DaiShiResearch commited on
Commit
7e1f126
1 Parent(s): f07af5a

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +125 -0
README.md CHANGED
@@ -1,3 +1,128 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ datasets:
4
+ - imagenet-1k
5
+ - ade20k
6
+ metrics:
7
+ - mean_iou
8
+ pipeline_tag: image-segmentation
9
+ tags:
10
+ - vision
11
+ language:
12
+ - en
13
+ library_name: pytorch
14
  ---
15
+
16
+ # TransNeXt
17
+
18
+ Official Model release
19
+ for ["TransNeXt: Robust Foveal Visual Perception for Vision Transformers"](https://arxiv.org/pdf/2311.17132.pdf) [CVPR 2024]
20
+ .
21
+ ## Model Details
22
+ - **Code:** https://github.com/DaiShiResearch/TransNeXt
23
+ - **Paper:** [TransNeXt: Robust Foveal Visual Perception for Vision Transformers](https://arxiv.org/abs/2311.17132)
24
+ - **Author:** [Dai Shi](https://github.com/DaiShiResearch)
25
+ - **Email:** daishiresearch@gmail.com
26
+
27
+ ## Methods
28
+ #### Pixel-focused attention (Left) & aggregated attention (Right):
29
+
30
+ ![pixel-focused_attention](https://github.com/DaiShiResearch/TransNeXt/blob/main/figures/pixel-focused_attention.jpg?raw=true "pixel-focused_attention")
31
+
32
+ #### Convolutional GLU (First on the right):
33
+
34
+ ![Convolutional GLU](https://github.com/DaiShiResearch/TransNeXt/blob/main/figures/feedforward_variants.jpg?raw=true "Convolutional GLU")
35
+
36
+ ## Results
37
+
38
+ #### Image Classification, Detection and Segmentation:
39
+
40
+ ![experiment_figure](https://github.com/DaiShiResearch/TransNeXt/blob/main/figures/experiment_figure.jpg?raw=true "experiment_figure")
41
+
42
+ #### Attention Visualization:
43
+
44
+ ![foveal_peripheral_vision](https://github.com/DaiShiResearch/TransNeXt/blob/main/figures/foveal_peripheral_vision.jpg?raw=true "foveal_peripheral_vision")
45
+
46
+ ## Model Zoo
47
+
48
+ ### Image Classification
49
+
50
+ ***Classification code & weights & configs & training logs are >>>[here](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/ )<<<.***
51
+
52
+ **ImageNet-1K 224x224 pre-trained models:**
53
+
54
+ | Model | #Params | #FLOPs |IN-1K | IN-A | IN-C&#8595; |IN-R|Sketch|IN-V2|Download |Config| Log |
55
+ |:---:|:---:|:---:|:---:| :---:|:---:|:---:|:---:| :---:|:---:|:---:|:---:|
56
+ | TransNeXt-Micro|12.8M|2.7G| 82.5 | 29.9 | 50.8|45.8|33.0|72.6|[model](https://huggingface.co/DaiShiResearch/transnext-micro-224-1k/resolve/main/transnext_micro_224_1k.pth?download=true) |[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/transnext_micro.py)|[log](https://huggingface.co/DaiShiResearch/transnext-micro-224-1k/raw/main/transnext_micro_224_1k.txt) |
57
+ | TransNeXt-Tiny |28.2M|5.7G| 84.0| 39.9| 46.5|49.6|37.6|73.8|[model](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/transnext_tiny.py)|[log](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/raw/main/transnext_tiny_224_1k.txt)|
58
+ | TransNeXt-Small |49.7M|10.3G| 84.7| 47.1| 43.9|52.5| 39.7|74.8 |[model](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/resolve/main/transnext_small_224_1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/transnext_small.py)|[log](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/raw/main/transnext_small_224_1k.txt)|
59
+ | TransNeXt-Base |89.7M|18.4G| 84.8| 50.6|43.5|53.9|41.4|75.1| [model](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/resolve/main/transnext_base_224_1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/transnext_base.py)|[log](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/raw/main/transnext_base_224_1k.txt)|
60
+
61
+ **ImageNet-1K 384x384 fine-tuned models:**
62
+
63
+ | Model | #Params | #FLOPs |IN-1K | IN-A |IN-R|Sketch|IN-V2| Download |Config|
64
+ |:---:|:---:|:---:|:---:| :---:|:---:|:---:| :---:|:---:|:---:|
65
+ | TransNeXt-Small |49.7M|32.1G| 86.0| 58.3|56.4|43.2|76.8| [model](https://huggingface.co/DaiShiResearch/transnext-small-384-1k-ft-1k/resolve/main/transnext_small_384_1k_ft_1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/finetune/transnext_small_384.py)|
66
+ | TransNeXt-Base |89.7M|56.3G| 86.2| 61.6|57.7|44.7|77.0| [model](https://huggingface.co/DaiShiResearch/transnext-base-384-1k-ft-1k/resolve/main/transnext_base_384_1k_ft_1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/finetune/transnext_base_384.py)|
67
+
68
+ **ImageNet-1K 256x256 pre-trained model fully utilizing aggregated attention at all stages:**
69
+
70
+ *(See Table.9 in Appendix D.6 for details)*
71
+
72
+ | Model |Token mixer| #Params | #FLOPs |IN-1K |Download |Config| Log |
73
+ |:---:|:---:|:---:|:---:| :---:|:---:|:---:|:---:|
74
+ |TransNeXt-Micro|**A-A-A-A**|13.1M|3.3G| 82.6 |[model](https://huggingface.co/DaiShiResearch/transnext-micro-AAAA-256-1k/resolve/main/transnext_micro_AAAA_256_1k.pth?download=true) |[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/transnext_micro_AAAA_256.py)|[log](https://huggingface.co/DaiShiResearch/transnext-micro-AAAA-256-1k/blob/main/transnext_micro_AAAA_256_1k.txt) |
75
+
76
+ ### Object Detection
77
+ ***Object detection code & weights & configs & training logs are >>>[here](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/ )<<<.***
78
+
79
+ **COCO object detection and instance segmentation results using the Mask R-CNN method:**
80
+
81
+ | Backbone | Pretrained Model| Lr Schd| box mAP | mask mAP | #Params | Download |Config| Log |
82
+ |:---:|:---:|:---:|:---:| :---:|:---:|:---:|:---:|:---:|
83
+ | TransNeXt-Tiny | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true) |1x|49.9|44.6|47.9M|[model](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-tiny-coco/resolve/main/mask_rcnn_transnext_tiny_fpn_1x_coco_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/maskrcnn/configs/mask_rcnn_transnext_tiny_fpn_1x_coco.py)|[log](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-tiny-coco/raw/main/mask_rcnn_transnext_tiny_fpn_1x_coco_in1k.log.json)|
84
+ | TransNeXt-Small | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/resolve/main/transnext_small_224_1k.pth?download=true) |1x|51.1|45.5|69.3M|[model](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-small-coco/resolve/main/mask_rcnn_transnext_small_fpn_1x_coco_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/maskrcnn/configs/mask_rcnn_transnext_small_fpn_1x_coco.py)|[log](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-small-coco/raw/main/mask_rcnn_transnext_small_fpn_1x_coco_in1k.log.json)|
85
+ | TransNeXt-Base | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/resolve/main/transnext_base_224_1k.pth?download=true) |1x|51.7|45.9|109.2M|[model](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-base-coco/resolve/main/mask_rcnn_transnext_base_fpn_1x_coco_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/maskrcnn/configs/mask_rcnn_transnext_base_fpn_1x_coco.py)|[log](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-base-coco/raw/main/mask_rcnn_transnext_base_fpn_1x_coco_in1k.log.json)|
86
+
87
+ **COCO object detection results using the DINO method:**
88
+
89
+ | Backbone | Pretrained Model| scales | epochs | box mAP | #Params | Download |Config| Log |
90
+ |:---:|:---:|:---:|:---:| :---:|:---:|:---:|:---:|:---:|
91
+ | TransNeXt-Tiny | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true)|4scale | 12|55.1|47.8M|[model](https://huggingface.co/DaiShiResearch/dino-4scale-transnext-tiny-coco/resolve/main/dino_4scale_transnext_tiny_12e_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/dino/configs/dino-4scale_transnext_tiny-12e_coco.py)|[log](https://huggingface.co/DaiShiResearch/dino-4scale-transnext-tiny-coco/raw/main/dino_4scale_transnext_tiny_12e_in1k.json)|
92
+ | TransNeXt-Tiny | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true)|5scale | 12|55.7|48.1M|[model](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-tiny-coco/resolve/main/dino_5scale_transnext_tiny_12e_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/dino/configs/dino-5scale_transnext_tiny-12e_coco.py)|[log](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-tiny-coco/raw/main/dino_5scale_transnext_tiny_12e_in1k.json)|
93
+ | TransNeXt-Small | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/resolve/main/transnext_small_224_1k.pth?download=true)|5scale | 12|56.6|69.6M|[model](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-small-coco/resolve/main/dino_5scale_transnext_small_12e_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/dino/configs/dino-5scale_transnext_small-12e_coco.py)|[log](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-small-coco/raw/main/dino_5scale_transnext_small_12e_in1k.json)|
94
+ | TransNeXt-Base | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/resolve/main/transnext_base_224_1k.pth?download=true)|5scale | 12|57.1|110M|[model](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-base-coco/resolve/main/dino_5scale_transnext_base_12e_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/dino/configs/dino-5scale_transnext_base-12e_coco.py)|[log](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-base-coco/raw/main/dino_5scale_transnext_base_12e_in1k.json)|
95
+
96
+ ### Semantic Segmentation
97
+ ***Semantic segmentation code & weights & configs & training logs are >>>[here](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/ )<<<.***
98
+
99
+ **ADE20K semantic segmentation results using the UPerNet method:**
100
+
101
+ | Backbone | Pretrained Model| Crop Size |Lr Schd| mIoU|mIoU (ms+flip)| #Params | Download |Config| Log |
102
+ |:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
103
+ | TransNeXt-Tiny | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true)|512x512|160K|51.1|51.5/51.7|59M|[model](https://huggingface.co/DaiShiResearch/upernet-transnext-tiny-ade/resolve/main/upernet_transnext_tiny_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/upernet/configs/upernet_transnext_tiny_512x512_160k_ade20k_ss.py)|[log](https://huggingface.co/DaiShiResearch/upernet-transnext-tiny-ade/blob/main/upernet_transnext_tiny_512x512_160k_ade20k_ss.log.json)|
104
+ | TransNeXt-Small | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/resolve/main/transnext_small_224_1k.pth?download=true)|512x512|160K|52.2|52.5/51.8|80M|[model](https://huggingface.co/DaiShiResearch/upernet-transnext-small-ade/resolve/main/upernet_transnext_small_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/upernet/configs/upernet_transnext_small_512x512_160k_ade20k_ss.py)|[log](https://huggingface.co/DaiShiResearch/upernet-transnext-small-ade/blob/main/upernet_transnext_small_512x512_160k_ade20k_ss.log.json)|
105
+ | TransNeXt-Base | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/resolve/main/transnext_base_224_1k.pth?download=true)|512x512|160K|53.0|53.5/53.7|121M|[model](https://huggingface.co/DaiShiResearch/upernet-transnext-base-ade/resolve/main/upernet_transnext_base_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/upernet/configs/upernet_transnext_base_512x512_160k_ade20k_ss.py)|[log](https://huggingface.co/DaiShiResearch/upernet-transnext-base-ade/blob/main/upernet_transnext_base_512x512_160k_ade20k_ss.log.json)|
106
+ * In the context of multi-scale evaluation, TransNeXt reports test results under two distinct scenarios: **interpolation** and **extrapolation** of relative position bias.
107
+
108
+ **ADE20K semantic segmentation results using the Mask2Former method:**
109
+
110
+ | Backbone | Pretrained Model| Crop Size |Lr Schd| mIoU| #Params | Download |Config| Log |
111
+ |:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
112
+ | TransNeXt-Tiny | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true)|512x512|160K|53.4|47.5M|[model](https://huggingface.co/DaiShiResearch/mask2former-transnext-tiny-ade/resolve/main/mask2former_transnext_tiny_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/mask2former/configs/mask2former_transnext_tiny_160k_ade20k-512x512.py)|[log](https://huggingface.co/DaiShiResearch/mask2former-transnext-tiny-ade/raw/main/mask2former_transnext_tiny_512x512_160k_ade20k_in1k.json)|
113
+ | TransNeXt-Small | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/resolve/main/transnext_small_224_1k.pth?download=true)|512x512|160K|54.1|69.0M|[model](https://huggingface.co/DaiShiResearch/mask2former-transnext-small-ade/resolve/main/mask2former_transnext_small_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/mask2former/configs/mask2former_transnext_small_160k_ade20k-512x512.py)|[log](https://huggingface.co/DaiShiResearch/mask2former-transnext-small-ade/raw/main/mask2former_transnext_small_512x512_160k_ade20k_in1k.json)|
114
+ | TransNeXt-Base | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/resolve/main/transnext_base_224_1k.pth?download=true)|512x512|160K|54.7|109M|[model](https://huggingface.co/DaiShiResearch/mask2former-transnext-base-ade/resolve/main/mask2former_transnext_base_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/mask2former/configs/mask2former_transnext_base_160k_ade20k-512x512.py)|[log](https://huggingface.co/DaiShiResearch/mask2former-transnext-base-ade/raw/main/mask2former_transnext_base_512x512_160k_ade20k_in1k.json)|
115
+
116
+ ## Citation
117
+
118
+ If you find our work helpful, please consider citing the following bibtex. We would greatly appreciate a star for this
119
+ project.
120
+
121
+ @misc{shi2023transnext,
122
+ author = {Dai Shi},
123
+ title = {TransNeXt: Robust Foveal Visual Perception for Vision Transformers},
124
+ year = {2023},
125
+ eprint = {arXiv:2311.17132},
126
+ archivePrefix={arXiv},
127
+ primaryClass={cs.CV}
128
+ }