delete_=True,
|
32 |
+
type='MultiModalYOLOBackbone',
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
text_model=dict(
|
35 |
+
type='HuggingCLIPLanguageBackbone',
|
36 |
+
model_name=text_model_name,
|
37 |
+
frozen_modules=['all'])),
|
38 |
+
neck=dict(type='YOLOWorldPAFPN',
|
39 |
+
guide_channels=text_channels,
|
40 |
+
embed_channels=neck_embed_channels,
|
41 |
+
num_heads=neck_num_heads,
|
42 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
43 |
+
bbox_head=dict(type='YOLOWorldHead',
|
44 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
text_transform = [
|
52 |
+
dict(type='RandomLoadText',
|
53 |
+
num_neg_samples=(num_classes, num_classes),
|
54 |
+
max_num_samples=num_training_classes,
|
55 |
+
padding_to_max=True,
|
56 |
+
padding_value=''),
|
57 |
+
dict(type='mmdet.PackDetInputs',
|
58 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
59 |
+
'flip_direction', 'texts'))
|
60 |
+
]
|
61 |
+
train_pipeline = [
|
62 |
+
*_base_.pre_transform,
|
63 |
+
dict(type='MultiModalMosaic',
|
64 |
+
img_scale=img_scale,
|
65 |
+
pad_val=114.0,
|
66 |
+
pre_transform=_base_.pre_transform),
|
67 |
+
dict(
|
68 |
+
type='YOLOv5RandomAffine',
|
69 |
+
max_rotate_degree=0.0,
|
70 |
+
max_shear_degree=0.0,
|
71 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
72 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
73 |
+
border=(-img_scale[0] // 2, -img_scale[1] // 2),
|
74 |
+
border_val=(114, 114, 114)),
|
75 |
+
*_base_.last_transform[:-1],
|
76 |
+
*text_transform,
|
77 |
+
]
|
78 |
+
|
79 |
+
train_pipeline_stage2 = [
|
80 |
+
*_base_.pre_transform,
|
81 |
+
dict(type='YOLOv5KeepRatioResize', scale=img_scale),
|
82 |
+
dict(
|
83 |
+
type='LetterResize',
|
84 |
+
scale=img_scale,
|
85 |
+
allow_scale_up=True,
|
86 |
+
pad_val=dict(img=114.0)),
|
87 |
+
dict(
|
88 |
+
type='YOLOv5RandomAffine',
|
89 |
+
max_rotate_degree=0.0,
|
90 |
+
max_shear_degree=0.0,
|
91 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
92 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
93 |
+
border_val=(114, 114, 114)),
|
94 |
+
*_base_.last_transform[:-1],
|
95 |
+
*text_transform
|
96 |
+
]
|
97 |
+
|
98 |
+
obj365v1_train_dataset = dict(
|
99 |
+
type='MultiModalDataset',
|
100 |
+
dataset=dict(
|
101 |
+
type='YOLOv5Objects365V1Dataset',
|
102 |
+
data_root='data/objects365v1/',
|
103 |
+
ann_file='annotations/objects365_train.json',
|
104 |
+
data_prefix=dict(img='train/'),
|
105 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
106 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
107 |
+
pipeline=train_pipeline)
|
108 |
+
|
109 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
110 |
+
data_root='data/mixed_grounding/',
|
111 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
112 |
+
data_prefix=dict(img='gqa/images/'),
|
113 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
114 |
+
pipeline=train_pipeline)
|
115 |
+
|
116 |
+
flickr_train_dataset = dict(
|
117 |
+
type='YOLOv5MixedGroundingDataset',
|
118 |
+
data_root='data/flickr/',
|
119 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
120 |
+
data_prefix=dict(img='full_images/'),
|
121 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
122 |
+
pipeline=train_pipeline)
|
123 |
+
|
124 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
125 |
+
collate_fn=dict(type='yolow_collate'),
|
126 |
+
dataset=dict(_delete_=True,
|
127 |
+
type='ConcatDataset',
|
128 |
+
datasets=[
|
129 |
+
obj365v1_train_dataset,
|
130 |
+
flickr_train_dataset, mg_train_dataset
|
131 |
+
],
|
132 |
+
ignore_keys=['classes', 'palette']))
|
133 |
+
|
134 |
+
test_pipeline = [
|
135 |
+
dict(type='LoadImageFromFile'),
|
136 |
+
dict(type='YOLOv5KeepRatioResize', scale=img_scale),
|
137 |
+
dict(
|
138 |
+
type='LetterResize',
|
139 |
+
scale=img_scale,
|
140 |
+
allow_scale_up=False,
|
141 |
+
pad_val=dict(img=114)),
|
142 |
+
dict(type='LoadAnnotations', with_bbox=True, _scope_='mmdet'),
|
143 |
+
dict(type='LoadText'),
|
144 |
+
dict(type='mmdet.PackDetInputs',
|
145 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
146 |
+
'scale_factor', 'pad_param', 'texts'))
|
147 |
+
]
|
148 |
+
|
149 |
+
coco_val_dataset = dict(
|
150 |
+
_delete_=True,
|
151 |
+
type='MultiModalDataset',
|
152 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
153 |
+
data_root='data/coco/',
|
154 |
+
test_mode=True,
|
155 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
156 |
+
data_prefix=dict(img=''),
|
157 |
+
batch_shapes_cfg=None),
|
158 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
159 |
+
pipeline=test_pipeline)
|
160 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
161 |
+
test_dataloader = val_dataloader
|
162 |
+
|
163 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
164 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
165 |
+
metric='bbox')
|
166 |
+
test_evaluator = val_evaluator
|
167 |
+
|
168 |
+
# training settings
|
169 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
170 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
171 |
+
rule='greater'))
|
172 |
+
custom_hooks = [
|
173 |
+
dict(type='EMAHook',
|
174 |
+
ema_type='ExpMomentumEMA',
|
175 |
+
momentum=0.0001,
|
176 |
+
update_buffers=True,
|
177 |
+
strict_load=False,
|
178 |
+
priority=49),
|
179 |
+
dict(type='mmdet.PipelineSwitchHook',
|
180 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
181 |
+
switch_pipeline=train_pipeline_stage2)
|
182 |
+
]
|
183 |
+
train_cfg = dict(max_epochs=max_epochs,
|
184 |
+
val_interval=10,
|
185 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
186 |
+
_base_.val_interval_stage2)])
|
187 |
+
|
188 |
+
optim_wrapper = dict(optimizer=dict(
|
189 |
+
_delete_=True,
|
190 |
+
type='AdamW',
|
191 |
+
lr=base_lr,
|
192 |
+
weight_decay=weight_decay,
|
193 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
194 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
195 |
+
norm_decay_mult=0.0,
|
196 |
+
custom_keys={
|
197 |
+
'backbone.text_model':
|
198 |
+
dict(lr_mult=0.01),
|
199 |
+
'logit_scale':
|
200 |
+
dict(weight_decay=0.0)
|
201 |
+
}),
|
202 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
19 |
+
text_model_name = 'openai/clip-vit-base-patch32'
|
20 |
+
# model settings
|
21 |
+
model = dict(
|
22 |
+
type='YOLOWorldDetector',
|
23 |
+
mm_neck=True,
|
24 |
+
num_train_classes=num_training_classes,
|
25 |
+
num_test_classes=num_classes,
|
26 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
27 |
+
backbone=dict(
|
28 |
+
_delete_=True,
|
29 |
+
type='MultiModalYOLOBackbone',
|
30 |
+
image_model={{_base_.model.backbone}},
|
31 |
+
text_model=dict(
|
32 |
+
type='HuggingCLIPLanguageBackbone',
|
33 |
+
model_name=text_model_name,
|
34 |
+
frozen_modules=['all'])),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
guide_channels=text_channels,
|
37 |
+
embed_channels=neck_embed_channels,
|
38 |
+
num_heads=neck_num_heads,
|
39 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
40 |
+
bbox_head=dict(type='YOLOWorldHead',
|
41 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
42 |
+
use_bn_head=True,
|
43 |
+
embed_dims=text_channels,
|
44 |
+
num_classes=num_training_classes)),
|
45 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
46 |
+
|
47 |
+
# dataset settings
|
48 |
+
text_transform = [
|
49 |
+
dict(type='RandomLoadText',
|
50 |
+
num_neg_samples=(num_classes, num_classes),
|
51 |
+
max_num_samples=num_training_classes,
|
52 |
+
padding_to_max=True,
|
53 |
+
padding_value=''),
|
54 |
+
dict(type='mmdet.PackDetInputs',
|
55 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
56 |
+
'flip_direction', 'texts'))
|
57 |
+
]
|
58 |
+
train_pipeline = [
|
59 |
+
*_base_.pre_transform,
|
60 |
+
dict(type='MultiModalMosaic',
|
61 |
+
img_scale=_base_.img_scale,
|
62 |
+
pad_val=114.0,
|
63 |
+
pre_transform=_base_.pre_transform),
|
64 |
+
dict(
|
65 |
+
type='YOLOv5RandomAffine',
|
66 |
+
max_rotate_degree=0.0,
|
67 |
+
max_shear_degree=0.0,
|
68 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
69 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
70 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
71 |
+
border_val=(114, 114, 114)),
|
72 |
+
*_base_.last_transform[:-1],
|
73 |
+
*text_transform,
|
74 |
+
]
|
75 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
76 |
+
obj365v1_train_dataset = dict(
|
77 |
+
type='MultiModalDataset',
|
78 |
+
dataset=dict(
|
79 |
+
type='YOLOv5Objects365V1Dataset',
|
80 |
+
data_root='data/objects365v1/',
|
81 |
+
ann_file='annotations/objects365_train.json',
|
82 |
+
data_prefix=dict(img='train/'),
|
83 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
84 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
85 |
+
pipeline=train_pipeline)
|
86 |
+
|
87 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
88 |
+
data_root='data/mixed_grounding/',
|
89 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
90 |
+
data_prefix=dict(img='gqa/images/'),
|
91 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
92 |
+
pipeline=train_pipeline)
|
93 |
+
|
94 |
+
flickr_train_dataset = dict(
|
95 |
+
type='YOLOv5MixedGroundingDataset',
|
96 |
+
data_root='data/flickr/',
|
97 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
98 |
+
data_prefix=dict(img='full_images/'),
|
99 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
100 |
+
pipeline=train_pipeline)
|
101 |
+
|
102 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
103 |
+
collate_fn=dict(type='yolow_collate'),
|
104 |
+
dataset=dict(_delete_=True,
|
105 |
+
type='ConcatDataset',
|
106 |
+
datasets=[
|
107 |
+
obj365v1_train_dataset,
|
108 |
+
flickr_train_dataset, mg_train_dataset
|
109 |
+
],
|
110 |
+
ignore_keys=['classes', 'palette']))
|
111 |
+
|
112 |
+
test_pipeline = [
|
113 |
+
*_base_.test_pipeline[:-1],
|
114 |
+
dict(type='LoadText'),
|
115 |
+
dict(type='mmdet.PackDetInputs',
|
116 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
117 |
+
'scale_factor', 'pad_param', 'texts'))
|
118 |
+
]
|
119 |
+
coco_val_dataset = dict(
|
120 |
+
_delete_=True,
|
121 |
+
type='MultiModalDataset',
|
122 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
123 |
+
data_root='data/coco/',
|
124 |
+
test_mode=True,
|
125 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
126 |
+
data_prefix=dict(img=''),
|
127 |
+
batch_shapes_cfg=None),
|
128 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
129 |
+
pipeline=test_pipeline)
|
130 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
131 |
+
test_dataloader = val_dataloader
|
132 |
+
|
133 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
134 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
135 |
+
metric='bbox')
|
136 |
+
test_evaluator = val_evaluator
|
137 |
+
|
138 |
+
# training settings
|
139 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
140 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
141 |
+
rule='greater'))
|
142 |
+
custom_hooks = [
|
143 |
+
dict(type='EMAHook',
|
144 |
+
ema_type='ExpMomentumEMA',
|
145 |
+
momentum=0.0001,
|
146 |
+
update_buffers=True,
|
147 |
+
strict_load=False,
|
148 |
+
priority=49),
|
149 |
+
dict(type='mmdet.PipelineSwitchHook',
|
150 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
151 |
+
switch_pipeline=train_pipeline_stage2)
|
152 |
+
]
|
153 |
+
train_cfg = dict(max_epochs=max_epochs,
|
154 |
+
val_interval=10,
|
155 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
156 |
+
_base_.val_interval_stage2)])
|
157 |
+
optim_wrapper = dict(optimizer=dict(
|
158 |
+
_delete_=True,
|
159 |
+
type='AdamW',
|
160 |
+
lr=base_lr,
|
161 |
+
weight_decay=weight_decay,
|
162 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
163 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
164 |
+
norm_decay_mult=0.0,
|
165 |
+
custom_keys={
|
166 |
+
'backbone.text_model':
|
167 |
+
dict(lr_mult=0.01),
|
168 |
+
'logit_scale':
|
169 |
+
dict(weight_decay=0.0)
|
170 |
+
}),
|
171 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_val.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
19 |
+
text_model_name = 'openai/clip-vit-base-patch32'
|
20 |
+
# model settings
|
21 |
+
model = dict(
|
22 |
+
type='YOLOWorldDetector',
|
23 |
+
mm_neck=True,
|
24 |
+
num_train_classes=num_training_classes,
|
25 |
+
num_test_classes=num_classes,
|
26 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
27 |
+
backbone=dict(
|
28 |
+
_delete_=True,
|
29 |
+
type='MultiModalYOLOBackbone',
|
30 |
+
image_model={{_base_.model.backbone}},
|
31 |
+
text_model=dict(
|
32 |
+
type='HuggingCLIPLanguageBackbone',
|
33 |
+
model_name=text_model_name,
|
34 |
+
frozen_modules=['all'])),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
guide_channels=text_channels,
|
37 |
+
embed_channels=neck_embed_channels,
|
38 |
+
num_heads=neck_num_heads,
|
39 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
40 |
+
bbox_head=dict(type='YOLOWorldHead',
|
41 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
42 |
+
use_bn_head=True,
|
43 |
+
embed_dims=text_channels,
|
44 |
+
num_classes=num_training_classes)),
|
45 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
46 |
+
|
47 |
+
# dataset settings
|
48 |
+
text_transform = [
|
49 |
+
dict(type='RandomLoadText',
|
50 |
+
num_neg_samples=(num_classes, num_classes),
|
51 |
+
max_num_samples=num_training_classes,
|
52 |
+
padding_to_max=True,
|
53 |
+
padding_value=''),
|
54 |
+
dict(type='mmdet.PackDetInputs',
|
55 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
56 |
+
'flip_direction', 'texts'))
|
57 |
+
]
|
58 |
+
train_pipeline = [
|
59 |
+
*_base_.pre_transform,
|
60 |
+
dict(type='MultiModalMosaic',
|
61 |
+
img_scale=_base_.img_scale,
|
62 |
+
pad_val=114.0,
|
63 |
+
pre_transform=_base_.pre_transform),
|
64 |
+
dict(
|
65 |
+
type='YOLOv5RandomAffine',
|
66 |
+
max_rotate_degree=0.0,
|
67 |
+
max_shear_degree=0.0,
|
68 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
69 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
70 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
71 |
+
border_val=(114, 114, 114)),
|
72 |
+
*_base_.last_transform[:-1],
|
73 |
+
*text_transform,
|
74 |
+
]
|
75 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
76 |
+
obj365v1_train_dataset = dict(
|
77 |
+
type='MultiModalDataset',
|
78 |
+
dataset=dict(
|
79 |
+
type='YOLOv5Objects365V1Dataset',
|
80 |
+
data_root='data/objects365v1/',
|
81 |
+
ann_file='annotations/objects365_train.json',
|
82 |
+
data_prefix=dict(img='train/'),
|
83 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
84 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
85 |
+
pipeline=train_pipeline)
|
86 |
+
|
87 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
88 |
+
data_root='data/mixed_grounding/',
|
89 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
90 |
+
data_prefix=dict(img='gqa/images/'),
|
91 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
92 |
+
pipeline=train_pipeline)
|
93 |
+
|
94 |
+
flickr_train_dataset = dict(
|
95 |
+
type='YOLOv5MixedGroundingDataset',
|
96 |
+
data_root='data/flickr/',
|
97 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
98 |
+
data_prefix=dict(img='full_images/'),
|
99 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
100 |
+
pipeline=train_pipeline)
|
101 |
+
|
102 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
103 |
+
collate_fn=dict(type='yolow_collate'),
|
104 |
+
dataset=dict(_delete_=True,
|
105 |
+
type='ConcatDataset',
|
106 |
+
datasets=[
|
107 |
+
obj365v1_train_dataset,
|
108 |
+
flickr_train_dataset, mg_train_dataset
|
109 |
+
],
|
110 |
+
ignore_keys=['classes', 'palette']))
|
111 |
+
|
112 |
+
test_pipeline = [
|
113 |
+
*_base_.test_pipeline[:-1],
|
114 |
+
dict(type='LoadText'),
|
115 |
+
dict(type='mmdet.PackDetInputs',
|
116 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
117 |
+
'scale_factor', 'pad_param', 'texts'))
|
118 |
+
]
|
119 |
+
coco_val_dataset = dict(
|
120 |
+
_delete_=True,
|
121 |
+
type='MultiModalDataset',
|
122 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
123 |
+
data_root='data/coco/',
|
124 |
+
test_mode=True,
|
125 |
+
ann_file='lvis/lvis_v1_val.json',
|
126 |
+
data_prefix=dict(img=''),
|
127 |
+
batch_shapes_cfg=None),
|
128 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
129 |
+
pipeline=test_pipeline)
|
130 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
131 |
+
test_dataloader = val_dataloader
|
132 |
+
|
133 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
134 |
+
ann_file='data/coco/lvis/lvis_v1_val.json',
|
135 |
+
metric='bbox')
|
136 |
+
test_evaluator = val_evaluator
|
137 |
+
|
138 |
+
# training settings
|
139 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
140 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
141 |
+
rule='greater'))
|
142 |
+
custom_hooks = [
|
143 |
+
dict(type='EMAHook',
|
144 |
+
ema_type='ExpMomentumEMA',
|
145 |
+
momentum=0.0001,
|
146 |
+
update_buffers=True,
|
147 |
+
strict_load=False,
|
148 |
+
priority=49),
|
149 |
+
dict(type='mmdet.PipelineSwitchHook',
|
150 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
151 |
+
switch_pipeline=train_pipeline_stage2)
|
152 |
+
]
|
153 |
+
train_cfg = dict(max_epochs=max_epochs,
|
154 |
+
val_interval=10,
|
155 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
156 |
+
_base_.val_interval_stage2)])
|
157 |
+
optim_wrapper = dict(optimizer=dict(
|
158 |
+
_delete_=True,
|
159 |
+
type='AdamW',
|
160 |
+
lr=base_lr,
|
161 |
+
weight_decay=weight_decay,
|
162 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
163 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
164 |
+
norm_decay_mult=0.0,
|
165 |
+
custom_keys={
|
166 |
+
'backbone.text_model':
|
167 |
+
dict(lr_mult=0.01),
|
168 |
+
'logit_scale':
|
169 |
+
dict(weight_decay=0.0)
|
170 |
+
}),
|
171 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_m_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py
ADDED
@@ -0,0 +1,198 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_m_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
19 |
+
img_scale = (1280, 1280)
|
20 |
+
|
21 |
+
# text_model_name = 'openai/clip-vit-base-patch32'
|
22 |
+
# model settings
|
23 |
+
model = dict(
|
24 |
+
type='YOLOWorldDetector',
|
25 |
+
mm_neck=True,
|
26 |
+
num_train_classes=num_training_classes,
|
27 |
+
num_test_classes=num_classes,
|
28 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
29 |
+
backbone=dict(
|
30 |
+
_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
image_model={{_base_.model.backbone}},
|
33 |
+
text_model=dict(
|
34 |
+
type='HuggingCLIPLanguageBackbone',
|
35 |
+
model_name=text_model_name,
|
36 |
+
frozen_modules=['all'])),
|
37 |
+
neck=dict(type='YOLOWorldPAFPN',
|
38 |
+
guide_channels=text_channels,
|
39 |
+
embed_channels=neck_embed_channels,
|
40 |
+
num_heads=neck_num_heads,
|
41 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
42 |
+
bbox_head=dict(type='YOLOWorldHead',
|
43 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
44 |
+
use_bn_head=True,
|
45 |
+
embed_dims=text_channels,
|
46 |
+
num_classes=num_training_classes)),
|
47 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
48 |
+
|
49 |
+
# dataset settings
|
50 |
+
text_transform = [
|
51 |
+
dict(type='RandomLoadText',
|
52 |
+
num_neg_samples=(num_classes, num_classes),
|
53 |
+
max_num_samples=num_training_classes,
|
54 |
+
padding_to_max=True,
|
55 |
+
padding_value=''),
|
56 |
+
dict(type='mmdet.PackDetInputs',
|
57 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
58 |
+
'flip_direction', 'texts'))
|
59 |
+
]
|
60 |
+
train_pipeline = [
|
61 |
+
*_base_.pre_transform,
|
62 |
+
dict(type='MultiModalMosaic',
|
63 |
+
img_scale=img_scale,
|
64 |
+
pad_val=114.0,
|
65 |
+
pre_transform=_base_.pre_transform),
|
66 |
+
dict(
|
67 |
+
type='YOLOv5RandomAffine',
|
68 |
+
max_rotate_degree=0.0,
|
69 |
+
max_shear_degree=0.0,
|
70 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
71 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
72 |
+
border=(-img_scale[0] // 2, -img_scale[1] // 2),
|
73 |
+
border_val=(114, 114, 114)),
|
74 |
+
*_base_.last_transform[:-1],
|
75 |
+
*text_transform,
|
76 |
+
]
|
77 |
+
|
78 |
+
train_pipeline_stage2 = [
|
79 |
+
*_base_.pre_transform,
|
80 |
+
dict(type='YOLOv5KeepRatioResize', scale=img_scale),
|
81 |
+
dict(
|
82 |
+
type='LetterResize',
|
83 |
+
scale=img_scale,
|
84 |
+
allow_scale_up=True,
|
85 |
+
pad_val=dict(img=114.0)),
|
86 |
+
dict(
|
87 |
+
type='YOLOv5RandomAffine',
|
88 |
+
max_rotate_degree=0.0,
|
89 |
+
max_shear_degree=0.0,
|
90 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
91 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
92 |
+
border_val=(114, 114, 114)),
|
93 |
+
*_base_.last_transform[:-1],
|
94 |
+
*text_transform
|
95 |
+
]
|
96 |
+
obj365v1_train_dataset = dict(
|
97 |
+
type='MultiModalDataset',
|
98 |
+
dataset=dict(
|
99 |
+
type='YOLOv5Objects365V1Dataset',
|
100 |
+
data_root='data/objects365v1/',
|
101 |
+
ann_file='annotations/objects365_train.json',
|
102 |
+
data_prefix=dict(img='train/'),
|
103 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
104 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
105 |
+
pipeline=train_pipeline)
|
106 |
+
|
107 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
108 |
+
data_root='data/mixed_grounding/',
|
109 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
110 |
+
data_prefix=dict(img='gqa/images/'),
|
111 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
112 |
+
pipeline=train_pipeline)
|
113 |
+
|
114 |
+
flickr_train_dataset = dict(
|
115 |
+
type='YOLOv5MixedGroundingDataset',
|
116 |
+
data_root='data/flickr/',
|
117 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
118 |
+
data_prefix=dict(img='full_images/'),
|
119 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
120 |
+
pipeline=train_pipeline)
|
121 |
+
|
122 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
123 |
+
collate_fn=dict(type='yolow_collate'),
|
124 |
+
dataset=dict(_delete_=True,
|
125 |
+
type='ConcatDataset',
|
126 |
+
datasets=[
|
127 |
+
obj365v1_train_dataset,
|
128 |
+
flickr_train_dataset, mg_train_dataset
|
129 |
+
],
|
130 |
+
ignore_keys=['classes', 'palette']))
|
131 |
+
|
132 |
+
test_pipeline = [
|
133 |
+
dict(type='LoadImageFromFile'),
|
134 |
+
dict(type='YOLOv5KeepRatioResize', scale=img_scale),
|
135 |
+
dict(
|
136 |
+
type='LetterResize',
|
137 |
+
scale=img_scale,
|
138 |
+
allow_scale_up=False,
|
139 |
+
pad_val=dict(img=114)),
|
140 |
+
dict(type='LoadAnnotations', with_bbox=True, _scope_='mmdet'),
|
141 |
+
dict(type='LoadText'),
|
142 |
+
dict(type='mmdet.PackDetInputs',
|
143 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
144 |
+
'scale_factor', 'pad_param', 'texts'))
|
145 |
+
]
|
146 |
+
coco_val_dataset = dict(
|
147 |
+
_delete_=True,
|
148 |
+
type='MultiModalDataset',
|
149 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
150 |
+
data_root='data/coco/',
|
151 |
+
test_mode=True,
|
152 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
153 |
+
data_prefix=dict(img=''),
|
154 |
+
batch_shapes_cfg=None),
|
155 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
156 |
+
pipeline=test_pipeline)
|
157 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
158 |
+
test_dataloader = val_dataloader
|
159 |
+
|
160 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
161 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
162 |
+
metric='bbox')
|
163 |
+
test_evaluator = val_evaluator
|
164 |
+
|
165 |
+
# training settings
|
166 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
167 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
168 |
+
rule='greater'))
|
169 |
+
custom_hooks = [
|
170 |
+
dict(type='EMAHook',
|
171 |
+
ema_type='ExpMomentumEMA',
|
172 |
+
momentum=0.0001,
|
173 |
+
update_buffers=True,
|
174 |
+
strict_load=False,
|
175 |
+
priority=49),
|
176 |
+
dict(type='mmdet.PipelineSwitchHook',
|
177 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
178 |
+
switch_pipeline=train_pipeline_stage2)
|
179 |
+
]
|
180 |
+
train_cfg = dict(max_epochs=max_epochs,
|
181 |
+
val_interval=10,
|
182 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
183 |
+
_base_.val_interval_stage2)])
|
184 |
+
optim_wrapper = dict(optimizer=dict(
|
185 |
+
_delete_=True,
|
186 |
+
type='AdamW',
|
187 |
+
lr=base_lr,
|
188 |
+
weight_decay=weight_decay,
|
189 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
190 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
191 |
+
norm_decay_mult=0.0,
|
192 |
+
custom_keys={
|
193 |
+
'backbone.text_model':
|
194 |
+
dict(lr_mult=0.01),
|
195 |
+
'logit_scale':
|
196 |
+
dict(weight_decay=0.0)
|
197 |
+
}),
|
198 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_m_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_m_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
19 |
+
# text_model_name = 'openai/clip-vit-base-patch32'
|
20 |
+
# model settings
|
21 |
+
model = dict(
|
22 |
+
type='YOLOWorldDetector',
|
23 |
+
mm_neck=True,
|
24 |
+
num_train_classes=num_training_classes,
|
25 |
+
num_test_classes=num_classes,
|
26 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
27 |
+
backbone=dict(
|
28 |
+
_delete_=True,
|
29 |
+
type='MultiModalYOLOBackbone',
|
30 |
+
image_model={{_base_.model.backbone}},
|
31 |
+
text_model=dict(
|
32 |
+
type='HuggingCLIPLanguageBackbone',
|
33 |
+
model_name=text_model_name,
|
34 |
+
frozen_modules=['all'])),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
guide_channels=text_channels,
|
37 |
+
embed_channels=neck_embed_channels,
|
38 |
+
num_heads=neck_num_heads,
|
39 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
40 |
+
bbox_head=dict(type='YOLOWorldHead',
|
41 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
42 |
+
use_bn_head=True,
|
43 |
+
embed_dims=text_channels,
|
44 |
+
num_classes=num_training_classes)),
|
45 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
46 |
+
|
47 |
+
# dataset settings
|
48 |
+
text_transform = [
|
49 |
+
dict(type='RandomLoadText',
|
50 |
+
num_neg_samples=(num_classes, num_classes),
|
51 |
+
max_num_samples=num_training_classes,
|
52 |
+
padding_to_max=True,
|
53 |
+
padding_value=''),
|
54 |
+
dict(type='mmdet.PackDetInputs',
|
55 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
56 |
+
'flip_direction', 'texts'))
|
57 |
+
]
|
58 |
+
train_pipeline = [
|
59 |
+
*_base_.pre_transform,
|
60 |
+
dict(type='MultiModalMosaic',
|
61 |
+
img_scale=_base_.img_scale,
|
62 |
+
pad_val=114.0,
|
63 |
+
pre_transform=_base_.pre_transform),
|
64 |
+
dict(
|
65 |
+
type='YOLOv5RandomAffine',
|
66 |
+
max_rotate_degree=0.0,
|
67 |
+
max_shear_degree=0.0,
|
68 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
69 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
70 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
71 |
+
border_val=(114, 114, 114)),
|
72 |
+
*_base_.last_transform[:-1],
|
73 |
+
*text_transform,
|
74 |
+
]
|
75 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
76 |
+
obj365v1_train_dataset = dict(
|
77 |
+
type='MultiModalDataset',
|
78 |
+
dataset=dict(
|
79 |
+
type='YOLOv5Objects365V1Dataset',
|
80 |
+
data_root='data/objects365v1/',
|
81 |
+
ann_file='annotations/objects365_train.json',
|
82 |
+
data_prefix=dict(img='train/'),
|
83 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
84 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
85 |
+
pipeline=train_pipeline)
|
86 |
+
|
87 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
88 |
+
data_root='data/mixed_grounding/',
|
89 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
90 |
+
data_prefix=dict(img='gqa/images/'),
|
91 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
92 |
+
pipeline=train_pipeline)
|
93 |
+
|
94 |
+
flickr_train_dataset = dict(
|
95 |
+
type='YOLOv5MixedGroundingDataset',
|
96 |
+
data_root='data/flickr/',
|
97 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
98 |
+
data_prefix=dict(img='full_images/'),
|
99 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
100 |
+
pipeline=train_pipeline)
|
101 |
+
|
102 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
103 |
+
collate_fn=dict(type='yolow_collate'),
|
104 |
+
dataset=dict(_delete_=True,
|
105 |
+
type='ConcatDataset',
|
106 |
+
datasets=[
|
107 |
+
obj365v1_train_dataset,
|
108 |
+
flickr_train_dataset, mg_train_dataset
|
109 |
+
],
|
110 |
+
ignore_keys=['classes', 'palette']))
|
111 |
+
|
112 |
+
test_pipeline = [
|
113 |
+
*_base_.test_pipeline[:-1],
|
114 |
+
dict(type='LoadText'),
|
115 |
+
dict(type='mmdet.PackDetInputs',
|
116 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
117 |
+
'scale_factor', 'pad_param', 'texts'))
|
118 |
+
]
|
119 |
+
coco_val_dataset = dict(
|
120 |
+
_delete_=True,
|
121 |
+
type='MultiModalDataset',
|
122 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
123 |
+
data_root='data/coco/',
|
124 |
+
test_mode=True,
|
125 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
126 |
+
data_prefix=dict(img=''),
|
127 |
+
batch_shapes_cfg=None),
|
128 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
129 |
+
pipeline=test_pipeline)
|
130 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
131 |
+
test_dataloader = val_dataloader
|
132 |
+
|
133 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
134 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
135 |
+
metric='bbox')
|
136 |
+
test_evaluator = val_evaluator
|
137 |
+
|
138 |
+
# training settings
|
139 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
140 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
141 |
+
rule='greater'))
|
142 |
+
custom_hooks = [
|
143 |
+
dict(type='EMAHook',
|
144 |
+
ema_type='ExpMomentumEMA',
|
145 |
+
momentum=0.0001,
|
146 |
+
update_buffers=True,
|
147 |
+
strict_load=False,
|
148 |
+
priority=49),
|
149 |
+
dict(type='mmdet.PipelineSwitchHook',
|
150 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
151 |
+
switch_pipeline=train_pipeline_stage2)
|
152 |
+
]
|
153 |
+
train_cfg = dict(max_epochs=max_epochs,
|
154 |
+
val_interval=10,
|
155 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
156 |
+
_base_.val_interval_stage2)])
|
157 |
+
optim_wrapper = dict(optimizer=dict(
|
158 |
+
_delete_=True,
|
159 |
+
type='AdamW',
|
160 |
+
lr=base_lr,
|
161 |
+
weight_decay=weight_decay,
|
162 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
163 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
164 |
+
norm_decay_mult=0.0,
|
165 |
+
custom_keys={
|
166 |
+
'backbone.text_model':
|
167 |
+
dict(lr_mult=0.01),
|
168 |
+
'logit_scale':
|
169 |
+
dict(weight_decay=0.0)
|
170 |
+
}),
|
171 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_m_vlpan_bn_noeinsum_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,176 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_m_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
19 |
+
# text_model_name = 'openai/clip-vit-base-patch32'
|
20 |
+
# model settings
|
21 |
+
model = dict(
|
22 |
+
type='YOLOWorldDetector',
|
23 |
+
mm_neck=True,
|
24 |
+
num_train_classes=num_training_classes,
|
25 |
+
num_test_classes=num_classes,
|
26 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
27 |
+
backbone=dict(
|
28 |
+
_delete_=True,
|
29 |
+
type='MultiModalYOLOBackbone',
|
30 |
+
image_model={{_base_.model.backbone}},
|
31 |
+
text_model=dict(
|
32 |
+
type='HuggingCLIPLanguageBackbone',
|
33 |
+
model_name=text_model_name,
|
34 |
+
frozen_modules=['all'])),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
guide_channels=text_channels,
|
37 |
+
embed_channels=neck_embed_channels,
|
38 |
+
num_heads=neck_num_heads,
|
39 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv',
|
40 |
+
use_einsum=False)),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
43 |
+
use_bn_head=True,
|
44 |
+
embed_dims=text_channels,
|
45 |
+
num_classes=num_training_classes,
|
46 |
+
use_einsum=False)),
|
47 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
48 |
+
|
49 |
+
# dataset settings
|
50 |
+
text_transform = [
|
51 |
+
dict(type='RandomLoadText',
|
52 |
+
num_neg_samples=(num_classes, num_classes),
|
53 |
+
max_num_samples=num_training_classes,
|
54 |
+
padding_to_max=True,
|
55 |
+
padding_value=''),
|
56 |
+
dict(type='mmdet.PackDetInputs',
|
57 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
58 |
+
'flip_direction', 'texts'))
|
59 |
+
]
|
60 |
+
|
61 |
+
train_pipeline = [
|
62 |
+
*_base_.pre_transform,
|
63 |
+
dict(type='MultiModalMosaic',
|
64 |
+
img_scale=_base_.img_scale,
|
65 |
+
pad_val=114.0,
|
66 |
+
pre_transform=_base_.pre_transform),
|
67 |
+
dict(
|
68 |
+
type='YOLOv5RandomAffine',
|
69 |
+
max_rotate_degree=0.0,
|
70 |
+
max_shear_degree=0.0,
|
71 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
72 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
73 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
74 |
+
border_val=(114, 114, 114)),
|
75 |
+
*_base_.last_transform[:-1],
|
76 |
+
*text_transform,
|
77 |
+
]
|
78 |
+
|
79 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
80 |
+
obj365v1_train_dataset = dict(
|
81 |
+
type='MultiModalDataset',
|
82 |
+
dataset=dict(
|
83 |
+
type='YOLOv5Objects365V1Dataset',
|
84 |
+
data_root='data/objects365v1/',
|
85 |
+
ann_file='annotations/objects365_train.json',
|
86 |
+
data_prefix=dict(img='train/'),
|
87 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
88 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
89 |
+
pipeline=train_pipeline)
|
90 |
+
|
91 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
92 |
+
data_root='data/mixed_grounding/',
|
93 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
94 |
+
data_prefix=dict(img='gqa/images/'),
|
95 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
96 |
+
pipeline=train_pipeline)
|
97 |
+
|
98 |
+
flickr_train_dataset = dict(
|
99 |
+
type='YOLOv5MixedGroundingDataset',
|
100 |
+
data_root='data/flickr/',
|
101 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
102 |
+
data_prefix=dict(img='full_images/'),
|
103 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
104 |
+
pipeline=train_pipeline)
|
105 |
+
|
106 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
107 |
+
collate_fn=dict(type='yolow_collate'),
|
108 |
+
dataset=dict(_delete_=True,
|
109 |
+
type='ConcatDataset',
|
110 |
+
datasets=[
|
111 |
+
obj365v1_train_dataset,
|
112 |
+
flickr_train_dataset, mg_train_dataset
|
113 |
+
],
|
114 |
+
ignore_keys=['classes', 'palette']))
|
115 |
+
|
116 |
+
test_pipeline = [
|
117 |
+
*_base_.test_pipeline[:-1],
|
118 |
+
dict(type='LoadText'),
|
119 |
+
dict(type='mmdet.PackDetInputs',
|
120 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
121 |
+
'scale_factor', 'pad_param', 'texts'))
|
122 |
+
]
|
123 |
+
|
124 |
+
coco_val_dataset = dict(
|
125 |
+
_delete_=True,
|
126 |
+
type='MultiModalDataset',
|
127 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
128 |
+
data_root='data/coco/',
|
129 |
+
test_mode=True,
|
130 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
131 |
+
data_prefix=dict(img=''),
|
132 |
+
batch_shapes_cfg=None),
|
133 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
134 |
+
pipeline=test_pipeline)
|
135 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
136 |
+
test_dataloader = val_dataloader
|
137 |
+
|
138 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
139 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
140 |
+
metric='bbox')
|
141 |
+
test_evaluator = val_evaluator
|
142 |
+
|
143 |
+
# training settings
|
144 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
145 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
146 |
+
rule='greater'))
|
147 |
+
custom_hooks = [
|
148 |
+
dict(type='EMAHook',
|
149 |
+
ema_type='ExpMomentumEMA',
|
150 |
+
momentum=0.0001,
|
151 |
+
update_buffers=True,
|
152 |
+
strict_load=False,
|
153 |
+
priority=49),
|
154 |
+
dict(type='mmdet.PipelineSwitchHook',
|
155 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
156 |
+
switch_pipeline=train_pipeline_stage2)
|
157 |
+
]
|
158 |
+
train_cfg = dict(max_epochs=max_epochs,
|
159 |
+
val_interval=10,
|
160 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
161 |
+
_base_.val_interval_stage2)])
|
162 |
+
optim_wrapper = dict(optimizer=dict(
|
163 |
+
_delete_=True,
|
164 |
+
type='AdamW',
|
165 |
+
lr=base_lr,
|
166 |
+
weight_decay=weight_decay,
|
167 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
168 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
169 |
+
norm_decay_mult=0.0,
|
170 |
+
custom_keys={
|
171 |
+
'backbone.text_model':
|
172 |
+
dict(lr_mult=0.01),
|
173 |
+
'logit_scale':
|
174 |
+
dict(weight_decay=0.0)
|
175 |
+
}),
|
176 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_s_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py
ADDED
@@ -0,0 +1,195 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_s_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-4
|
16 |
+
weight_decay = 0.025
|
17 |
+
train_batch_size_per_gpu = 4
|
18 |
+
img_scale = (1280, 1280)
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(
|
22 |
+
type='YOLOWorldDetector',
|
23 |
+
mm_neck=True,
|
24 |
+
num_train_classes=num_training_classes,
|
25 |
+
num_test_classes=num_classes,
|
26 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
27 |
+
backbone=dict(
|
28 |
+
_delete_=True,
|
29 |
+
type='MultiModalYOLOBackbone',
|
30 |
+
image_model={{_base_.model.backbone}},
|
31 |
+
text_model=dict(
|
32 |
+
type='HuggingCLIPLanguageBackbone',
|
33 |
+
model_name='openai/clip-vit-base-patch32',
|
34 |
+
frozen_modules=['all'])),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
guide_channels=text_channels,
|
37 |
+
embed_channels=neck_embed_channels,
|
38 |
+
num_heads=neck_num_heads,
|
39 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
40 |
+
bbox_head=dict(type='YOLOWorldHead',
|
41 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
42 |
+
use_bn_head=True,
|
43 |
+
embed_dims=text_channels,
|
44 |
+
num_classes=num_training_classes)),
|
45 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
46 |
+
|
47 |
+
# dataset settings
|
48 |
+
text_transform = [
|
49 |
+
dict(type='RandomLoadText',
|
50 |
+
num_neg_samples=(num_classes, num_classes),
|
51 |
+
max_num_samples=num_training_classes,
|
52 |
+
padding_to_max=True,
|
53 |
+
padding_value=''),
|
54 |
+
dict(type='mmdet.PackDetInputs',
|
55 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
56 |
+
'flip_direction', 'texts'))
|
57 |
+
]
|
58 |
+
train_pipeline = [
|
59 |
+
*_base_.pre_transform,
|
60 |
+
dict(type='MultiModalMosaic',
|
61 |
+
img_scale=img_scale,
|
62 |
+
pad_val=114.0,
|
63 |
+
pre_transform=_base_.pre_transform),
|
64 |
+
dict(
|
65 |
+
type='YOLOv5RandomAffine',
|
66 |
+
max_rotate_degree=0.0,
|
67 |
+
max_shear_degree=0.0,
|
68 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
69 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
70 |
+
border=(-img_scale[0] // 2, -img_scale[1] // 2),
|
71 |
+
border_val=(114, 114, 114)),
|
72 |
+
*_base_.last_transform[:-1],
|
73 |
+
*text_transform,
|
74 |
+
]
|
75 |
+
train_pipeline_stage2 = [
|
76 |
+
*_base_.pre_transform,
|
77 |
+
dict(type='YOLOv5KeepRatioResize', scale=img_scale),
|
78 |
+
dict(
|
79 |
+
type='LetterResize',
|
80 |
+
scale=img_scale,
|
81 |
+
allow_scale_up=True,
|
82 |
+
pad_val=dict(img=114.0)),
|
83 |
+
dict(
|
84 |
+
type='YOLOv5RandomAffine',
|
85 |
+
max_rotate_degree=0.0,
|
86 |
+
max_shear_degree=0.0,
|
87 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
88 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
89 |
+
border_val=(114, 114, 114)),
|
90 |
+
*_base_.last_transform[:-1],
|
91 |
+
*text_transform
|
92 |
+
]
|
93 |
+
obj365v1_train_dataset = dict(
|
94 |
+
type='MultiModalDataset',
|
95 |
+
dataset=dict(
|
96 |
+
type='YOLOv5Objects365V1Dataset',
|
97 |
+
data_root='data/objects365v1/',
|
98 |
+
ann_file='annotations/objects365_train.json',
|
99 |
+
data_prefix=dict(img='train/'),
|
100 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
101 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
102 |
+
pipeline=train_pipeline)
|
103 |
+
|
104 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
105 |
+
data_root='data/mixed_grounding/',
|
106 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
107 |
+
data_prefix=dict(img='gqa/images/'),
|
108 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
109 |
+
pipeline=train_pipeline)
|
110 |
+
|
111 |
+
flickr_train_dataset = dict(
|
112 |
+
type='YOLOv5MixedGroundingDataset',
|
113 |
+
data_root='data/flickr/',
|
114 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
115 |
+
data_prefix=dict(img='full_images/'),
|
116 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
117 |
+
pipeline=train_pipeline)
|
118 |
+
|
119 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
120 |
+
collate_fn=dict(type='yolow_collate'),
|
121 |
+
dataset=dict(_delete_=True,
|
122 |
+
type='ConcatDataset',
|
123 |
+
datasets=[
|
124 |
+
obj365v1_train_dataset,
|
125 |
+
flickr_train_dataset, mg_train_dataset
|
126 |
+
],
|
127 |
+
ignore_keys=['classes', 'palette']))
|
128 |
+
test_pipeline = [
|
129 |
+
dict(type='LoadImageFromFile'),
|
130 |
+
dict(type='YOLOv5KeepRatioResize', scale=img_scale),
|
131 |
+
dict(
|
132 |
+
type='LetterResize',
|
133 |
+
scale=img_scale,
|
134 |
+
allow_scale_up=False,
|
135 |
+
pad_val=dict(img=114)),
|
136 |
+
dict(type='LoadAnnotations', with_bbox=True, _scope_='mmdet'),
|
137 |
+
dict(type='LoadText'),
|
138 |
+
dict(type='mmdet.PackDetInputs',
|
139 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
140 |
+
'scale_factor', 'pad_param', 'texts'))
|
141 |
+
]
|
142 |
+
|
143 |
+
coco_val_dataset = dict(
|
144 |
+
_delete_=True,
|
145 |
+
type='MultiModalDataset',
|
146 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
147 |
+
data_root='data/coco/',
|
148 |
+
test_mode=True,
|
149 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
150 |
+
data_prefix=dict(img=''),
|
151 |
+
batch_shapes_cfg=None),
|
152 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
153 |
+
pipeline=test_pipeline)
|
154 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
155 |
+
test_dataloader = val_dataloader
|
156 |
+
|
157 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
158 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
159 |
+
metric='bbox')
|
160 |
+
test_evaluator = val_evaluator
|
161 |
+
|
162 |
+
# training settings
|
163 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
164 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
165 |
+
rule='greater'))
|
166 |
+
custom_hooks = [
|
167 |
+
dict(type='EMAHook',
|
168 |
+
ema_type='ExpMomentumEMA',
|
169 |
+
momentum=0.0001,
|
170 |
+
update_buffers=True,
|
171 |
+
strict_load=False,
|
172 |
+
priority=49),
|
173 |
+
dict(type='mmdet.PipelineSwitchHook',
|
174 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
175 |
+
switch_pipeline=train_pipeline_stage2)
|
176 |
+
]
|
177 |
+
train_cfg = dict(max_epochs=max_epochs,
|
178 |
+
val_interval=10,
|
179 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
180 |
+
_base_.val_interval_stage2)])
|
181 |
+
optim_wrapper = dict(optimizer=dict(
|
182 |
+
_delete_=True,
|
183 |
+
type='AdamW',
|
184 |
+
lr=base_lr,
|
185 |
+
weight_decay=weight_decay,
|
186 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
187 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
188 |
+
norm_decay_mult=0.0,
|
189 |
+
custom_keys={
|
190 |
+
'backbone.text_model':
|
191 |
+
dict(lr_mult=0.01),
|
192 |
+
'logit_scale':
|
193 |
+
dict(weight_decay=0.0)
|
194 |
+
}),
|
195 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_s_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,170 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_s_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
|
19 |
+
# model settings
|
20 |
+
model = dict(
|
21 |
+
type='YOLOWorldDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
26 |
+
backbone=dict(
|
27 |
+
_delete_=True,
|
28 |
+
type='MultiModalYOLOBackbone',
|
29 |
+
image_model={{_base_.model.backbone}},
|
30 |
+
text_model=dict(
|
31 |
+
type='HuggingCLIPLanguageBackbone',
|
32 |
+
model_name='openai/clip-vit-base-patch32',
|
33 |
+
frozen_modules=['all'])),
|
34 |
+
neck=dict(type='YOLOWorldPAFPN',
|
35 |
+
guide_channels=text_channels,
|
36 |
+
embed_channels=neck_embed_channels,
|
37 |
+
num_heads=neck_num_heads,
|
38 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
39 |
+
bbox_head=dict(type='YOLOWorldHead',
|
40 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
41 |
+
use_bn_head=True,
|
42 |
+
embed_dims=text_channels,
|
43 |
+
num_classes=num_training_classes)),
|
44 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
45 |
+
|
46 |
+
# dataset settings
|
47 |
+
text_transform = [
|
48 |
+
dict(type='RandomLoadText',
|
49 |
+
num_neg_samples=(num_classes, num_classes),
|
50 |
+
max_num_samples=num_training_classes,
|
51 |
+
padding_to_max=True,
|
52 |
+
padding_value=''),
|
53 |
+
dict(type='mmdet.PackDetInputs',
|
54 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
55 |
+
'flip_direction', 'texts'))
|
56 |
+
]
|
57 |
+
train_pipeline = [
|
58 |
+
*_base_.pre_transform,
|
59 |
+
dict(type='MultiModalMosaic',
|
60 |
+
img_scale=_base_.img_scale,
|
61 |
+
pad_val=114.0,
|
62 |
+
pre_transform=_base_.pre_transform),
|
63 |
+
dict(
|
64 |
+
type='YOLOv5RandomAffine',
|
65 |
+
max_rotate_degree=0.0,
|
66 |
+
max_shear_degree=0.0,
|
67 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
68 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
69 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
70 |
+
border_val=(114, 114, 114)),
|
71 |
+
*_base_.last_transform[:-1],
|
72 |
+
*text_transform,
|
73 |
+
]
|
74 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
75 |
+
obj365v1_train_dataset = dict(
|
76 |
+
type='MultiModalDataset',
|
77 |
+
dataset=dict(
|
78 |
+
type='YOLOv5Objects365V1Dataset',
|
79 |
+
data_root='data/objects365v1/',
|
80 |
+
ann_file='annotations/objects365_train.json',
|
81 |
+
data_prefix=dict(img='train/'),
|
82 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
83 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
84 |
+
pipeline=train_pipeline)
|
85 |
+
|
86 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
87 |
+
data_root='data/mixed_grounding/',
|
88 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
89 |
+
data_prefix=dict(img='gqa/images/'),
|
90 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
91 |
+
pipeline=train_pipeline)
|
92 |
+
|
93 |
+
flickr_train_dataset = dict(
|
94 |
+
type='YOLOv5MixedGroundingDataset',
|
95 |
+
data_root='data/flickr/',
|
96 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
97 |
+
data_prefix=dict(img='full_images/'),
|
98 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
99 |
+
pipeline=train_pipeline)
|
100 |
+
|
101 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
102 |
+
collate_fn=dict(type='yolow_collate'),
|
103 |
+
dataset=dict(_delete_=True,
|
104 |
+
type='ConcatDataset',
|
105 |
+
datasets=[
|
106 |
+
obj365v1_train_dataset,
|
107 |
+
flickr_train_dataset, mg_train_dataset
|
108 |
+
],
|
109 |
+
ignore_keys=['classes', 'palette']))
|
110 |
+
|
111 |
+
test_pipeline = [
|
112 |
+
*_base_.test_pipeline[:-1],
|
113 |
+
dict(type='LoadText'),
|
114 |
+
dict(type='mmdet.PackDetInputs',
|
115 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
116 |
+
'scale_factor', 'pad_param', 'texts'))
|
117 |
+
]
|
118 |
+
coco_val_dataset = dict(
|
119 |
+
_delete_=True,
|
120 |
+
type='MultiModalDataset',
|
121 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
122 |
+
data_root='data/coco/',
|
123 |
+
test_mode=True,
|
124 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
125 |
+
data_prefix=dict(img=''),
|
126 |
+
batch_shapes_cfg=None),
|
127 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
128 |
+
pipeline=test_pipeline)
|
129 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
130 |
+
test_dataloader = val_dataloader
|
131 |
+
|
132 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
133 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
134 |
+
metric='bbox')
|
135 |
+
test_evaluator = val_evaluator
|
136 |
+
|
137 |
+
# training settings
|
138 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
139 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
140 |
+
rule='greater'))
|
141 |
+
custom_hooks = [
|
142 |
+
dict(type='EMAHook',
|
143 |
+
ema_type='ExpMomentumEMA',
|
144 |
+
momentum=0.0001,
|
145 |
+
update_buffers=True,
|
146 |
+
strict_load=False,
|
147 |
+
priority=49),
|
148 |
+
dict(type='mmdet.PipelineSwitchHook',
|
149 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
150 |
+
switch_pipeline=train_pipeline_stage2)
|
151 |
+
]
|
152 |
+
train_cfg = dict(max_epochs=max_epochs,
|
153 |
+
val_interval=10,
|
154 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
155 |
+
_base_.val_interval_stage2)])
|
156 |
+
optim_wrapper = dict(optimizer=dict(
|
157 |
+
_delete_=True,
|
158 |
+
type='AdamW',
|
159 |
+
lr=base_lr,
|
160 |
+
weight_decay=weight_decay,
|
161 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
162 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
163 |
+
norm_decay_mult=0.0,
|
164 |
+
custom_keys={
|
165 |
+
'backbone.text_model':
|
166 |
+
dict(lr_mult=0.01),
|
167 |
+
'logit_scale':
|
168 |
+
dict(weight_decay=0.0)
|
169 |
+
}),
|
170 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_x_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_x_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
19 |
+
text_model_name = 'openai/clip-vit-base-patch32'
|
20 |
+
# model settings
|
21 |
+
model = dict(
|
22 |
+
type='YOLOWorldDetector',
|
23 |
+
mm_neck=True,
|
24 |
+
num_train_classes=num_training_classes,
|
25 |
+
num_test_classes=num_classes,
|
26 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
27 |
+
backbone=dict(
|
28 |
+
_delete_=True,
|
29 |
+
type='MultiModalYOLOBackbone',
|
30 |
+
image_model={{_base_.model.backbone}},
|
31 |
+
text_model=dict(
|
32 |
+
type='HuggingCLIPLanguageBackbone',
|
33 |
+
model_name=text_model_name,
|
34 |
+
frozen_modules=['all'])),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
guide_channels=text_channels,
|
37 |
+
embed_channels=neck_embed_channels,
|
38 |
+
num_heads=neck_num_heads,
|
39 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
40 |
+
bbox_head=dict(type='YOLOWorldHead',
|
41 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
42 |
+
use_bn_head=True,
|
43 |
+
embed_dims=text_channels,
|
44 |
+
num_classes=num_training_classes)),
|
45 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
46 |
+
|
47 |
+
# dataset settings
|
48 |
+
text_transform = [
|
49 |
+
dict(type='RandomLoadText',
|
50 |
+
num_neg_samples=(num_classes, num_classes),
|
51 |
+
max_num_samples=num_training_classes,
|
52 |
+
padding_to_max=True,
|
53 |
+
padding_value=''),
|
54 |
+
dict(type='mmdet.PackDetInputs',
|
55 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
56 |
+
'flip_direction', 'texts'))
|
57 |
+
]
|
58 |
+
train_pipeline = [
|
59 |
+
*_base_.pre_transform,
|
60 |
+
dict(type='MultiModalMosaic',
|
61 |
+
img_scale=_base_.img_scale,
|
62 |
+
pad_val=114.0,
|
63 |
+
pre_transform=_base_.pre_transform),
|
64 |
+
dict(
|
65 |
+
type='YOLOv5RandomAffine',
|
66 |
+
max_rotate_degree=0.0,
|
67 |
+
max_shear_degree=0.0,
|
68 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
69 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
70 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
71 |
+
border_val=(114, 114, 114)),
|
72 |
+
*_base_.last_transform[:-1],
|
73 |
+
*text_transform,
|
74 |
+
]
|
75 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
76 |
+
obj365v1_train_dataset = dict(
|
77 |
+
type='MultiModalDataset',
|
78 |
+
dataset=dict(
|
79 |
+
type='YOLOv5Objects365V1Dataset',
|
80 |
+
data_root='data/objects365v1/',
|
81 |
+
ann_file='annotations/objects365_train.json',
|
82 |
+
data_prefix=dict(img='train/'),
|
83 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
84 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
85 |
+
pipeline=train_pipeline)
|
86 |
+
|
87 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
88 |
+
data_root='data/mixed_grounding/',
|
89 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
90 |
+
data_prefix=dict(img='gqa/images/'),
|
91 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
92 |
+
pipeline=train_pipeline)
|
93 |
+
|
94 |
+
flickr_train_dataset = dict(
|
95 |
+
type='YOLOv5MixedGroundingDataset',
|
96 |
+
data_root='data/flickr/',
|
97 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
98 |
+
data_prefix=dict(img='full_images/'),
|
99 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
100 |
+
pipeline=train_pipeline)
|
101 |
+
|
102 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
103 |
+
collate_fn=dict(type='yolow_collate'),
|
104 |
+
dataset=dict(_delete_=True,
|
105 |
+
type='ConcatDataset',
|
106 |
+
datasets=[
|
107 |
+
obj365v1_train_dataset,
|
108 |
+
flickr_train_dataset, mg_train_dataset
|
109 |
+
],
|
110 |
+
ignore_keys=['classes', 'palette']))
|
111 |
+
|
112 |
+
test_pipeline = [
|
113 |
+
*_base_.test_pipeline[:-1],
|
114 |
+
dict(type='LoadText'),
|
115 |
+
dict(type='mmdet.PackDetInputs',
|
116 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
117 |
+
'scale_factor', 'pad_param', 'texts'))
|
118 |
+
]
|
119 |
+
coco_val_dataset = dict(
|
120 |
+
_delete_=True,
|
121 |
+
type='MultiModalDataset',
|
122 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
123 |
+
data_root='data/coco/',
|
124 |
+
test_mode=True,
|
125 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
126 |
+
data_prefix=dict(img=''),
|
127 |
+
batch_shapes_cfg=None),
|
128 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
129 |
+
pipeline=test_pipeline)
|
130 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
131 |
+
test_dataloader = val_dataloader
|
132 |
+
|
133 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
134 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
135 |
+
metric='bbox')
|
136 |
+
test_evaluator = val_evaluator
|
137 |
+
|
138 |
+
# training settings
|
139 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
140 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
141 |
+
rule='greater'))
|
142 |
+
custom_hooks = [
|
143 |
+
dict(type='EMAHook',
|
144 |
+
ema_type='ExpMomentumEMA',
|
145 |
+
momentum=0.0001,
|
146 |
+
update_buffers=True,
|
147 |
+
strict_load=False,
|
148 |
+
priority=49),
|
149 |
+
dict(type='mmdet.PipelineSwitchHook',
|
150 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
151 |
+
switch_pipeline=train_pipeline_stage2)
|
152 |
+
]
|
153 |
+
train_cfg = dict(max_epochs=max_epochs,
|
154 |
+
val_interval=10,
|
155 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
156 |
+
_base_.val_interval_stage2)])
|
157 |
+
optim_wrapper = dict(optimizer=dict(
|
158 |
+
_delete_=True,
|
159 |
+
type='AdamW',
|
160 |
+
lr=base_lr,
|
161 |
+
weight_decay=weight_decay,
|
162 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
163 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
164 |
+
norm_decay_mult=0.0,
|
165 |
+
custom_keys={
|
166 |
+
'backbone.text_model':
|
167 |
+
dict(lr_mult=0.01),
|
168 |
+
'logit_scale':
|
169 |
+
dict(weight_decay=0.0)
|
170 |
+
}),
|
171 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain/yolo_world_v2_xl_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,185 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_x_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
19 |
+
# text_model_name = 'openai/clip-vit-base-patch32'
|
20 |
+
|
21 |
+
# scaling model from X to XL
|
22 |
+
deepen_factor = 1.0
|
23 |
+
widen_factor = 1.5
|
24 |
+
|
25 |
+
backbone = _base_.model.backbone
|
26 |
+
backbone.update(
|
27 |
+
deepen_factor=deepen_factor,
|
28 |
+
widen_factor=widen_factor
|
29 |
+
)
|
30 |
+
|
31 |
+
# model settings
|
32 |
+
model = dict(
|
33 |
+
type='YOLOWorldDetector',
|
34 |
+
mm_neck=True,
|
35 |
+
num_train_classes=num_training_classes,
|
36 |
+
num_test_classes=num_classes,
|
37 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
38 |
+
backbone=dict(
|
39 |
+
_delete_=True,
|
40 |
+
type='MultiModalYOLOBackbone',
|
41 |
+
image_model=backbone,
|
42 |
+
text_model=dict(
|
43 |
+
type='HuggingCLIPLanguageBackbone',
|
44 |
+
model_name=text_model_name,
|
45 |
+
frozen_modules=['all'])),
|
46 |
+
neck=dict(type='YOLOWorldPAFPN',
|
47 |
+
deepen_factor=deepen_factor,
|
48 |
+
widen_factor=widen_factor,
|
49 |
+
guide_channels=text_channels,
|
50 |
+
embed_channels=neck_embed_channels,
|
51 |
+
num_heads=neck_num_heads,
|
52 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
53 |
+
bbox_head=dict(type='YOLOWorldHead',
|
54 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
55 |
+
widen_factor=widen_factor,
|
56 |
+
use_bn_head=True,
|
57 |
+
embed_dims=text_channels,
|
58 |
+
num_classes=num_training_classes)),
|
59 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
60 |
+
|
61 |
+
# dataset settings
|
62 |
+
text_transform = [
|
63 |
+
dict(type='RandomLoadText',
|
64 |
+
num_neg_samples=(num_classes, num_classes),
|
65 |
+
max_num_samples=num_training_classes,
|
66 |
+
padding_to_max=True,
|
67 |
+
padding_value=''),
|
68 |
+
dict(type='mmdet.PackDetInputs',
|
69 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
70 |
+
'flip_direction', 'texts'))
|
71 |
+
]
|
72 |
+
train_pipeline = [
|
73 |
+
*_base_.pre_transform,
|
74 |
+
dict(type='MultiModalMosaic',
|
75 |
+
img_scale=_base_.img_scale,
|
76 |
+
pad_val=114.0,
|
77 |
+
pre_transform=_base_.pre_transform),
|
78 |
+
dict(
|
79 |
+
type='YOLOv5RandomAffine',
|
80 |
+
max_rotate_degree=0.0,
|
81 |
+
max_shear_degree=0.0,
|
82 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
83 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
84 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
85 |
+
border_val=(114, 114, 114)),
|
86 |
+
*_base_.last_transform[:-1],
|
87 |
+
*text_transform,
|
88 |
+
]
|
89 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
90 |
+
obj365v1_train_dataset = dict(
|
91 |
+
type='MultiModalDataset',
|
92 |
+
dataset=dict(
|
93 |
+
type='YOLOv5Objects365V1Dataset',
|
94 |
+
data_root='data/objects365v1/',
|
95 |
+
ann_file='annotations/objects365_train.json',
|
96 |
+
data_prefix=dict(img='train/'),
|
97 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
98 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
99 |
+
pipeline=train_pipeline)
|
100 |
+
|
101 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
102 |
+
data_root='data/mixed_grounding/',
|
103 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
104 |
+
data_prefix=dict(img='gqa/images/'),
|
105 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
106 |
+
pipeline=train_pipeline)
|
107 |
+
|
108 |
+
flickr_train_dataset = dict(
|
109 |
+
type='YOLOv5MixedGroundingDataset',
|
110 |
+
data_root='data/flickr/',
|
111 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
112 |
+
data_prefix=dict(img='full_images/'),
|
113 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
114 |
+
pipeline=train_pipeline)
|
115 |
+
|
116 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
117 |
+
collate_fn=dict(type='yolow_collate'),
|
118 |
+
dataset=dict(_delete_=True,
|
119 |
+
type='ConcatDataset',
|
120 |
+
datasets=[
|
121 |
+
obj365v1_train_dataset,
|
122 |
+
flickr_train_dataset, mg_train_dataset
|
123 |
+
],
|
124 |
+
ignore_keys=['classes', 'palette']))
|
125 |
+
|
126 |
+
test_pipeline = [
|
127 |
+
*_base_.test_pipeline[:-1],
|
128 |
+
dict(type='LoadText'),
|
129 |
+
dict(type='mmdet.PackDetInputs',
|
130 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
131 |
+
'scale_factor', 'pad_param', 'texts'))
|
132 |
+
]
|
133 |
+
coco_val_dataset = dict(
|
134 |
+
_delete_=True,
|
135 |
+
type='MultiModalDataset',
|
136 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
137 |
+
data_root='data/coco/',
|
138 |
+
test_mode=True,
|
139 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
140 |
+
data_prefix=dict(img=''),
|
141 |
+
batch_shapes_cfg=None),
|
142 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
143 |
+
pipeline=test_pipeline)
|
144 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
145 |
+
test_dataloader = val_dataloader
|
146 |
+
|
147 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
148 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
149 |
+
metric='bbox')
|
150 |
+
test_evaluator = val_evaluator
|
151 |
+
|
152 |
+
# training settings
|
153 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
154 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
155 |
+
rule='greater'))
|
156 |
+
custom_hooks = [
|
157 |
+
dict(type='EMAHook',
|
158 |
+
ema_type='ExpMomentumEMA',
|
159 |
+
momentum=0.0001,
|
160 |
+
update_buffers=True,
|
161 |
+
strict_load=False,
|
162 |
+
priority=49),
|
163 |
+
dict(type='mmdet.PipelineSwitchHook',
|
164 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
165 |
+
switch_pipeline=train_pipeline_stage2)
|
166 |
+
]
|
167 |
+
train_cfg = dict(max_epochs=max_epochs,
|
168 |
+
val_interval=10,
|
169 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
170 |
+
_base_.val_interval_stage2)])
|
171 |
+
optim_wrapper = dict(optimizer=dict(
|
172 |
+
_delete_=True,
|
173 |
+
type='AdamW',
|
174 |
+
lr=base_lr,
|
175 |
+
weight_decay=weight_decay,
|
176 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
177 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
178 |
+
norm_decay_mult=0.0,
|
179 |
+
custom_keys={
|
180 |
+
'backbone.text_model':
|
181 |
+
dict(lr_mult=0.01),
|
182 |
+
'logit_scale':
|
183 |
+
dict(weight_decay=0.0)
|
184 |
+
}),
|
185 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain_v1/README.md
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## Pre-training YOLO-World-v1
|
2 |
+
|
3 |
+
> The YOLO-World-v1 is an initial version, and now is nearly deprecated! We strongly suggest you use the [latest version](../pretrain/).
|
4 |
+
|
5 |
+
|
6 |
+
|
7 |
+
### Zero-shot Inference on LVIS dataset
|
8 |
+
|
9 |
+
| model | Pre-train Data | Size | AP<sup>mini</su> | AP<sub>r</sub> | AP<sub>c</sub> | AP<sub>f</sub> | AP<sup>val</su> | AP<sub>r</sub> | AP<sub>c</sub> | AP<sub>f</sub> | weights |
|
10 |
+
| :------------------------------------------------------------------------------------------------------------------- | :------------------- | :----------------- | :--------------: | :------------: | :------------: | :------------: | :-------------: | :------------: | :------------: | :------------: | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
|
11 |
+
| [YOLO-World-S](./yolo_world_s_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG | 640 | 24.3 | 16.6 | 22.1 | 27.7 | 17.8 | 11.0 | 14.8 | 24.0 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/resolve/main/yolo_world_s_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-18bea4d2.pth) |
|
12 |
+
| [YOLO-World-M](./yolo_world_m_dual_l2norm_2e-4_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG | 640 | 28.6 | 19.7 | 26.6 | 31.9 | 22.3 | 16.2 | 19.0 | 28.7 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/resolve/main/yolo_world_m_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-2b7bd1be.pth) |
|
13 |
+
| [YOLO-World-L](./yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG | 640 | 32.5 | 22.3 | 30.6 | 36.1 | 24.8 | 17.8 | 22.4 | 32.5 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/resolve/main/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth) |
|
14 |
+
| [YOLO-World-L](./yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG+CC3M-Lite | 640 | 33.0 | 23.6 | 32.0 | 35.5 | 25.3 | 18.0 | 22.1 | 32.1 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_cc3mlite_train_pretrained-7a5eea3b.pth) |
|
15 |
+
| [YOLO-World-X](./yolo_world_x_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG+CC3M-Lite | 640 | 33.4 | 24.4 | 31.6 | 36.6 | 26.6 | 19.2 | 23.5 | 33.2 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_x_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_cc3mlite_train_pretrained-8cf6b025.pth) |
|
16 |
+
|
17 |
+
|
18 |
+
**NOTE:**
|
19 |
+
1. AP<sup>mini</sup>: evaluated on LVIS `minival`.
|
20 |
+
3. AP<sup>val</sup>: evaluated on LVIS `val 1.0`.
|
21 |
+
4. [HuggingFace Mirror](https://hf-mirror.com/) provides the mirror of HuggingFace, which is a choice for users who are unable to reach.
|
configs/pretrain_v1/yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,172 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
|
19 |
+
# model settings
|
20 |
+
model = dict(
|
21 |
+
type='YOLOWorldDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
26 |
+
backbone=dict(
|
27 |
+
_delete_=True,
|
28 |
+
type='MultiModalYOLOBackbone',
|
29 |
+
image_model={{_base_.model.backbone}},
|
30 |
+
text_model=dict(
|
31 |
+
type='HuggingCLIPLanguageBackbone',
|
32 |
+
model_name='openai/clip-vit-base-patch32',
|
33 |
+
frozen_modules=['all'])),
|
34 |
+
neck=dict(type='YOLOWorldDualPAFPN',
|
35 |
+
guide_channels=text_channels,
|
36 |
+
embed_channels=neck_embed_channels,
|
37 |
+
num_heads=neck_num_heads,
|
38 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
|
39 |
+
text_enhancder=dict(type='ImagePoolingAttentionModule',
|
40 |
+
embed_channels=256,
|
41 |
+
num_heads=8)),
|
42 |
+
bbox_head=dict(type='YOLOWorldHead',
|
43 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
44 |
+
embed_dims=text_channels,
|
45 |
+
num_classes=num_training_classes)),
|
46 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
47 |
+
|
48 |
+
# dataset settings
|
49 |
+
text_transform = [
|
50 |
+
dict(type='RandomLoadText',
|
51 |
+
num_neg_samples=(num_classes, num_classes),
|
52 |
+
max_num_samples=num_training_classes,
|
53 |
+
padding_to_max=True,
|
54 |
+
padding_value=''),
|
55 |
+
dict(type='mmdet.PackDetInputs',
|
56 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
57 |
+
'flip_direction', 'texts'))
|
58 |
+
]
|
59 |
+
train_pipeline = [
|
60 |
+
*_base_.pre_transform,
|
61 |
+
dict(type='MultiModalMosaic',
|
62 |
+
img_scale=_base_.img_scale,
|
63 |
+
pad_val=114.0,
|
64 |
+
pre_transform=_base_.pre_transform),
|
65 |
+
dict(
|
66 |
+
type='YOLOv5RandomAffine',
|
67 |
+
max_rotate_degree=0.0,
|
68 |
+
max_shear_degree=0.0,
|
69 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
70 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
71 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
72 |
+
border_val=(114, 114, 114)),
|
73 |
+
*_base_.last_transform[:-1],
|
74 |
+
*text_transform,
|
75 |
+
]
|
76 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
77 |
+
obj365v1_train_dataset = dict(
|
78 |
+
type='MultiModalDataset',
|
79 |
+
dataset=dict(
|
80 |
+
type='YOLOv5Objects365V1Dataset',
|
81 |
+
data_root='data/objects365v1/',
|
82 |
+
ann_file='annotations/objects365_train.json',
|
83 |
+
data_prefix=dict(img='train/'),
|
84 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
85 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
86 |
+
pipeline=train_pipeline)
|
87 |
+
|
88 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
89 |
+
data_root='data/mixed_grounding/',
|
90 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
91 |
+
data_prefix=dict(img='gqa/images/'),
|
92 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
93 |
+
pipeline=train_pipeline)
|
94 |
+
|
95 |
+
flickr_train_dataset = dict(
|
96 |
+
type='YOLOv5MixedGroundingDataset',
|
97 |
+
data_root='data/flickr/',
|
98 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
99 |
+
data_prefix=dict(img='full_images/'),
|
100 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
101 |
+
pipeline=train_pipeline)
|
102 |
+
|
103 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
104 |
+
collate_fn=dict(type='yolow_collate'),
|
105 |
+
dataset=dict(_delete_=True,
|
106 |
+
type='ConcatDataset',
|
107 |
+
datasets=[
|
108 |
+
obj365v1_train_dataset,
|
109 |
+
flickr_train_dataset, mg_train_dataset
|
110 |
+
],
|
111 |
+
ignore_keys=['classes', 'palette']))
|
112 |
+
|
113 |
+
test_pipeline = [
|
114 |
+
*_base_.test_pipeline[:-1],
|
115 |
+
dict(type='LoadText'),
|
116 |
+
dict(type='mmdet.PackDetInputs',
|
117 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
118 |
+
'scale_factor', 'pad_param', 'texts'))
|
119 |
+
]
|
120 |
+
coco_val_dataset = dict(
|
121 |
+
_delete_=True,
|
122 |
+
type='MultiModalDataset',
|
123 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
124 |
+
data_root='data/coco/',
|
125 |
+
test_mode=True,
|
126 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
127 |
+
data_prefix=dict(img=''),
|
128 |
+
batch_shapes_cfg=None),
|
129 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
130 |
+
pipeline=test_pipeline)
|
131 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
132 |
+
test_dataloader = val_dataloader
|
133 |
+
|
134 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
135 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
136 |
+
metric='bbox')
|
137 |
+
test_evaluator = val_evaluator
|
138 |
+
|
139 |
+
# training settings
|
140 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
141 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
142 |
+
rule='greater'))
|
143 |
+
custom_hooks = [
|
144 |
+
dict(type='EMAHook',
|
145 |
+
ema_type='ExpMomentumEMA',
|
146 |
+
momentum=0.0001,
|
147 |
+
update_buffers=True,
|
148 |
+
strict_load=False,
|
149 |
+
priority=49),
|
150 |
+
dict(type='mmdet.PipelineSwitchHook',
|
151 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
152 |
+
switch_pipeline=train_pipeline_stage2)
|
153 |
+
]
|
154 |
+
train_cfg = dict(max_epochs=max_epochs,
|
155 |
+
val_interval=10,
|
156 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
157 |
+
_base_.val_interval_stage2)])
|
158 |
+
optim_wrapper = dict(optimizer=dict(
|
159 |
+
_delete_=True,
|
160 |
+
type='AdamW',
|
161 |
+
lr=base_lr,
|
162 |
+
weight_decay=weight_decay,
|
163 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
164 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
165 |
+
norm_decay_mult=0.0,
|
166 |
+
custom_keys={
|
167 |
+
'backbone.text_model':
|
168 |
+
dict(lr_mult=0.01),
|
169 |
+
'logit_scale':
|
170 |
+
dict(weight_decay=0.0)
|
171 |
+
}),
|
172 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain_v1/yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_val.py
ADDED
@@ -0,0 +1,172 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
|
19 |
+
# model settings
|
20 |
+
model = dict(
|
21 |
+
type='YOLOWorldDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
26 |
+
backbone=dict(
|
27 |
+
_delete_=True,
|
28 |
+
type='MultiModalYOLOBackbone',
|
29 |
+
image_model={{_base_.model.backbone}},
|
30 |
+
text_model=dict(
|
31 |
+
type='HuggingCLIPLanguageBackbone',
|
32 |
+
model_name='openai/clip-vit-base-patch32',
|
33 |
+
frozen_modules=['all'])),
|
34 |
+
neck=dict(type='YOLOWorldDualPAFPN',
|
35 |
+
guide_channels=text_channels,
|
36 |
+
embed_channels=neck_embed_channels,
|
37 |
+
num_heads=neck_num_heads,
|
38 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
|
39 |
+
text_enhancder=dict(type='ImagePoolingAttentionModule',
|
40 |
+
embed_channels=256,
|
41 |
+
num_heads=8)),
|
42 |
+
bbox_head=dict(type='YOLOWorldHead',
|
43 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
44 |
+
embed_dims=text_channels,
|
45 |
+
num_classes=num_training_classes)),
|
46 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
47 |
+
|
48 |
+
# dataset settings
|
49 |
+
text_transform = [
|
50 |
+
dict(type='RandomLoadText',
|
51 |
+
num_neg_samples=(num_classes, num_classes),
|
52 |
+
max_num_samples=num_training_classes,
|
53 |
+
padding_to_max=True,
|
54 |
+
padding_value=''),
|
55 |
+
dict(type='mmdet.PackDetInputs',
|
56 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
57 |
+
'flip_direction', 'texts'))
|
58 |
+
]
|
59 |
+
train_pipeline = [
|
60 |
+
*_base_.pre_transform,
|
61 |
+
dict(type='MultiModalMosaic',
|
62 |
+
img_scale=_base_.img_scale,
|
63 |
+
pad_val=114.0,
|
64 |
+
pre_transform=_base_.pre_transform),
|
65 |
+
dict(
|
66 |
+
type='YOLOv5RandomAffine',
|
67 |
+
max_rotate_degree=0.0,
|
68 |
+
max_shear_degree=0.0,
|
69 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
70 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
71 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
72 |
+
border_val=(114, 114, 114)),
|
73 |
+
*_base_.last_transform[:-1],
|
74 |
+
*text_transform,
|
75 |
+
]
|
76 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
77 |
+
obj365v1_train_dataset = dict(
|
78 |
+
type='MultiModalDataset',
|
79 |
+
dataset=dict(
|
80 |
+
type='YOLOv5Objects365V1Dataset',
|
81 |
+
data_root='data/objects365v1/',
|
82 |
+
ann_file='annotations/objects365_train.json',
|
83 |
+
data_prefix=dict(img='train/'),
|
84 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
85 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
86 |
+
pipeline=train_pipeline)
|
87 |
+
|
88 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
89 |
+
data_root='data/mixed_grounding/',
|
90 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
91 |
+
data_prefix=dict(img='gqa/images/'),
|
92 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
93 |
+
pipeline=train_pipeline)
|
94 |
+
|
95 |
+
flickr_train_dataset = dict(
|
96 |
+
type='YOLOv5MixedGroundingDataset',
|
97 |
+
data_root='data/flickr/',
|
98 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
99 |
+
data_prefix=dict(img='full_images/'),
|
100 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
101 |
+
pipeline=train_pipeline)
|
102 |
+
|
103 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
104 |
+
collate_fn=dict(type='yolow_collate'),
|
105 |
+
dataset=dict(_delete_=True,
|
106 |
+
type='ConcatDataset',
|
107 |
+
datasets=[
|
108 |
+
obj365v1_train_dataset,
|
109 |
+
flickr_train_dataset, mg_train_dataset
|
110 |
+
],
|
111 |
+
ignore_keys=['classes', 'palette']))
|
112 |
+
|
113 |
+
test_pipeline = [
|
114 |
+
*_base_.test_pipeline[:-1],
|
115 |
+
dict(type='LoadText'),
|
116 |
+
dict(type='mmdet.PackDetInputs',
|
117 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
118 |
+
'scale_factor', 'pad_param', 'texts'))
|
119 |
+
]
|
120 |
+
coco_val_dataset = dict(
|
121 |
+
_delete_=True,
|
122 |
+
type='MultiModalDataset',
|
123 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
124 |
+
data_root='data/coco/',
|
125 |
+
test_mode=True,
|
126 |
+
ann_file='lvis/lvis_v1_val.json',
|
127 |
+
data_prefix=dict(img=''),
|
128 |
+
batch_shapes_cfg=None),
|
129 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
130 |
+
pipeline=test_pipeline)
|
131 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
132 |
+
test_dataloader = val_dataloader
|
133 |
+
|
134 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
135 |
+
ann_file='data/coco/lvis/lvis_v1_val.json',
|
136 |
+
metric='bbox')
|
137 |
+
test_evaluator = val_evaluator
|
138 |
+
|
139 |
+
# training settings
|
140 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
141 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
142 |
+
rule='greater'))
|
143 |
+
custom_hooks = [
|
144 |
+
dict(type='EMAHook',
|
145 |
+
ema_type='ExpMomentumEMA',
|
146 |
+
momentum=0.0001,
|
147 |
+
update_buffers=True,
|
148 |
+
strict_load=False,
|
149 |
+
priority=49),
|
150 |
+
dict(type='mmdet.PipelineSwitchHook',
|
151 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
152 |
+
switch_pipeline=train_pipeline_stage2)
|
153 |
+
]
|
154 |
+
train_cfg = dict(max_epochs=max_epochs,
|
155 |
+
val_interval=10,
|
156 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
157 |
+
_base_.val_interval_stage2)])
|
158 |
+
optim_wrapper = dict(optimizer=dict(
|
159 |
+
_delete_=True,
|
160 |
+
type='AdamW',
|
161 |
+
lr=base_lr,
|
162 |
+
weight_decay=weight_decay,
|
163 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
164 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
165 |
+
norm_decay_mult=0.0,
|
166 |
+
custom_keys={
|
167 |
+
'backbone.text_model':
|
168 |
+
dict(lr_mult=0.01),
|
169 |
+
'logit_scale':
|
170 |
+
dict(weight_decay=0.0)
|
171 |
+
}),
|
172 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain_v1/yolo_world_m_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,172 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_m_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
|
19 |
+
# model settings
|
20 |
+
model = dict(
|
21 |
+
type='YOLOWorldDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
26 |
+
backbone=dict(
|
27 |
+
_delete_=True,
|
28 |
+
type='MultiModalYOLOBackbone',
|
29 |
+
image_model={{_base_.model.backbone}},
|
30 |
+
text_model=dict(
|
31 |
+
type='HuggingCLIPLanguageBackbone',
|
32 |
+
model_name='openai/clip-vit-base-patch32',
|
33 |
+
frozen_modules=['all'])),
|
34 |
+
neck=dict(type='YOLOWorldDualPAFPN',
|
35 |
+
guide_channels=text_channels,
|
36 |
+
embed_channels=neck_embed_channels,
|
37 |
+
num_heads=neck_num_heads,
|
38 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
|
39 |
+
text_enhancder=dict(type='ImagePoolingAttentionModule',
|
40 |
+
embed_channels=256,
|
41 |
+
num_heads=8)),
|
42 |
+
bbox_head=dict(type='YOLOWorldHead',
|
43 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
44 |
+
embed_dims=text_channels,
|
45 |
+
num_classes=num_training_classes)),
|
46 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
47 |
+
|
48 |
+
# dataset settings
|
49 |
+
text_transform = [
|
50 |
+
dict(type='RandomLoadText',
|
51 |
+
num_neg_samples=(num_classes, num_classes),
|
52 |
+
max_num_samples=num_training_classes,
|
53 |
+
padding_to_max=True,
|
54 |
+
padding_value=''),
|
55 |
+
dict(type='mmdet.PackDetInputs',
|
56 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
57 |
+
'flip_direction', 'texts'))
|
58 |
+
]
|
59 |
+
train_pipeline = [
|
60 |
+
*_base_.pre_transform,
|
61 |
+
dict(type='MultiModalMosaic',
|
62 |
+
img_scale=_base_.img_scale,
|
63 |
+
pad_val=114.0,
|
64 |
+
pre_transform=_base_.pre_transform),
|
65 |
+
dict(
|
66 |
+
type='YOLOv5RandomAffine',
|
67 |
+
max_rotate_degree=0.0,
|
68 |
+
max_shear_degree=0.0,
|
69 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
70 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
71 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
72 |
+
border_val=(114, 114, 114)),
|
73 |
+
*_base_.last_transform[:-1],
|
74 |
+
*text_transform,
|
75 |
+
]
|
76 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
77 |
+
obj365v1_train_dataset = dict(
|
78 |
+
type='MultiModalDataset',
|
79 |
+
dataset=dict(
|
80 |
+
type='YOLOv5Objects365V1Dataset',
|
81 |
+
data_root='data/objects365v1/',
|
82 |
+
ann_file='annotations/objects365_train.json',
|
83 |
+
data_prefix=dict(img='train/'),
|
84 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
85 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
86 |
+
pipeline=train_pipeline)
|
87 |
+
|
88 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
89 |
+
data_root='data/mixed_grounding/',
|
90 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
91 |
+
data_prefix=dict(img='gqa/images/'),
|
92 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
93 |
+
pipeline=train_pipeline)
|
94 |
+
|
95 |
+
flickr_train_dataset = dict(
|
96 |
+
type='YOLOv5MixedGroundingDataset',
|
97 |
+
data_root='data/flickr/',
|
98 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
99 |
+
data_prefix=dict(img='full_images/'),
|
100 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
101 |
+
pipeline=train_pipeline)
|
102 |
+
|
103 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
104 |
+
collate_fn=dict(type='yolow_collate'),
|
105 |
+
dataset=dict(_delete_=True,
|
106 |
+
type='ConcatDataset',
|
107 |
+
datasets=[
|
108 |
+
obj365v1_train_dataset,
|
109 |
+
flickr_train_dataset, mg_train_dataset
|
110 |
+
],
|
111 |
+
ignore_keys=['classes', 'palette']))
|
112 |
+
|
113 |
+
test_pipeline = [
|
114 |
+
*_base_.test_pipeline[:-1],
|
115 |
+
dict(type='LoadText'),
|
116 |
+
dict(type='mmdet.PackDetInputs',
|
117 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
118 |
+
'scale_factor', 'pad_param', 'texts'))
|
119 |
+
]
|
120 |
+
coco_val_dataset = dict(
|
121 |
+
_delete_=True,
|
122 |
+
type='MultiModalDataset',
|
123 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
124 |
+
data_root='data/coco/',
|
125 |
+
test_mode=True,
|
126 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
127 |
+
data_prefix=dict(img=''),
|
128 |
+
batch_shapes_cfg=None),
|
129 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
130 |
+
pipeline=test_pipeline)
|
131 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
132 |
+
test_dataloader = val_dataloader
|
133 |
+
|
134 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
135 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
136 |
+
metric='bbox')
|
137 |
+
test_evaluator = val_evaluator
|
138 |
+
|
139 |
+
# training settings
|
140 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
141 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
142 |
+
rule='greater'))
|
143 |
+
custom_hooks = [
|
144 |
+
dict(type='EMAHook',
|
145 |
+
ema_type='ExpMomentumEMA',
|
146 |
+
momentum=0.0001,
|
147 |
+
update_buffers=True,
|
148 |
+
strict_load=False,
|
149 |
+
priority=49),
|
150 |
+
dict(type='mmdet.PipelineSwitchHook',
|
151 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
152 |
+
switch_pipeline=train_pipeline_stage2)
|
153 |
+
]
|
154 |
+
train_cfg = dict(max_epochs=max_epochs,
|
155 |
+
val_interval=10,
|
156 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
157 |
+
_base_.val_interval_stage2)])
|
158 |
+
optim_wrapper = dict(optimizer=dict(
|
159 |
+
_delete_=True,
|
160 |
+
type='AdamW',
|
161 |
+
lr=base_lr,
|
162 |
+
weight_decay=weight_decay,
|
163 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
164 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
165 |
+
norm_decay_mult=0.0,
|
166 |
+
custom_keys={
|
167 |
+
'backbone.text_model':
|
168 |
+
dict(lr_mult=0.01),
|
169 |
+
'logit_scale':
|
170 |
+
dict(weight_decay=0.0)
|
171 |
+
}),
|
172 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain_v1/yolo_world_s_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,172 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_s_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
|
19 |
+
# model settings
|
20 |
+
model = dict(
|
21 |
+
type='YOLOWorldDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
26 |
+
backbone=dict(
|
27 |
+
_delete_=True,
|
28 |
+
type='MultiModalYOLOBackbone',
|
29 |
+
image_model={{_base_.model.backbone}},
|
30 |
+
text_model=dict(
|
31 |
+
type='HuggingCLIPLanguageBackbone',
|
32 |
+
model_name='openai/clip-vit-base-patch32',
|
33 |
+
frozen_modules=['all'])),
|
34 |
+
neck=dict(type='YOLOWorldDualPAFPN',
|
35 |
+
guide_channels=text_channels,
|
36 |
+
embed_channels=neck_embed_channels,
|
37 |
+
num_heads=neck_num_heads,
|
38 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
|
39 |
+
text_enhancder=dict(type='ImagePoolingAttentionModule',
|
40 |
+
embed_channels=256,
|
41 |
+
num_heads=8)),
|
42 |
+
bbox_head=dict(type='YOLOWorldHead',
|
43 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
44 |
+
embed_dims=text_channels,
|
45 |
+
num_classes=num_training_classes)),
|
46 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
47 |
+
|
48 |
+
# dataset settings
|
49 |
+
text_transform = [
|
50 |
+
dict(type='RandomLoadText',
|
51 |
+
num_neg_samples=(num_classes, num_classes),
|
52 |
+
max_num_samples=num_training_classes,
|
53 |
+
padding_to_max=True,
|
54 |
+
padding_value=''),
|
55 |
+
dict(type='mmdet.PackDetInputs',
|
56 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
57 |
+
'flip_direction', 'texts'))
|
58 |
+
]
|
59 |
+
train_pipeline = [
|
60 |
+
*_base_.pre_transform,
|
61 |
+
dict(type='MultiModalMosaic',
|
62 |
+
img_scale=_base_.img_scale,
|
63 |
+
pad_val=114.0,
|
64 |
+
pre_transform=_base_.pre_transform),
|
65 |
+
dict(
|
66 |
+
type='YOLOv5RandomAffine',
|
67 |
+
max_rotate_degree=0.0,
|
68 |
+
max_shear_degree=0.0,
|
69 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
70 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
71 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
72 |
+
border_val=(114, 114, 114)),
|
73 |
+
*_base_.last_transform[:-1],
|
74 |
+
*text_transform,
|
75 |
+
]
|
76 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
77 |
+
obj365v1_train_dataset = dict(
|
78 |
+
type='MultiModalDataset',
|
79 |
+
dataset=dict(
|
80 |
+
type='YOLOv5Objects365V1Dataset',
|
81 |
+
data_root='data/objects365v1/',
|
82 |
+
ann_file='annotations/objects365_train.json',
|
83 |
+
data_prefix=dict(img='train/'),
|
84 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
85 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
86 |
+
pipeline=train_pipeline)
|
87 |
+
|
88 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
89 |
+
data_root='data/mixed_grounding/',
|
90 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
91 |
+
data_prefix=dict(img='gqa/images/'),
|
92 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
93 |
+
pipeline=train_pipeline)
|
94 |
+
|
95 |
+
flickr_train_dataset = dict(
|
96 |
+
type='YOLOv5MixedGroundingDataset',
|
97 |
+
data_root='data/flickr/',
|
98 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
99 |
+
data_prefix=dict(img='full_images/'),
|
100 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
101 |
+
pipeline=train_pipeline)
|
102 |
+
|
103 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
104 |
+
collate_fn=dict(type='yolow_collate'),
|
105 |
+
dataset=dict(_delete_=True,
|
106 |
+
type='ConcatDataset',
|
107 |
+
datasets=[
|
108 |
+
obj365v1_train_dataset,
|
109 |
+
flickr_train_dataset, mg_train_dataset
|
110 |
+
],
|
111 |
+
ignore_keys=['classes', 'palette']))
|
112 |
+
|
113 |
+
test_pipeline = [
|
114 |
+
*_base_.test_pipeline[:-1],
|
115 |
+
dict(type='LoadText'),
|
116 |
+
dict(type='mmdet.PackDetInputs',
|
117 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
118 |
+
'scale_factor', 'pad_param', 'texts'))
|
119 |
+
]
|
120 |
+
coco_val_dataset = dict(
|
121 |
+
_delete_=True,
|
122 |
+
type='MultiModalDataset',
|
123 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
124 |
+
data_root='data/coco/',
|
125 |
+
test_mode=True,
|
126 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
127 |
+
data_prefix=dict(img=''),
|
128 |
+
batch_shapes_cfg=None),
|
129 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
130 |
+
pipeline=test_pipeline)
|
131 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
132 |
+
test_dataloader = val_dataloader
|
133 |
+
|
134 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
135 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
136 |
+
metric='bbox')
|
137 |
+
test_evaluator = val_evaluator
|
138 |
+
|
139 |
+
# training settings
|
140 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
141 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
142 |
+
rule='greater'))
|
143 |
+
custom_hooks = [
|
144 |
+
dict(type='EMAHook',
|
145 |
+
ema_type='ExpMomentumEMA',
|
146 |
+
momentum=0.0001,
|
147 |
+
update_buffers=True,
|
148 |
+
strict_load=False,
|
149 |
+
priority=49),
|
150 |
+
dict(type='mmdet.PipelineSwitchHook',
|
151 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
152 |
+
switch_pipeline=train_pipeline_stage2)
|
153 |
+
]
|
154 |
+
train_cfg = dict(max_epochs=max_epochs,
|
155 |
+
val_interval=10,
|
156 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
157 |
+
_base_.val_interval_stage2)])
|
158 |
+
optim_wrapper = dict(optimizer=dict(
|
159 |
+
_delete_=True,
|
160 |
+
type='AdamW',
|
161 |
+
lr=base_lr,
|
162 |
+
weight_decay=weight_decay,
|
163 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
164 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
165 |
+
norm_decay_mult=0.0,
|
166 |
+
custom_keys={
|
167 |
+
'backbone.text_model':
|
168 |
+
dict(lr_mult=0.01),
|
169 |
+
'logit_scale':
|
170 |
+
dict(weight_decay=0.0)
|
171 |
+
}),
|
172 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/pretrain_v1/yolo_world_x_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py
ADDED
@@ -0,0 +1,172 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_x_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'],
|
4 |
+
allow_failed_imports=False)
|
5 |
+
|
6 |
+
# hyper-parameters
|
7 |
+
num_classes = 1203
|
8 |
+
num_training_classes = 80
|
9 |
+
max_epochs = 100 # Maximum training epochs
|
10 |
+
close_mosaic_epochs = 2
|
11 |
+
save_epoch_intervals = 2
|
12 |
+
text_channels = 512
|
13 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
14 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
15 |
+
base_lr = 2e-3
|
16 |
+
weight_decay = 0.05 / 2
|
17 |
+
train_batch_size_per_gpu = 16
|
18 |
+
|
19 |
+
# model settings
|
20 |
+
model = dict(
|
21 |
+
type='YOLOWorldDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
26 |
+
backbone=dict(
|
27 |
+
_delete_=True,
|
28 |
+
type='MultiModalYOLOBackbone',
|
29 |
+
image_model={{_base_.model.backbone}},
|
30 |
+
text_model=dict(
|
31 |
+
type='HuggingCLIPLanguageBackbone',
|
32 |
+
model_name='openai/clip-vit-base-patch32',
|
33 |
+
frozen_modules=['all'])),
|
34 |
+
neck=dict(type='YOLOWorldDualPAFPN',
|
35 |
+
guide_channels=text_channels,
|
36 |
+
embed_channels=neck_embed_channels,
|
37 |
+
num_heads=neck_num_heads,
|
38 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
|
39 |
+
text_enhancder=dict(type='ImagePoolingAttentionModule',
|
40 |
+
embed_channels=256,
|
41 |
+
num_heads=8)),
|
42 |
+
bbox_head=dict(type='YOLOWorldHead',
|
43 |
+
head_module=dict(type='YOLOWorldHeadModule',
|
44 |
+
embed_dims=text_channels,
|
45 |
+
num_classes=num_training_classes)),
|
46 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
47 |
+
|
48 |
+
# dataset settings
|
49 |
+
text_transform = [
|
50 |
+
dict(type='RandomLoadText',
|
51 |
+
num_neg_samples=(num_classes, num_classes),
|
52 |
+
max_num_samples=num_training_classes,
|
53 |
+
padding_to_max=True,
|
54 |
+
padding_value=''),
|
55 |
+
dict(type='mmdet.PackDetInputs',
|
56 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
57 |
+
'flip_direction', 'texts'))
|
58 |
+
]
|
59 |
+
train_pipeline = [
|
60 |
+
*_base_.pre_transform,
|
61 |
+
dict(type='MultiModalMosaic',
|
62 |
+
img_scale=_base_.img_scale,
|
63 |
+
pad_val=114.0,
|
64 |
+
pre_transform=_base_.pre_transform),
|
65 |
+
dict(
|
66 |
+
type='YOLOv5RandomAffine',
|
67 |
+
max_rotate_degree=0.0,
|
68 |
+
max_shear_degree=0.0,
|
69 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
70 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
71 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
72 |
+
border_val=(114, 114, 114)),
|
73 |
+
*_base_.last_transform[:-1],
|
74 |
+
*text_transform,
|
75 |
+
]
|
76 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
|
77 |
+
obj365v1_train_dataset = dict(
|
78 |
+
type='MultiModalDataset',
|
79 |
+
dataset=dict(
|
80 |
+
type='YOLOv5Objects365V1Dataset',
|
81 |
+
data_root='data/objects365v1/',
|
82 |
+
ann_file='annotations/objects365_train.json',
|
83 |
+
data_prefix=dict(img='train/'),
|
84 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32)),
|
85 |
+
class_text_path='data/texts/obj365v1_class_texts.json',
|
86 |
+
pipeline=train_pipeline)
|
87 |
+
|
88 |
+
mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
|
89 |
+
data_root='data/mixed_grounding/',
|
90 |
+
ann_file='annotations/final_mixed_train_no_coco.json',
|
91 |
+
data_prefix=dict(img='gqa/images/'),
|
92 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
93 |
+
pipeline=train_pipeline)
|
94 |
+
|
95 |
+
flickr_train_dataset = dict(
|
96 |
+
type='YOLOv5MixedGroundingDataset',
|
97 |
+
data_root='data/flickr/',
|
98 |
+
ann_file='annotations/final_flickr_separateGT_train.json',
|
99 |
+
data_prefix=dict(img='full_images/'),
|
100 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32),
|
101 |
+
pipeline=train_pipeline)
|
102 |
+
|
103 |
+
train_dataloader = dict(batch_size=train_batch_size_per_gpu,
|
104 |
+
collate_fn=dict(type='yolow_collate'),
|
105 |
+
dataset=dict(_delete_=True,
|
106 |
+
type='ConcatDataset',
|
107 |
+
datasets=[
|
108 |
+
obj365v1_train_dataset,
|
109 |
+
flickr_train_dataset, mg_train_dataset
|
110 |
+
],
|
111 |
+
ignore_keys=['classes', 'palette']))
|
112 |
+
|
113 |
+
test_pipeline = [
|
114 |
+
*_base_.test_pipeline[:-1],
|
115 |
+
dict(type='LoadText'),
|
116 |
+
dict(type='mmdet.PackDetInputs',
|
117 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
118 |
+
'scale_factor', 'pad_param', 'texts'))
|
119 |
+
]
|
120 |
+
coco_val_dataset = dict(
|
121 |
+
_delete_=True,
|
122 |
+
type='MultiModalDataset',
|
123 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
124 |
+
data_root='data/coco/',
|
125 |
+
test_mode=True,
|
126 |
+
ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
|
127 |
+
data_prefix=dict(img=''),
|
128 |
+
batch_shapes_cfg=None),
|
129 |
+
class_text_path='data/texts/lvis_v1_class_texts.json',
|
130 |
+
pipeline=test_pipeline)
|
131 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
132 |
+
test_dataloader = val_dataloader
|
133 |
+
|
134 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
135 |
+
ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
|
136 |
+
metric='bbox')
|
137 |
+
test_evaluator = val_evaluator
|
138 |
+
|
139 |
+
# training settings
|
140 |
+
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
|
141 |
+
checkpoint=dict(interval=save_epoch_intervals,
|
142 |
+
rule='greater'))
|
143 |
+
custom_hooks = [
|
144 |
+
dict(type='EMAHook',
|
145 |
+
ema_type='ExpMomentumEMA',
|
146 |
+
momentum=0.0001,
|
147 |
+
update_buffers=True,
|
148 |
+
strict_load=False,
|
149 |
+
priority=49),
|
150 |
+
dict(type='mmdet.PipelineSwitchHook',
|
151 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
152 |
+
switch_pipeline=train_pipeline_stage2)
|
153 |
+
]
|
154 |
+
train_cfg = dict(max_epochs=max_epochs,
|
155 |
+
val_interval=10,
|
156 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
157 |
+
_base_.val_interval_stage2)])
|
158 |
+
optim_wrapper = dict(optimizer=dict(
|
159 |
+
_delete_=True,
|
160 |
+
type='AdamW',
|
161 |
+
lr=base_lr,
|
162 |
+
weight_decay=weight_decay,
|
163 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
164 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
165 |
+
norm_decay_mult=0.0,
|
166 |
+
custom_keys={
|
167 |
+
'backbone.text_model':
|
168 |
+
dict(lr_mult=0.01),
|
169 |
+
'logit_scale':
|
170 |
+
dict(weight_decay=0.0)
|
171 |
+
}),
|
172 |
+
constructor='YOLOWv5OptimizerConstructor')
|
configs/prompt_tuning_coco/READEME.md
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## Prompt Tuning for YOLO-World
|
2 |
+
|
3 |
+
### NOTE:
|
4 |
+
|
5 |
+
This folder contains many experimental config files, which will be removed later!!
|
6 |
+
|
7 |
+
### Experimental Results
|
8 |
+
|
9 |
+
| Model | Config | AP | AP50 | AP75 | APS | APM | APL |
|
10 |
+
| :---- | :----: | :--: | :--: | :---: | :-: | :-: | :-: |
|
11 |
+
| YOLO-World-v2-L | Zero-shot | 45.7 | 61.6 | 49.8 | 29.9 | 50.0 | 60.8 |
|
12 |
+
| [YOLO-World-v2-L](./../configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_prompt_tuning_coco.py) | Prompt tuning | 47.9 | 64.3 | 52.5 | 31.9 | 52.6 | 61.3 |
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_all_fine_tuning_coco.py
ADDED
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 2e-4
|
15 |
+
weight_decay = 0.05
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
freeze_prompt=True,
|
29 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
30 |
+
backbone=dict(_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
text_model=None,
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=False,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=False,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
52 |
+
data_root='data/coco',
|
53 |
+
ann_file='annotations/instances_train2017.json',
|
54 |
+
data_prefix=dict(img='train2017/'),
|
55 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
56 |
+
pipeline=_base_.train_pipeline)
|
57 |
+
|
58 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
59 |
+
batch_size=train_batch_size_per_gpu,
|
60 |
+
collate_fn=dict(type='yolow_collate'),
|
61 |
+
dataset=coco_train_dataset)
|
62 |
+
|
63 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
64 |
+
data_root='data/coco',
|
65 |
+
ann_file='annotations/instances_val2017.json',
|
66 |
+
data_prefix=dict(img='val2017/'),
|
67 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
68 |
+
pipeline=_base_.test_pipeline)
|
69 |
+
|
70 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
71 |
+
test_dataloader = val_dataloader
|
72 |
+
# training settings
|
73 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
74 |
+
lr_factor=0.01,
|
75 |
+
max_epochs=max_epochs),
|
76 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
77 |
+
save_best=None,
|
78 |
+
interval=save_epoch_intervals))
|
79 |
+
custom_hooks = [
|
80 |
+
dict(type='EMAHook',
|
81 |
+
ema_type='ExpMomentumEMA',
|
82 |
+
momentum=0.0001,
|
83 |
+
update_buffers=True,
|
84 |
+
strict_load=False,
|
85 |
+
priority=49),
|
86 |
+
dict(type='mmdet.PipelineSwitchHook',
|
87 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
88 |
+
switch_pipeline=_base_.train_pipeline_stage2)
|
89 |
+
]
|
90 |
+
train_cfg = dict(max_epochs=max_epochs,
|
91 |
+
val_interval=5,
|
92 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
93 |
+
_base_.val_interval_stage2)])
|
94 |
+
|
95 |
+
optim_wrapper = dict(optimizer=dict(
|
96 |
+
_delete_=True,
|
97 |
+
type='AdamW',
|
98 |
+
lr=base_lr,
|
99 |
+
weight_decay=weight_decay,
|
100 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
101 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
102 |
+
norm_decay_mult=0.0,
|
103 |
+
custom_keys={
|
104 |
+
'backbone.text_model':
|
105 |
+
dict(lr_mult=0.01),
|
106 |
+
'logit_scale':
|
107 |
+
dict(weight_decay=0.0),
|
108 |
+
'embeddings':
|
109 |
+
dict(weight_decay=0.0)
|
110 |
+
}),
|
111 |
+
constructor='YOLOWv5OptimizerConstructor')
|
112 |
+
|
113 |
+
# evaluation settings
|
114 |
+
val_evaluator = dict(_delete_=True,
|
115 |
+
type='mmdet.CocoMetric',
|
116 |
+
proposal_nums=(100, 1, 10),
|
117 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
118 |
+
metric='bbox')
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_all_fine_tuning_rmdecay_rmmosaic_coco.py
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 70
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 2e-4
|
15 |
+
weight_decay = 0.05
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
freeze_prompt=True,
|
29 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
30 |
+
backbone=dict(_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
text_model=None,
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=False,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=False,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
52 |
+
data_root='data/coco',
|
53 |
+
ann_file='annotations/instances_train2017.json',
|
54 |
+
data_prefix=dict(img='train2017/'),
|
55 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
56 |
+
pipeline=_base_.train_pipeline)
|
57 |
+
|
58 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
59 |
+
batch_size=train_batch_size_per_gpu,
|
60 |
+
collate_fn=dict(type='yolow_collate'),
|
61 |
+
dataset=coco_train_dataset)
|
62 |
+
|
63 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
64 |
+
data_root='data/coco',
|
65 |
+
ann_file='annotations/instances_val2017.json',
|
66 |
+
data_prefix=dict(img='val2017/'),
|
67 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
68 |
+
pipeline=_base_.test_pipeline)
|
69 |
+
|
70 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
71 |
+
test_dataloader = val_dataloader
|
72 |
+
# training settings
|
73 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
74 |
+
lr_factor=0.01,
|
75 |
+
max_epochs=max_epochs),
|
76 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
77 |
+
save_best=None,
|
78 |
+
interval=save_epoch_intervals))
|
79 |
+
custom_hooks = [
|
80 |
+
dict(type='EMAHook',
|
81 |
+
ema_type='ExpMomentumEMA',
|
82 |
+
momentum=0.0001,
|
83 |
+
update_buffers=True,
|
84 |
+
strict_load=False,
|
85 |
+
priority=49),
|
86 |
+
dict(type='mmdet.PipelineSwitchHook',
|
87 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
88 |
+
switch_pipeline=_base_.train_pipeline_stage2)
|
89 |
+
]
|
90 |
+
train_cfg = dict(max_epochs=max_epochs,
|
91 |
+
val_interval=5,
|
92 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
93 |
+
_base_.val_interval_stage2)])
|
94 |
+
|
95 |
+
optim_wrapper = dict(optimizer=dict(
|
96 |
+
_delete_=True,
|
97 |
+
type='AdamW',
|
98 |
+
lr=base_lr,
|
99 |
+
weight_decay=weight_decay,
|
100 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
101 |
+
paramwise_cfg=dict(
|
102 |
+
custom_keys={
|
103 |
+
'backbone.text_model': dict(lr_mult=0.01),
|
104 |
+
'logit_scale': dict(weight_decay=0.0),
|
105 |
+
'embeddings': dict(weight_decay=0.0)
|
106 |
+
}),
|
107 |
+
constructor='YOLOWv5OptimizerConstructor')
|
108 |
+
|
109 |
+
# evaluation settings
|
110 |
+
val_evaluator = dict(_delete_=True,
|
111 |
+
type='mmdet.CocoMetric',
|
112 |
+
proposal_nums=(100, 1, 10),
|
113 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
114 |
+
metric='bbox')
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_all_fine_tuning_coco.py
ADDED
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 2e-4
|
15 |
+
weight_decay = 0.05
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
freeze_prompt=True,
|
29 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
30 |
+
backbone=dict(_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
text_model=None,
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=False,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=False,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
final_transform = [
|
52 |
+
dict(type='mmdet.PackDetInputs',
|
53 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
54 |
+
'flip_direction'))
|
55 |
+
]
|
56 |
+
mosaic_affine_transform = [
|
57 |
+
dict(type='Mosaic',
|
58 |
+
img_scale=_base_.img_scale,
|
59 |
+
pad_val=114.0,
|
60 |
+
pre_transform=_base_.pre_transform),
|
61 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
62 |
+
dict(
|
63 |
+
type='YOLOv5RandomAffine',
|
64 |
+
max_rotate_degree=0.0,
|
65 |
+
max_shear_degree=0.0,
|
66 |
+
max_aspect_ratio=100.,
|
67 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
68 |
+
# img_scale is (width, height)
|
69 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
70 |
+
border_val=(114, 114, 114),
|
71 |
+
min_area_ratio=_base_.min_area_ratio,
|
72 |
+
use_mask_refine=_base_.use_mask2refine)
|
73 |
+
]
|
74 |
+
train_pipeline = [
|
75 |
+
*_base_.pre_transform, *mosaic_affine_transform,
|
76 |
+
dict(type='YOLOv5MixUp',
|
77 |
+
prob=_base_.mixup_prob,
|
78 |
+
pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
|
79 |
+
*_base_.last_transform[:-1], *final_transform
|
80 |
+
]
|
81 |
+
|
82 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *final_transform]
|
83 |
+
|
84 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
85 |
+
data_root='data/coco',
|
86 |
+
ann_file='annotations/instances_train2017.json',
|
87 |
+
data_prefix=dict(img='train2017/'),
|
88 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
89 |
+
pipeline=train_pipeline)
|
90 |
+
|
91 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
92 |
+
batch_size=train_batch_size_per_gpu,
|
93 |
+
collate_fn=dict(type='yolow_collate'),
|
94 |
+
dataset=coco_train_dataset)
|
95 |
+
|
96 |
+
test_pipeline = [
|
97 |
+
*_base_.test_pipeline[:-1],
|
98 |
+
dict(type='mmdet.PackDetInputs',
|
99 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
100 |
+
'scale_factor', 'pad_param'))
|
101 |
+
]
|
102 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
103 |
+
data_root='data/coco',
|
104 |
+
ann_file='annotations/instances_val2017.json',
|
105 |
+
data_prefix=dict(img='val2017/'),
|
106 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
107 |
+
pipeline=test_pipeline)
|
108 |
+
|
109 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
110 |
+
test_dataloader = val_dataloader
|
111 |
+
# training settings
|
112 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
113 |
+
lr_factor=0.01,
|
114 |
+
max_epochs=max_epochs),
|
115 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
116 |
+
save_best=None,
|
117 |
+
interval=save_epoch_intervals))
|
118 |
+
custom_hooks = [
|
119 |
+
dict(type='EMAHook',
|
120 |
+
ema_type='ExpMomentumEMA',
|
121 |
+
momentum=0.0001,
|
122 |
+
update_buffers=True,
|
123 |
+
strict_load=False,
|
124 |
+
priority=49),
|
125 |
+
dict(type='mmdet.PipelineSwitchHook',
|
126 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
127 |
+
switch_pipeline=train_pipeline_stage2)
|
128 |
+
]
|
129 |
+
train_cfg = dict(max_epochs=max_epochs,
|
130 |
+
val_interval=5,
|
131 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
132 |
+
_base_.val_interval_stage2)])
|
133 |
+
optim_wrapper = dict(optimizer=dict(
|
134 |
+
_delete_=True,
|
135 |
+
type='AdamW',
|
136 |
+
lr=base_lr,
|
137 |
+
weight_decay=weight_decay,
|
138 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
139 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
140 |
+
norm_decay_mult=0.0,
|
141 |
+
custom_keys={
|
142 |
+
'backbone.text_model':
|
143 |
+
dict(lr_mult=0.01),
|
144 |
+
'logit_scale':
|
145 |
+
dict(weight_decay=0.0),
|
146 |
+
'embeddings':
|
147 |
+
dict(weight_decay=0.0)
|
148 |
+
}),
|
149 |
+
constructor='YOLOWv5OptimizerConstructor')
|
150 |
+
|
151 |
+
# evaluation settings
|
152 |
+
val_evaluator = dict(_delete_=True,
|
153 |
+
type='mmdet.CocoMetric',
|
154 |
+
proposal_nums=(100, 1, 10),
|
155 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
156 |
+
metric='bbox')
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_fine_prompt_tuning_coco.py
ADDED
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 2e-4
|
15 |
+
weight_decay = 0.05
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
freeze_prompt=False,
|
29 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
30 |
+
backbone=dict(_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
text_model=None,
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=False,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=False,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
final_transform = [
|
52 |
+
dict(type='mmdet.PackDetInputs',
|
53 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
54 |
+
'flip_direction'))
|
55 |
+
]
|
56 |
+
mosaic_affine_transform = [
|
57 |
+
dict(type='Mosaic',
|
58 |
+
img_scale=_base_.img_scale,
|
59 |
+
pad_val=114.0,
|
60 |
+
pre_transform=_base_.pre_transform),
|
61 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
62 |
+
dict(
|
63 |
+
type='YOLOv5RandomAffine',
|
64 |
+
max_rotate_degree=0.0,
|
65 |
+
max_shear_degree=0.0,
|
66 |
+
max_aspect_ratio=100.,
|
67 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
68 |
+
# img_scale is (width, height)
|
69 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
70 |
+
border_val=(114, 114, 114),
|
71 |
+
min_area_ratio=_base_.min_area_ratio,
|
72 |
+
use_mask_refine=_base_.use_mask2refine)
|
73 |
+
]
|
74 |
+
train_pipeline = [
|
75 |
+
*_base_.pre_transform, *mosaic_affine_transform,
|
76 |
+
dict(type='YOLOv5MixUp',
|
77 |
+
prob=_base_.mixup_prob,
|
78 |
+
pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
|
79 |
+
*_base_.last_transform[:-1], *final_transform
|
80 |
+
]
|
81 |
+
|
82 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *final_transform]
|
83 |
+
|
84 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
85 |
+
data_root='data/coco',
|
86 |
+
ann_file='annotations/instances_train2017.json',
|
87 |
+
data_prefix=dict(img='train2017/'),
|
88 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
89 |
+
pipeline=train_pipeline)
|
90 |
+
|
91 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
92 |
+
batch_size=train_batch_size_per_gpu,
|
93 |
+
collate_fn=dict(type='yolow_collate'),
|
94 |
+
dataset=coco_train_dataset)
|
95 |
+
|
96 |
+
test_pipeline = [
|
97 |
+
*_base_.test_pipeline[:-1],
|
98 |
+
dict(type='mmdet.PackDetInputs',
|
99 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
100 |
+
'scale_factor', 'pad_param'))
|
101 |
+
]
|
102 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
103 |
+
data_root='data/coco',
|
104 |
+
ann_file='annotations/instances_val2017.json',
|
105 |
+
data_prefix=dict(img='val2017/'),
|
106 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
107 |
+
pipeline=test_pipeline)
|
108 |
+
|
109 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
110 |
+
test_dataloader = val_dataloader
|
111 |
+
# training settings
|
112 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
113 |
+
lr_factor=0.01,
|
114 |
+
max_epochs=max_epochs),
|
115 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
116 |
+
save_best=None,
|
117 |
+
interval=save_epoch_intervals))
|
118 |
+
custom_hooks = [
|
119 |
+
dict(type='EMAHook',
|
120 |
+
ema_type='ExpMomentumEMA',
|
121 |
+
momentum=0.0001,
|
122 |
+
update_buffers=True,
|
123 |
+
strict_load=False,
|
124 |
+
priority=49),
|
125 |
+
dict(type='mmdet.PipelineSwitchHook',
|
126 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
127 |
+
switch_pipeline=train_pipeline_stage2)
|
128 |
+
]
|
129 |
+
train_cfg = dict(max_epochs=max_epochs,
|
130 |
+
val_interval=5,
|
131 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
132 |
+
_base_.val_interval_stage2)])
|
133 |
+
optim_wrapper = dict(optimizer=dict(
|
134 |
+
_delete_=True,
|
135 |
+
type='AdamW',
|
136 |
+
lr=base_lr,
|
137 |
+
weight_decay=weight_decay,
|
138 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
139 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
140 |
+
norm_decay_mult=0.0,
|
141 |
+
custom_keys={
|
142 |
+
'backbone.text_model':
|
143 |
+
dict(lr_mult=0.01),
|
144 |
+
'logit_scale':
|
145 |
+
dict(weight_decay=0.0),
|
146 |
+
'embeddings':
|
147 |
+
dict(weight_decay=0.0)
|
148 |
+
}),
|
149 |
+
constructor='YOLOWv5OptimizerConstructor')
|
150 |
+
|
151 |
+
# evaluation settings
|
152 |
+
val_evaluator = dict(_delete_=True,
|
153 |
+
type='mmdet.CocoMetric',
|
154 |
+
proposal_nums=(100, 1, 10),
|
155 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
156 |
+
metric='bbox')
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_prompt_tuning_coco.py
ADDED
@@ -0,0 +1,161 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 2e-3
|
15 |
+
weight_decay = 0.05
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
29 |
+
backbone=dict(_delete_=True,
|
30 |
+
type='MultiModalYOLOBackbone',
|
31 |
+
text_model=None,
|
32 |
+
image_model={{_base_.model.backbone}},
|
33 |
+
frozen_stages=4,
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=True,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=True,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
final_transform = [
|
52 |
+
dict(type='mmdet.PackDetInputs',
|
53 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
54 |
+
'flip_direction'))
|
55 |
+
]
|
56 |
+
mosaic_affine_transform = [
|
57 |
+
dict(type='Mosaic',
|
58 |
+
img_scale=_base_.img_scale,
|
59 |
+
pad_val=114.0,
|
60 |
+
pre_transform=_base_.pre_transform),
|
61 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
62 |
+
dict(
|
63 |
+
type='YOLOv5RandomAffine',
|
64 |
+
max_rotate_degree=0.0,
|
65 |
+
max_shear_degree=0.0,
|
66 |
+
max_aspect_ratio=100.,
|
67 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
68 |
+
# img_scale is (width, height)
|
69 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
70 |
+
border_val=(114, 114, 114),
|
71 |
+
min_area_ratio=_base_.min_area_ratio,
|
72 |
+
use_mask_refine=_base_.use_mask2refine)
|
73 |
+
]
|
74 |
+
train_pipeline = [
|
75 |
+
*_base_.pre_transform, *mosaic_affine_transform,
|
76 |
+
dict(type='YOLOv5MixUp',
|
77 |
+
prob=_base_.mixup_prob,
|
78 |
+
pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
|
79 |
+
*_base_.last_transform[:-1], *final_transform
|
80 |
+
]
|
81 |
+
|
82 |
+
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *final_transform]
|
83 |
+
|
84 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
85 |
+
data_root='data/coco',
|
86 |
+
ann_file='annotations/instances_train2017.json',
|
87 |
+
data_prefix=dict(img='train2017/'),
|
88 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
89 |
+
pipeline=train_pipeline)
|
90 |
+
|
91 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
92 |
+
batch_size=train_batch_size_per_gpu,
|
93 |
+
collate_fn=dict(type='yolow_collate'),
|
94 |
+
dataset=coco_train_dataset)
|
95 |
+
|
96 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
97 |
+
batch_size=train_batch_size_per_gpu,
|
98 |
+
collate_fn=dict(type='yolow_collate'),
|
99 |
+
dataset=coco_train_dataset)
|
100 |
+
test_pipeline = [
|
101 |
+
*_base_.test_pipeline[:-1],
|
102 |
+
dict(type='mmdet.PackDetInputs',
|
103 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
104 |
+
'scale_factor', 'pad_param'))
|
105 |
+
]
|
106 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
107 |
+
data_root='data/coco',
|
108 |
+
ann_file='annotations/instances_val2017.json',
|
109 |
+
data_prefix=dict(img='val2017/'),
|
110 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
111 |
+
pipeline=test_pipeline)
|
112 |
+
|
113 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
114 |
+
test_dataloader = val_dataloader
|
115 |
+
# training settings
|
116 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
117 |
+
lr_factor=0.01,
|
118 |
+
max_epochs=max_epochs),
|
119 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
120 |
+
save_best=None,
|
121 |
+
interval=save_epoch_intervals))
|
122 |
+
custom_hooks = [
|
123 |
+
dict(type='EMAHook',
|
124 |
+
ema_type='ExpMomentumEMA',
|
125 |
+
momentum=0.0001,
|
126 |
+
update_buffers=True,
|
127 |
+
strict_load=False,
|
128 |
+
priority=49),
|
129 |
+
dict(type='mmdet.PipelineSwitchHook',
|
130 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
131 |
+
switch_pipeline=train_pipeline_stage2)
|
132 |
+
]
|
133 |
+
train_cfg = dict(max_epochs=max_epochs,
|
134 |
+
val_interval=5,
|
135 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
136 |
+
_base_.val_interval_stage2)])
|
137 |
+
optim_wrapper = dict(optimizer=dict(
|
138 |
+
_delete_=True,
|
139 |
+
type='AdamW',
|
140 |
+
lr=base_lr,
|
141 |
+
weight_decay=weight_decay,
|
142 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
143 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
144 |
+
norm_decay_mult=0.0,
|
145 |
+
custom_keys={
|
146 |
+
'backbone.text_model':
|
147 |
+
dict(lr_mult=0.01),
|
148 |
+
'logit_scale':
|
149 |
+
dict(weight_decay=0.0),
|
150 |
+
'embeddings':
|
151 |
+
dict(weight_decay=0.0)
|
152 |
+
}),
|
153 |
+
constructor='YOLOWv5OptimizerConstructor')
|
154 |
+
|
155 |
+
# evaluation settings
|
156 |
+
val_evaluator = dict(_delete_=True,
|
157 |
+
type='mmdet.CocoMetric',
|
158 |
+
proposal_nums=(100, 1, 10),
|
159 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
160 |
+
metric='bbox')
|
161 |
+
find_unused_parameters = True
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_20e_8gpus_all_fine_tuning_rmdecay_coco.py
ADDED
@@ -0,0 +1,113 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 20 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 1e-3
|
15 |
+
weight_decay = 0.0005
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
freeze_prompt=True,
|
29 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
30 |
+
backbone=dict(_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
text_model=None,
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=False,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=False,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
52 |
+
data_root='data/coco',
|
53 |
+
ann_file='annotations/instances_train2017.json',
|
54 |
+
data_prefix=dict(img='train2017/'),
|
55 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
56 |
+
pipeline=_base_.train_pipeline)
|
57 |
+
|
58 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
59 |
+
batch_size=train_batch_size_per_gpu,
|
60 |
+
collate_fn=dict(type='yolow_collate'),
|
61 |
+
dataset=coco_train_dataset)
|
62 |
+
|
63 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
64 |
+
data_root='data/coco',
|
65 |
+
ann_file='annotations/instances_val2017.json',
|
66 |
+
data_prefix=dict(img='val2017/'),
|
67 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
68 |
+
pipeline=_base_.test_pipeline)
|
69 |
+
|
70 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
71 |
+
test_dataloader = val_dataloader
|
72 |
+
# training settings
|
73 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
74 |
+
lr_factor=0.01,
|
75 |
+
max_epochs=max_epochs),
|
76 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
77 |
+
save_best=None,
|
78 |
+
interval=save_epoch_intervals))
|
79 |
+
custom_hooks = [
|
80 |
+
dict(type='EMAHook',
|
81 |
+
ema_type='ExpMomentumEMA',
|
82 |
+
momentum=0.0001,
|
83 |
+
update_buffers=True,
|
84 |
+
strict_load=False,
|
85 |
+
priority=49),
|
86 |
+
dict(type='mmdet.PipelineSwitchHook',
|
87 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
88 |
+
switch_pipeline=_base_.train_pipeline_stage2)
|
89 |
+
]
|
90 |
+
train_cfg = dict(max_epochs=max_epochs,
|
91 |
+
val_interval=5,
|
92 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
93 |
+
_base_.val_interval_stage2)])
|
94 |
+
|
95 |
+
optim_wrapper = dict(
|
96 |
+
optimizer=dict(_delete_=True,
|
97 |
+
type='SGD',
|
98 |
+
lr=base_lr,
|
99 |
+
momentum=0.937,
|
100 |
+
nesterov=True,
|
101 |
+
weight_decay=weight_decay,
|
102 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
103 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
104 |
+
norm_decay_mult=0.0,
|
105 |
+
custom_keys={'logit_scale': dict(weight_decay=0.0)}),
|
106 |
+
constructor='YOLOWv5OptimizerConstructor')
|
107 |
+
|
108 |
+
# evaluation settings
|
109 |
+
val_evaluator = dict(_delete_=True,
|
110 |
+
type='mmdet.CocoMetric',
|
111 |
+
proposal_nums=(100, 1, 10),
|
112 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
113 |
+
metric='bbox')
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_all_fine_tuning_rmdecay_coco.py
ADDED
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 40 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 30
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 1e-3
|
15 |
+
weight_decay = 0.0005
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
freeze_prompt=True,
|
29 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
30 |
+
backbone=dict(_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
text_model=None,
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=False,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=False,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
52 |
+
data_root='data/coco',
|
53 |
+
ann_file='annotations/instances_train2017.json',
|
54 |
+
data_prefix=dict(img='train2017/'),
|
55 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
56 |
+
pipeline=_base_.train_pipeline)
|
57 |
+
|
58 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
59 |
+
batch_size=train_batch_size_per_gpu,
|
60 |
+
collate_fn=dict(type='yolow_collate'),
|
61 |
+
dataset=coco_train_dataset)
|
62 |
+
|
63 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
64 |
+
data_root='data/coco',
|
65 |
+
ann_file='annotations/instances_val2017.json',
|
66 |
+
data_prefix=dict(img='val2017/'),
|
67 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
68 |
+
pipeline=_base_.test_pipeline)
|
69 |
+
|
70 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
71 |
+
test_dataloader = val_dataloader
|
72 |
+
# training settings
|
73 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
74 |
+
lr_factor=0.01,
|
75 |
+
max_epochs=max_epochs),
|
76 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
77 |
+
save_best=None,
|
78 |
+
interval=save_epoch_intervals))
|
79 |
+
custom_hooks = [
|
80 |
+
dict(type='EMAHook',
|
81 |
+
ema_type='ExpMomentumEMA',
|
82 |
+
momentum=0.0001,
|
83 |
+
update_buffers=True,
|
84 |
+
strict_load=False,
|
85 |
+
priority=49),
|
86 |
+
dict(type='mmdet.PipelineSwitchHook',
|
87 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
88 |
+
switch_pipeline=_base_.train_pipeline_stage2)
|
89 |
+
]
|
90 |
+
train_cfg = dict(max_epochs=max_epochs,
|
91 |
+
val_interval=5,
|
92 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
93 |
+
_base_.val_interval_stage2)])
|
94 |
+
|
95 |
+
optim_wrapper = dict(
|
96 |
+
optimizer=dict(_delete_=True,
|
97 |
+
type='SGD',
|
98 |
+
lr=base_lr,
|
99 |
+
momentum=0.937,
|
100 |
+
nesterov=True,
|
101 |
+
weight_decay=weight_decay,
|
102 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
103 |
+
paramwise_cfg=dict(custom_keys={'logit_scale': dict(weight_decay=0.0)}),
|
104 |
+
constructor='YOLOWv5OptimizerConstructor')
|
105 |
+
|
106 |
+
# evaluation settings
|
107 |
+
val_evaluator = dict(_delete_=True,
|
108 |
+
type='mmdet.CocoMetric',
|
109 |
+
proposal_nums=(100, 1, 10),
|
110 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
111 |
+
metric='bbox')
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_80e_8gpus_all_fine_tuning_coco.py
ADDED
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 1e-3
|
15 |
+
weight_decay = 0.0005
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
freeze_prompt=True,
|
29 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
30 |
+
backbone=dict(_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
text_model=None,
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=False,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=False,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
52 |
+
data_root='data/coco',
|
53 |
+
ann_file='annotations/instances_train2017.json',
|
54 |
+
data_prefix=dict(img='train2017/'),
|
55 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
56 |
+
pipeline=_base_.train_pipeline)
|
57 |
+
|
58 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
59 |
+
batch_size=train_batch_size_per_gpu,
|
60 |
+
collate_fn=dict(type='yolow_collate'),
|
61 |
+
dataset=coco_train_dataset)
|
62 |
+
|
63 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
64 |
+
data_root='data/coco',
|
65 |
+
ann_file='annotations/instances_val2017.json',
|
66 |
+
data_prefix=dict(img='val2017/'),
|
67 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
68 |
+
pipeline=_base_.test_pipeline)
|
69 |
+
|
70 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
71 |
+
test_dataloader = val_dataloader
|
72 |
+
# training settings
|
73 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
74 |
+
lr_factor=0.01,
|
75 |
+
max_epochs=max_epochs),
|
76 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
77 |
+
save_best=None,
|
78 |
+
interval=save_epoch_intervals))
|
79 |
+
custom_hooks = [
|
80 |
+
dict(type='EMAHook',
|
81 |
+
ema_type='ExpMomentumEMA',
|
82 |
+
momentum=0.0001,
|
83 |
+
update_buffers=True,
|
84 |
+
strict_load=False,
|
85 |
+
priority=49),
|
86 |
+
dict(type='mmdet.PipelineSwitchHook',
|
87 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
88 |
+
switch_pipeline=_base_.train_pipeline_stage2)
|
89 |
+
]
|
90 |
+
train_cfg = dict(max_epochs=max_epochs,
|
91 |
+
val_interval=5,
|
92 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
93 |
+
_base_.val_interval_stage2)])
|
94 |
+
|
95 |
+
optim_wrapper = dict(optimizer=dict(
|
96 |
+
_delete_=True,
|
97 |
+
type='SGD',
|
98 |
+
lr=base_lr,
|
99 |
+
momentum=0.937,
|
100 |
+
nesterov=True,
|
101 |
+
weight_decay=weight_decay,
|
102 |
+
batch_size_per_gpu=train_batch_size_per_gpu))
|
103 |
+
|
104 |
+
# evaluation settings
|
105 |
+
val_evaluator = dict(_delete_=True,
|
106 |
+
type='mmdet.CocoMetric',
|
107 |
+
proposal_nums=(100, 1, 10),
|
108 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
109 |
+
metric='bbox')
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_80e_8gpus_all_fine_tuning_rmdecay_coco.py
ADDED
@@ -0,0 +1,113 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 1e-3
|
15 |
+
weight_decay = 0.0005
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
freeze_prompt=True,
|
29 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
30 |
+
backbone=dict(_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
text_model=None,
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=False,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=False,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
52 |
+
data_root='data/coco',
|
53 |
+
ann_file='annotations/instances_train2017.json',
|
54 |
+
data_prefix=dict(img='train2017/'),
|
55 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
56 |
+
pipeline=_base_.train_pipeline)
|
57 |
+
|
58 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
59 |
+
batch_size=train_batch_size_per_gpu,
|
60 |
+
collate_fn=dict(type='yolow_collate'),
|
61 |
+
dataset=coco_train_dataset)
|
62 |
+
|
63 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
64 |
+
data_root='data/coco',
|
65 |
+
ann_file='annotations/instances_val2017.json',
|
66 |
+
data_prefix=dict(img='val2017/'),
|
67 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
68 |
+
pipeline=_base_.test_pipeline)
|
69 |
+
|
70 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
71 |
+
test_dataloader = val_dataloader
|
72 |
+
# training settings
|
73 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
74 |
+
lr_factor=0.01,
|
75 |
+
max_epochs=max_epochs),
|
76 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
77 |
+
save_best=None,
|
78 |
+
interval=save_epoch_intervals))
|
79 |
+
custom_hooks = [
|
80 |
+
dict(type='EMAHook',
|
81 |
+
ema_type='ExpMomentumEMA',
|
82 |
+
momentum=0.0001,
|
83 |
+
update_buffers=True,
|
84 |
+
strict_load=False,
|
85 |
+
priority=49),
|
86 |
+
dict(type='mmdet.PipelineSwitchHook',
|
87 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
88 |
+
switch_pipeline=_base_.train_pipeline_stage2)
|
89 |
+
]
|
90 |
+
train_cfg = dict(max_epochs=max_epochs,
|
91 |
+
val_interval=5,
|
92 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
93 |
+
_base_.val_interval_stage2)])
|
94 |
+
|
95 |
+
optim_wrapper = dict(
|
96 |
+
optimizer=dict(_delete_=True,
|
97 |
+
type='SGD',
|
98 |
+
lr=base_lr,
|
99 |
+
momentum=0.937,
|
100 |
+
nesterov=True,
|
101 |
+
weight_decay=weight_decay,
|
102 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
103 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
104 |
+
norm_decay_mult=0.0,
|
105 |
+
custom_keys={'logit_scale': dict(weight_decay=0.0)}),
|
106 |
+
constructor='YOLOWv5OptimizerConstructor')
|
107 |
+
|
108 |
+
# evaluation settings
|
109 |
+
val_evaluator = dict(_delete_=True,
|
110 |
+
type='mmdet.CocoMetric',
|
111 |
+
proposal_nums=(100, 1, 10),
|
112 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
113 |
+
metric='bbox')
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_80e_8gpus_all_fine_tuning_rmdecay_coco_fixed.py
ADDED
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 70
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 1e-3
|
15 |
+
weight_decay = 0.0005
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
freeze_prompt=True,
|
29 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
30 |
+
backbone=dict(_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
text_model=None,
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=False,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=False,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
52 |
+
data_root='data/coco',
|
53 |
+
ann_file='annotations/instances_train2017.json',
|
54 |
+
data_prefix=dict(img='train2017/'),
|
55 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
56 |
+
pipeline=_base_.train_pipeline)
|
57 |
+
|
58 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
59 |
+
batch_size=train_batch_size_per_gpu,
|
60 |
+
collate_fn=dict(type='yolow_collate'),
|
61 |
+
dataset=coco_train_dataset)
|
62 |
+
|
63 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
64 |
+
data_root='data/coco',
|
65 |
+
ann_file='annotations/instances_val2017.json',
|
66 |
+
data_prefix=dict(img='val2017/'),
|
67 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
68 |
+
pipeline=_base_.test_pipeline)
|
69 |
+
|
70 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
71 |
+
test_dataloader = val_dataloader
|
72 |
+
# training settings
|
73 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
74 |
+
lr_factor=0.01,
|
75 |
+
max_epochs=max_epochs),
|
76 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
77 |
+
save_best=None,
|
78 |
+
interval=save_epoch_intervals))
|
79 |
+
custom_hooks = [
|
80 |
+
dict(type='EMAHook',
|
81 |
+
ema_type='ExpMomentumEMA',
|
82 |
+
momentum=0.0001,
|
83 |
+
update_buffers=True,
|
84 |
+
strict_load=False,
|
85 |
+
priority=49),
|
86 |
+
dict(type='mmdet.PipelineSwitchHook',
|
87 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
88 |
+
switch_pipeline=_base_.train_pipeline_stage2)
|
89 |
+
]
|
90 |
+
train_cfg = dict(max_epochs=max_epochs,
|
91 |
+
val_interval=5,
|
92 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
93 |
+
_base_.val_interval_stage2)])
|
94 |
+
|
95 |
+
optim_wrapper = dict(
|
96 |
+
optimizer=dict(_delete_=True,
|
97 |
+
type='SGD',
|
98 |
+
lr=base_lr,
|
99 |
+
momentum=0.937,
|
100 |
+
nesterov=True,
|
101 |
+
weight_decay=weight_decay,
|
102 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
103 |
+
paramwise_cfg=dict(custom_keys={'logit_scale': dict(weight_decay=0.0)}),
|
104 |
+
constructor='YOLOWv5OptimizerConstructor')
|
105 |
+
|
106 |
+
# evaluation settings
|
107 |
+
val_evaluator = dict(_delete_=True,
|
108 |
+
type='mmdet.CocoMetric',
|
109 |
+
proposal_nums=(100, 1, 10),
|
110 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
111 |
+
metric='bbox')
|
configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-4_80e_8gpus_all_fine_tuning_coco.py
ADDED
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = ('../../third_party/mmyolo/configs/yolov8/'
|
2 |
+
'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
|
3 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
4 |
+
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 80
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 1e-3
|
15 |
+
weight_decay = 0.0005
|
16 |
+
train_batch_size_per_gpu = 16
|
17 |
+
load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
|
18 |
+
persistent_workers = False
|
19 |
+
|
20 |
+
# model settings
|
21 |
+
model = dict(type='YOLOWorldPromptDetector',
|
22 |
+
mm_neck=True,
|
23 |
+
num_train_classes=num_training_classes,
|
24 |
+
num_test_classes=num_classes,
|
25 |
+
embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
|
26 |
+
prompt_dim=text_channels,
|
27 |
+
num_prompts=80,
|
28 |
+
freeze_prompt=True,
|
29 |
+
data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
|
30 |
+
backbone=dict(_delete_=True,
|
31 |
+
type='MultiModalYOLOBackbone',
|
32 |
+
text_model=None,
|
33 |
+
image_model={{_base_.model.backbone}},
|
34 |
+
with_text_model=False),
|
35 |
+
neck=dict(type='YOLOWorldPAFPN',
|
36 |
+
freeze_all=False,
|
37 |
+
guide_channels=text_channels,
|
38 |
+
embed_channels=neck_embed_channels,
|
39 |
+
num_heads=neck_num_heads,
|
40 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
|
41 |
+
bbox_head=dict(type='YOLOWorldHead',
|
42 |
+
head_module=dict(
|
43 |
+
type='YOLOWorldHeadModule',
|
44 |
+
freeze_all=False,
|
45 |
+
use_bn_head=True,
|
46 |
+
embed_dims=text_channels,
|
47 |
+
num_classes=num_training_classes)),
|
48 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
|
49 |
+
|
50 |
+
# dataset settings
|
51 |
+
coco_train_dataset = dict(type='YOLOv5CocoDataset',
|
52 |
+
data_root='data/coco',
|
53 |
+
ann_file='annotations/instances_train2017.json',
|
54 |
+
data_prefix=dict(img='train2017/'),
|
55 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
56 |
+
pipeline=_base_.train_pipeline)
|
57 |
+
|
58 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
59 |
+
batch_size=train_batch_size_per_gpu,
|
60 |
+
collate_fn=dict(type='yolow_collate'),
|
61 |
+
dataset=coco_train_dataset)
|
62 |
+
|
63 |
+
coco_val_dataset = dict(type='YOLOv5CocoDataset',
|
64 |
+
data_root='data/coco',
|
65 |
+
ann_file='annotations/instances_val2017.json',
|
66 |
+
data_prefix=dict(img='val2017/'),
|
67 |
+
filter_cfg=dict(filter_empty_gt=False, min_size=32),
|
68 |
+
pipeline=_base_.test_pipeline)
|
69 |
+
|
70 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
71 |
+
test_dataloader = val_dataloader
|
72 |
+
# training settings
|
73 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
74 |
+
lr_factor=0.01,
|
75 |
+
max_epochs=max_epochs),
|
76 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
77 |
+
save_best=None,
|
78 |
+
interval=save_epoch_intervals))
|
79 |
+
custom_hooks = [
|
80 |
+
dict(type='EMAHook',
|
81 |
+
ema_type='ExpMomentumEMA',
|
82 |
+
momentum=0.0001,
|
83 |
+
update_buffers=True,
|
84 |
+
strict_load=False,
|
85 |
+
priority=49),
|
86 |
+
dict(type='mmdet.PipelineSwitchHook',
|
87 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
88 |
+
switch_pipeline=_base_.train_pipeline_stage2)
|
89 |
+
]
|
90 |
+
train_cfg = dict(max_epochs=max_epochs,
|
91 |
+
val_interval=5,
|
92 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
93 |
+
_base_.val_interval_stage2)])
|
94 |
+
|
95 |
+
optim_wrapper = dict(optimizer=dict(
|
96 |
+
_delete_=True,
|
97 |
+
type='SGD',
|
98 |
+
lr=base_lr,
|
99 |
+
momentum=0.937,
|
100 |
+
nesterov=True,
|
101 |
+
weight_decay=weight_decay,
|
102 |
+
batch_size_per_gpu=train_batch_size_per_gpu))
|
103 |
+
|
104 |
+
# evaluation settings
|
105 |
+
val_evaluator = dict(_delete_=True,
|
106 |
+
type='mmdet.CocoMetric',
|
107 |
+
proposal_nums=(100, 1, 10),
|
108 |
+
ann_file='data/coco/annotations/instances_val2017.json',
|
109 |
+
metric='bbox')
|
configs/segmentation/README.md
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## Fine-tuning YOLO-World for Instance Segmentation
|
2 |
+
|
3 |
+
|
4 |
+
### Models
|
5 |
+
|
6 |
+
We fine-tune YOLO-World on LVIS (`LVIS-Base`) with mask annotations for open-vocabulary (zero-shot) instance segmentation.
|
7 |
+
|
8 |
+
We provide two fine-tuning strategies YOLO-World towards open-vocabulary instance segmentation:
|
9 |
+
|
10 |
+
* fine-tuning `all modules`: leads to better LVIS segmentation accuracy but affects the zero-shot performance.
|
11 |
+
|
12 |
+
* fine-tuning the `segmentation head`: maintains the zero-shot performanc but lowers LVIS segmentation accuracy.
|
13 |
+
|
14 |
+
| Model | Fine-tuning Data | Fine-tuning Modules| AP<sup>mask</su> | AP<sub>r</sub> | AP<sub>c</sub> | AP<sub>f</sub> | Weights |
|
15 |
+
| :---- | :--------------- | :----------------: | :--------------: | :------------: | :------------: | :------------: | :-----: |
|
16 |
+
| [YOLO-World-Seg-M](./yolo_world_seg_m_dual_vlpan_2e-4_80e_8gpus_allmodules_finetune_lvis.py) | `LVIS-Base` | `all modules` | 25.9 | 13.4 | 24.9 | 32.6 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_seg_m_dual_vlpan_2e-4_80e_8gpus_allmodules_finetune_lvis-ca465825.pth) |
|
17 |
+
| [YOLO-World-v2-Seg-M](./yolo_world_seg_m_dual_vlpan_2e-4_80e_8gpus_allmodules_finetune_lvis.py) | `LVIS-Base` | `all modules` | 25.9 | 13.4 | 24.9 | 32.6 | [HF Checkpoints 🤗]() |
|
18 |
+
| [YOLO-World-Seg-L](./yolo_world_seg_l_dual_vlpan_2e-4_80e_8gpus_allmodules_finetune_lvis.py) | `LVIS-Base` | `all modules` | 28.7 | 15.0 | 28.3 | 35.2| [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_seg_l_dual_vlpan_2e-4_80e_8gpus_allmodules_finetune_lvis-8c58c916.pth) |
|
19 |
+
| [YOLO-World-v2-Seg-L](./yolo_world_seg_l_dual_vlpan_2e-4_80e_8gpus_allmodules_finetune_lvis.py) | `LVIS-Base` | `all modules` | 28.7 | 15.0 | 28.3 | 35.2| [HF Checkpoints 🤗]() |
|
20 |
+
| [YOLO-World-Seg-M](./yolo_seg_world_m_dual_vlpan_2e-4_80e_8gpus_seghead_finetune_lvis.py) | `LVIS-Base` | `seg head` | 16.7 | 12.6 | 14.6 | 20.8 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_seg_m_dual_vlpan_2e-4_80e_8gpus_seghead_finetune_lvis-7bca59a7.pth) |
|
21 |
+
| [YOLO-World-v2-Seg-M](./yolo_world_v2_seg_m_vlpan_bn_2e-4_80e_8gpus_seghead_finetune_lvis.py) | `LVIS-Base` | `seg head` | 17.8 | 13.9 | 15.5 | 22.0 | [HF Checkpoints 🤗]() |
|
22 |
+
| [YOLO-World-Seg-L](yolo_seg_world_l_dual_vlpan_2e-4_80e_8gpus_seghead_finetune_lvis.py) | `LVIS-Base` | `seg head` | 19.1 | 14.2 | 17.2 | 23.5 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_seg_l_dual_vlpan_2e-4_80e_8gpus_seghead_finetune_lvis-5a642d30.pth) |
|
23 |
+
| [YOLO-World-v2-Seg-L](./yolo_world_v2_seg_l_vlpan_bn_2e-4_80e_8gpus_seghead_finetune_lvis.py) | `LVIS-Base` | `seg head` | 19.8 | 17.2 | 17.5 | 23.6 | [HF Checkpoints 🤗]() |
|
24 |
+
**NOTE:**
|
25 |
+
1. The mask AP are evaluated on the LVIS `val 1.0`.
|
26 |
+
2. All models are fine-tuned for 80 epochs on `LVIS-Base` (866 categories, `common + frequent`).
|
27 |
+
3. The YOLO-World-Seg with only `seg head` fine-tuned maintains the original zero-shot detection capability and segments objects.
|
configs/segmentation/yolo_world_seg_l_dual_vlpan_2e-4_80e_8gpus_allmodules_finetune_lvis.py
ADDED
@@ -0,0 +1,227 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = (
|
2 |
+
'../../third_party/mmyolo/configs/yolov8/yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py'
|
3 |
+
)
|
4 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 1203
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 2e-4
|
15 |
+
|
16 |
+
weight_decay = 0.05
|
17 |
+
train_batch_size_per_gpu = 8
|
18 |
+
load_from = 'pretrained_models/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth'
|
19 |
+
persistent_workers = False
|
20 |
+
text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
|
21 |
+
# text_model_name = 'openai/clip-vit-base-patch32'
|
22 |
+
# Polygon2Mask
|
23 |
+
downsample_ratio = 4
|
24 |
+
mask_overlap = False
|
25 |
+
use_mask2refine = True
|
26 |
+
max_aspect_ratio = 100
|
27 |
+
min_area_ratio = 0.01
|
28 |
+
|
29 |
+
# model settings
|
30 |
+
model = dict(
|
31 |
+
type='YOLOWorldDetector',
|
32 |
+
mm_neck=True,
|
33 |
+
num_train_classes=num_training_classes,
|
34 |
+
num_test_classes=num_classes,
|
35 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
36 |
+
backbone=dict(
|
37 |
+
_delete_=True,
|
38 |
+
type='MultiModalYOLOBackbone',
|
39 |
+
image_model={{_base_.model.backbone}},
|
40 |
+
text_model=dict(
|
41 |
+
type='HuggingCLIPLanguageBackbone',
|
42 |
+
model_name=text_model_name,
|
43 |
+
frozen_modules=[])),
|
44 |
+
neck=dict(type='YOLOWorldDualPAFPN',
|
45 |
+
guide_channels=text_channels,
|
46 |
+
embed_channels=neck_embed_channels,
|
47 |
+
num_heads=neck_num_heads,
|
48 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
|
49 |
+
text_enhancder=dict(type='ImagePoolingAttentionModule',
|
50 |
+
embed_channels=256,
|
51 |
+
num_heads=8)),
|
52 |
+
bbox_head=dict(type='YOLOWorldSegHead',
|
53 |
+
head_module=dict(type='YOLOWorldSegHeadModule',
|
54 |
+
embed_dims=text_channels,
|
55 |
+
num_classes=num_training_classes,
|
56 |
+
mask_channels=32,
|
57 |
+
proto_channels=256),
|
58 |
+
mask_overlap=mask_overlap,
|
59 |
+
loss_mask=dict(type='mmdet.CrossEntropyLoss',
|
60 |
+
use_sigmoid=True,
|
61 |
+
reduction='none'),
|
62 |
+
loss_mask_weight=1.0),
|
63 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)),
|
64 |
+
test_cfg=dict(mask_thr_binary=0.5, fast_test=True))
|
65 |
+
|
66 |
+
pre_transform = [
|
67 |
+
dict(type='LoadImageFromFile', backend_args=_base_.backend_args),
|
68 |
+
dict(type='LoadAnnotations',
|
69 |
+
with_bbox=True,
|
70 |
+
with_mask=True,
|
71 |
+
mask2bbox=True)
|
72 |
+
]
|
73 |
+
|
74 |
+
last_transform = [
|
75 |
+
dict(type='mmdet.Albu',
|
76 |
+
transforms=_base_.albu_train_transforms,
|
77 |
+
bbox_params=dict(type='BboxParams',
|
78 |
+
format='pascal_voc',
|
79 |
+
label_fields=['gt_bboxes_labels',
|
80 |
+
'gt_ignore_flags']),
|
81 |
+
keymap={
|
82 |
+
'img': 'image',
|
83 |
+
'gt_bboxes': 'bboxes'
|
84 |
+
}),
|
85 |
+
dict(type='YOLOv5HSVRandomAug'),
|
86 |
+
dict(type='mmdet.RandomFlip', prob=0.5),
|
87 |
+
dict(type='Polygon2Mask',
|
88 |
+
downsample_ratio=downsample_ratio,
|
89 |
+
mask_overlap=mask_overlap),
|
90 |
+
]
|
91 |
+
|
92 |
+
# dataset settings
|
93 |
+
text_transform = [
|
94 |
+
dict(type='RandomLoadText',
|
95 |
+
num_neg_samples=(num_classes, num_classes),
|
96 |
+
max_num_samples=num_training_classes,
|
97 |
+
padding_to_max=True,
|
98 |
+
padding_value=''),
|
99 |
+
dict(type='PackDetInputs',
|
100 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
101 |
+
'flip_direction', 'texts'))
|
102 |
+
]
|
103 |
+
mosaic_affine_transform = [
|
104 |
+
dict(type='MultiModalMosaic',
|
105 |
+
img_scale=_base_.img_scale,
|
106 |
+
pad_val=114.0,
|
107 |
+
pre_transform=pre_transform),
|
108 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
109 |
+
dict(
|
110 |
+
type='YOLOv5RandomAffine',
|
111 |
+
max_rotate_degree=0.0,
|
112 |
+
max_shear_degree=0.0,
|
113 |
+
max_aspect_ratio=100.,
|
114 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
115 |
+
# img_scale is (width, height)
|
116 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
117 |
+
border_val=(114, 114, 114),
|
118 |
+
min_area_ratio=_base_.min_area_ratio,
|
119 |
+
use_mask_refine=True)
|
120 |
+
]
|
121 |
+
train_pipeline = [
|
122 |
+
*pre_transform, *mosaic_affine_transform,
|
123 |
+
dict(type='YOLOv5MultiModalMixUp',
|
124 |
+
prob=_base_.mixup_prob,
|
125 |
+
pre_transform=[*pre_transform, *mosaic_affine_transform]),
|
126 |
+
*last_transform, *text_transform
|
127 |
+
]
|
128 |
+
|
129 |
+
_train_pipeline_stage2 = [
|
130 |
+
*pre_transform,
|
131 |
+
dict(type='YOLOv5KeepRatioResize', scale=_base_.img_scale),
|
132 |
+
dict(type='LetterResize',
|
133 |
+
scale=_base_.img_scale,
|
134 |
+
allow_scale_up=True,
|
135 |
+
pad_val=dict(img=114.0)),
|
136 |
+
dict(type='YOLOv5RandomAffine',
|
137 |
+
max_rotate_degree=0.0,
|
138 |
+
max_shear_degree=0.0,
|
139 |
+
scaling_ratio_range=(1 - _base_.affine_scale,
|
140 |
+
1 + _base_.affine_scale),
|
141 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
142 |
+
border_val=(114, 114, 114),
|
143 |
+
min_area_ratio=min_area_ratio,
|
144 |
+
use_mask_refine=use_mask2refine), *last_transform
|
145 |
+
]
|
146 |
+
train_pipeline_stage2 = [*_train_pipeline_stage2, *text_transform]
|
147 |
+
coco_train_dataset = dict(
|
148 |
+
_delete_=True,
|
149 |
+
type='MultiModalDataset',
|
150 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
151 |
+
data_root='data/coco',
|
152 |
+
ann_file='lvis/lvis_v1_train_base.json',
|
153 |
+
data_prefix=dict(img=''),
|
154 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32)),
|
155 |
+
class_text_path='data/texts/lvis_v1_base_class_texts.json',
|
156 |
+
pipeline=train_pipeline)
|
157 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
158 |
+
batch_size=train_batch_size_per_gpu,
|
159 |
+
collate_fn=dict(type='yolow_collate'),
|
160 |
+
dataset=coco_train_dataset)
|
161 |
+
|
162 |
+
test_pipeline = [
|
163 |
+
*_base_.test_pipeline[:-1],
|
164 |
+
dict(type='LoadText'),
|
165 |
+
dict(type='mmdet.PackDetInputs',
|
166 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
167 |
+
'scale_factor', 'pad_param', 'texts'))
|
168 |
+
]
|
169 |
+
|
170 |
+
# training settings
|
171 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
172 |
+
lr_factor=0.01,
|
173 |
+
max_epochs=max_epochs),
|
174 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
175 |
+
save_best=None,
|
176 |
+
interval=save_epoch_intervals))
|
177 |
+
custom_hooks = [
|
178 |
+
dict(type='EMAHook',
|
179 |
+
ema_type='ExpMomentumEMA',
|
180 |
+
momentum=0.0001,
|
181 |
+
update_buffers=True,
|
182 |
+
strict_load=False,
|
183 |
+
priority=49),
|
184 |
+
dict(type='mmdet.PipelineSwitchHook',
|
185 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
186 |
+
switch_pipeline=train_pipeline_stage2)
|
187 |
+
]
|
188 |
+
train_cfg = dict(max_epochs=max_epochs,
|
189 |
+
val_interval=5,
|
190 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
191 |
+
_base_.val_interval_stage2)])
|
192 |
+
optim_wrapper = dict(optimizer=dict(
|
193 |
+
_delete_=True,
|
194 |
+
type='AdamW',
|
195 |
+
lr=base_lr,
|
196 |
+
weight_decay=weight_decay,
|
197 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
198 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
199 |
+
norm_decay_mult=0.0,
|
200 |
+
custom_keys={
|
201 |
+
'backbone.text_model':
|
202 |
+
dict(lr_mult=0.01),
|
203 |
+
'logit_scale':
|
204 |
+
dict(weight_decay=0.0),
|
205 |
+
}),
|
206 |
+
constructor='YOLOWv5OptimizerConstructor')
|
207 |
+
|
208 |
+
# evaluation settings
|
209 |
+
coco_val_dataset = dict(
|
210 |
+
_delete_=True,
|
211 |
+
type='MultiModalDataset',
|
212 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
213 |
+
data_root='data/coco/',
|
214 |
+
test_mode=True,
|
215 |
+
ann_file='lvis/lvis_v1_val.json',
|
216 |
+
data_prefix=dict(img=''),
|
217 |
+
batch_shapes_cfg=None),
|
218 |
+
class_text_path='data/captions/lvis_v1_class_captions.json',
|
219 |
+
pipeline=test_pipeline)
|
220 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
221 |
+
test_dataloader = val_dataloader
|
222 |
+
|
223 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
224 |
+
ann_file='data/coco/lvis/lvis_v1_val.json',
|
225 |
+
metric=['bbox', 'segm'])
|
226 |
+
test_evaluator = val_evaluator
|
227 |
+
find_unused_parameters = True
|
configs/segmentation/yolo_world_seg_l_dual_vlpan_2e-4_80e_8gpus_seghead_finetune_lvis.py
ADDED
@@ -0,0 +1,237 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
_base_ = (
|
2 |
+
'../../third_party/mmyolo/configs/yolov8/yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py'
|
3 |
+
)
|
4 |
+
custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
|
5 |
+
# hyper-parameters
|
6 |
+
num_classes = 1203
|
7 |
+
num_training_classes = 80
|
8 |
+
max_epochs = 80 # Maximum training epochs
|
9 |
+
close_mosaic_epochs = 10
|
10 |
+
save_epoch_intervals = 5
|
11 |
+
text_channels = 512
|
12 |
+
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
|
13 |
+
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
|
14 |
+
base_lr = 2e-4
|
15 |
+
|
16 |
+
weight_decay = 0.05
|
17 |
+
train_batch_size_per_gpu = 8
|
18 |
+
load_from = 'pretrained_models/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth'
|
19 |
+
persistent_workers = False
|
20 |
+
|
21 |
+
# Polygon2Mask
|
22 |
+
downsample_ratio = 4
|
23 |
+
mask_overlap = False
|
24 |
+
use_mask2refine = True
|
25 |
+
max_aspect_ratio = 100
|
26 |
+
min_area_ratio = 0.01
|
27 |
+
|
28 |
+
# model settings
|
29 |
+
model = dict(
|
30 |
+
type='YOLOWorldDetector',
|
31 |
+
mm_neck=True,
|
32 |
+
num_train_classes=num_training_classes,
|
33 |
+
num_test_classes=num_classes,
|
34 |
+
data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
|
35 |
+
backbone=dict(
|
36 |
+
_delete_=True,
|
37 |
+
type='MultiModalYOLOBackbone',
|
38 |
+
image_model={{_base_.model.backbone}},
|
39 |
+
frozen_stages=4, # frozen the image backbone
|
40 |
+
text_model=dict(
|
41 |
+
type='HuggingCLIPLanguageBackbone',
|
42 |
+
model_name='openai/clip-vit-base-patch32',
|
43 |
+
frozen_modules=['all'])),
|
44 |
+
neck=dict(type='YOLOWorldDualPAFPN',
|
45 |
+
freeze_all=True,
|
46 |
+
guide_channels=text_channels,
|
47 |
+
embed_channels=neck_embed_channels,
|
48 |
+
num_heads=neck_num_heads,
|
49 |
+
block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
|
50 |
+
text_enhancder=dict(type='ImagePoolingAttentionModule',
|
51 |
+
embed_channels=256,
|
52 |
+
num_heads=8)),
|
53 |
+
bbox_head=dict(type='YOLOWorldSegHead',
|
54 |
+
head_module=dict(type='YOLOWorldSegHeadModule',
|
55 |
+
embed_dims=text_channels,
|
56 |
+
num_classes=num_training_classes,
|
57 |
+
mask_channels=32,
|
58 |
+
proto_channels=256,
|
59 |
+
freeze_bbox=True),
|
60 |
+
mask_overlap=mask_overlap,
|
61 |
+
loss_mask=dict(type='mmdet.CrossEntropyLoss',
|
62 |
+
use_sigmoid=True,
|
63 |
+
reduction='none'),
|
64 |
+
loss_mask_weight=1.0),
|
65 |
+
train_cfg=dict(assigner=dict(num_classes=num_training_classes)),
|
66 |
+
test_cfg=dict(mask_thr_binary=0.5, fast_test=True))
|
67 |
+
|
68 |
+
pre_transform = [
|
69 |
+
dict(type='LoadImageFromFile', backend_args=_base_.backend_args),
|
70 |
+
dict(type='LoadAnnotations',
|
71 |
+
with_bbox=True,
|
72 |
+
with_mask=True,
|
73 |
+
mask2bbox=True)
|
74 |
+
]
|
75 |
+
|
76 |
+
last_transform = [
|
77 |
+
dict(type='mmdet.Albu',
|
78 |
+
transforms=_base_.albu_train_transforms,
|
79 |
+
bbox_params=dict(type='BboxParams',
|
80 |
+
format='pascal_voc',
|
81 |
+
label_fields=['gt_bboxes_labels',
|
82 |
+
'gt_ignore_flags']),
|
83 |
+
keymap={
|
84 |
+
'img': 'image',
|
85 |
+
'gt_bboxes': 'bboxes'
|
86 |
+
}),
|
87 |
+
dict(type='YOLOv5HSVRandomAug'),
|
88 |
+
dict(type='mmdet.RandomFlip', prob=0.5),
|
89 |
+
dict(type='Polygon2Mask',
|
90 |
+
downsample_ratio=downsample_ratio,
|
91 |
+
mask_overlap=mask_overlap),
|
92 |
+
]
|
93 |
+
|
94 |
+
# dataset settings
|
95 |
+
text_transform = [
|
96 |
+
dict(type='RandomLoadText',
|
97 |
+
num_neg_samples=(num_classes, num_classes),
|
98 |
+
max_num_samples=num_training_classes,
|
99 |
+
padding_to_max=True,
|
100 |
+
padding_value=''),
|
101 |
+
dict(type='PackDetInputs',
|
102 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
|
103 |
+
'flip_direction', 'texts'))
|
104 |
+
]
|
105 |
+
mosaic_affine_transform = [
|
106 |
+
dict(type='MultiModalMosaic',
|
107 |
+
img_scale=_base_.img_scale,
|
108 |
+
pad_val=114.0,
|
109 |
+
pre_transform=pre_transform),
|
110 |
+
dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
|
111 |
+
dict(
|
112 |
+
type='YOLOv5RandomAffine',
|
113 |
+
max_rotate_degree=0.0,
|
114 |
+
max_shear_degree=0.0,
|
115 |
+
max_aspect_ratio=100.,
|
116 |
+
scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
|
117 |
+
# img_scale is (width, height)
|
118 |
+
border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
|
119 |
+
border_val=(114, 114, 114),
|
120 |
+
min_area_ratio=_base_.min_area_ratio,
|
121 |
+
use_mask_refine=True)
|
122 |
+
]
|
123 |
+
train_pipeline = [
|
124 |
+
*pre_transform, *mosaic_affine_transform,
|
125 |
+
dict(type='YOLOv5MultiModalMixUp',
|
126 |
+
prob=_base_.mixup_prob,
|
127 |
+
pre_transform=[*pre_transform, *mosaic_affine_transform]),
|
128 |
+
*last_transform, *text_transform
|
129 |
+
]
|
130 |
+
|
131 |
+
_train_pipeline_stage2 = [
|
132 |
+
*pre_transform,
|
133 |
+
dict(type='YOLOv5KeepRatioResize', scale=_base_.img_scale),
|
134 |
+
dict(type='LetterResize',
|
135 |
+
scale=_base_.img_scale,
|
136 |
+
allow_scale_up=True,
|
137 |
+
pad_val=dict(img=114.0)),
|
138 |
+
dict(type='YOLOv5RandomAffine',
|
139 |
+
max_rotate_degree=0.0,
|
140 |
+
max_shear_degree=0.0,
|
141 |
+
scaling_ratio_range=(1 - _base_.affine_scale,
|
142 |
+
1 + _base_.affine_scale),
|
143 |
+
max_aspect_ratio=_base_.max_aspect_ratio,
|
144 |
+
border_val=(114, 114, 114),
|
145 |
+
min_area_ratio=min_area_ratio,
|
146 |
+
use_mask_refine=use_mask2refine), *last_transform
|
147 |
+
]
|
148 |
+
train_pipeline_stage2 = [*_train_pipeline_stage2, *text_transform]
|
149 |
+
coco_train_dataset = dict(
|
150 |
+
_delete_=True,
|
151 |
+
type='MultiModalDataset',
|
152 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
153 |
+
data_root='data/coco',
|
154 |
+
ann_file='lvis/lvis_v1_train_base.json',
|
155 |
+
data_prefix=dict(img=''),
|
156 |
+
filter_cfg=dict(filter_empty_gt=True, min_size=32)),
|
157 |
+
class_text_path='data/texts/lvis_v1_base_class_texts.json',
|
158 |
+
pipeline=train_pipeline)
|
159 |
+
train_dataloader = dict(persistent_workers=persistent_workers,
|
160 |
+
batch_size=train_batch_size_per_gpu,
|
161 |
+
collate_fn=dict(type='yolow_collate'),
|
162 |
+
dataset=coco_train_dataset)
|
163 |
+
|
164 |
+
test_pipeline = [
|
165 |
+
*_base_.test_pipeline[:-1],
|
166 |
+
dict(type='LoadText'),
|
167 |
+
dict(type='mmdet.PackDetInputs',
|
168 |
+
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
|
169 |
+
'scale_factor', 'pad_param', 'texts'))
|
170 |
+
]
|
171 |
+
|
172 |
+
# training settings
|
173 |
+
default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
|
174 |
+
lr_factor=0.01,
|
175 |
+
max_epochs=max_epochs),
|
176 |
+
checkpoint=dict(max_keep_ckpts=-1,
|
177 |
+
save_best=None,
|
178 |
+
interval=save_epoch_intervals))
|
179 |
+
custom_hooks = [
|
180 |
+
dict(type='EMAHook',
|
181 |
+
ema_type='ExpMomentumEMA',
|
182 |
+
momentum=0.0001,
|
183 |
+
update_buffers=True,
|
184 |
+
strict_load=False,
|
185 |
+
priority=49),
|
186 |
+
dict(type='mmdet.PipelineSwitchHook',
|
187 |
+
switch_epoch=max_epochs - close_mosaic_epochs,
|
188 |
+
switch_pipeline=train_pipeline_stage2)
|
189 |
+
]
|
190 |
+
train_cfg = dict(max_epochs=max_epochs,
|
191 |
+
val_interval=5,
|
192 |
+
dynamic_intervals=[((max_epochs - close_mosaic_epochs),
|
193 |
+
_base_.val_interval_stage2)])
|
194 |
+
optim_wrapper = dict(optimizer=dict(
|
195 |
+
_delete_=True,
|
196 |
+
type='AdamW',
|
197 |
+
lr=base_lr,
|
198 |
+
weight_decay=weight_decay,
|
199 |
+
batch_size_per_gpu=train_batch_size_per_gpu),
|
200 |
+
paramwise_cfg=dict(bias_decay_mult=0.0,
|
201 |
+
norm_decay_mult=0.0,
|
202 |
+
custom_keys={
|
203 |
+
'backbone.text_model':
|
204 |
+
dict(lr_mult=0.01),
|
205 |
+
'logit_scale':
|
206 |
+
dict(weight_decay=0.0),
|
207 |
+
'neck':
|
208 |
+
dict(lr_mult=0.0),
|
209 |
+
'head.head_module.reg_preds':
|
210 |
+
dict(lr_mult=0.0),
|
211 |
+
'head.head_module.cls_preds':
|
212 |
+
dict(lr_mult=0.0),
|
213 |
+
'head.head_module.cls_contrasts':
|
214 |
+
dict(lr_mult=0.0)
|
215 |
+
}),
|
216 |
+
constructor='YOLOWv5OptimizerConstructor')
|
217 |
+
|
218 |
+
# evaluation settings
|
219 |
+
coco_val_dataset = dict(
|
220 |
+
_delete_=True,
|
221 |
+
type='MultiModalDataset',
|
222 |
+
dataset=dict(type='YOLOv5LVISV1Dataset',
|
223 |
+
data_root='data/coco/',
|
224 |
+
test_mode=True,
|
225 |
+
ann_file='lvis/lvis_v1_val.json',
|
226 |
+
data_prefix=dict(img=''),
|
227 |
+
batch_shapes_cfg=None),
|
228 |
+
class_text_path='data/captions/lvis_v1_class_captions.json',
|
229 |
+
pipeline=test_pipeline)
|
230 |
+
val_dataloader = dict(dataset=coco_val_dataset)
|
231 |
+
test_dataloader = val_dataloader
|
232 |
+
|
233 |
+
val_evaluator = dict(type='mmdet.LVISMetric',
|
234 |
+
ann_file='data/coco/lvis/lvis_v1_val.json',
|
235 |
+
metric=['bbox', 'segm'])
|
236 |
+
test_evaluator = val_evaluator
|
237 |
+
find_unused_parameters = True
|