clean

Dockerfile_back

@@ -1,103 +0,0 @@
-ENV PATH=/opt/conda/bin:/opt/cmake-3.14.6-Linux-x86_64/bin/:/usr/local/mpi/bin:/usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/local/ucx/bin:/opt/tensorrt/bin
-
-ENV CUDA_VERSION=11.3.1.005
-
-ENV CUDA_DRIVER_VERSION=465.19.01
-
-ENV CUDA_CACHE_DISABLE=1
-
-ENV _CUDA_COMPAT_PATH=/usr/local/cuda/compat
-
-ENV ENV=/etc/shinit_v2
-
-ENV BASH_ENV=/etc/bash.bashrc
-
-ENV NVIDIA_REQUIRE_CUDA=cuda>=9.0
-
-ENV NCCL_VERSION=2.9.9
-
-ENV CUBLAS_VERSION=11.5.1.109
-
-ENV CUFFT_VERSION=10.4.2.109
-
-ENV CURAND_VERSION=10.2.4.109
-
-ENV CUSPARSE_VERSION=11.6.0.109
-
-ENV CUSOLVER_VERSION=11.1.2.109
-
-ENV NPP_VERSION=11.3.3.95
-
-ENV NVJPEG_VERSION=11.5.0.109
-
-ENV CUDNN_VERSION=8.2.1.32
-
-ENV TRT_VERSION=7.2.3.4+cuda11.1.0.024
-
-ENV TRTOSS_VERSION=21.06
-
-ENV NSIGHT_SYSTEMS_VERSION=2021.2.1.58
-
-ENV NSIGHT_COMPUTE_VERSION=2021.1.1.5
-
-ENV DALI_VERSION=1.2.0
-
-ENV DALI_BUILD=2356513
-
-ENV DLPROF_VERSION=21.06
-
-ENV LD_LIBRARY_PATH=/opt/hpcx/nccl_rdma_sharp_plugin/lib:/usr/local/cuda/compat/lib:/usr/local/nvidia/lib:/usr/local/nvidia/lib64
-
-ENV NVIDIA_VISIBLE_DEVICES=all
-
-ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility,video
-
-ENV GDRCOPY_VERSION=2.2
-
-ENV HPCX_VERSION=2.8.2rc3
-
-ENV MOFED_VERSION=5.2-rdmacore32.1
-
-ENV OPENUCX_VERSION=1.10.1
-
-ENV OPENMPI_VERSION=4.1.1rc1
-
-ENV RDMACORE_VERSION=32.1
-
-ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:
-
-ENV PYTORCH_BUILD_VERSION=1.9.0a0+c3d40fd
-
-ENV PYTORCH_VERSION=1.9.0a0+c3d40fd
-
-ENV PYTORCH_BUILD_NUMBER=0
-
-ENV NVIDIA_PYTORCH_VERSION=21.06
-
-ENV NVM_DIR=/usr/local/nvm
-
-ENV JUPYTER_PORT=8888
-
-ENV TENSORBOARD_PORT=6006
-
-ENV COCOAPI_VERSION=2.0+nv0.5.1
-
-ENV PYTHONIOENCODING=utf-8
-
-ENV LC_ALL=C.UTF-8
-
-ENV NVIDIA_BUILD_ID=24129275
-
-ENV NCCL_ASYNC_ERROR_HANDLING=1
-
-EXPOSE 6006/tcp
-
-EXPOSE 8888/tcp
-
-WORKDIR /workspace
-ENTRYPOINT ["/usr/local/bin/nvidia_entrypoint.sh"]
-
-
-COPY . .
-
-CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "7860"]

LICENSE

@@ -0,0 +1,22 @@
+    MIT License
+
+    Copyright (c) Liunian Harold Li, Zi-Yi Dou.
+    Copyright (c) Microsoft Corporation.
+
+    Permission is hereby granted, free of charge, to any person obtaining a copy
+    of this software and associated documentation files (the "Software"), to deal
+    in the Software without restriction, including without limitation the rights
+    to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+    copies of the Software, and to permit persons to whom the Software is
+    furnished to do so, subject to the following conditions:
+
+    The above copyright notice and this permission notice shall be included in all
+    copies or substantial portions of the Software.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+    OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+    SOFTWARE

app.py

@@ -1,19 +1,4 @@
-'''import gradio as gr
-def update(name):
-    return f"Welcome to Gradio, {name}!"
-
-with gr.Blocks() as demo:
-    gr.Markdown("Start typing below and then click **Run** to see the output.")
-    with gr.Row():
-        inp = gr.Textbox(placeholder="What is your name?")
-        out = gr.Textbox()
-    btn = gr.Button("Run")
-    btn.click(fn=update, inputs=inp, outputs=out)
-
-demo.launch()
-
-
-exit()#'''
+# Reference: https://huggingface.co/spaces/haotiz/glip-zeroshot-demo/blob/main/app.py 
 
 import requests
 import os
@@ -27,27 +12,14 @@ import warnings
 
 warnings.filterwarnings("ignore")
 
-# os.system(
-#     "pip install einops shapely timm yacs tensorboardX ftfy prettytable pymongo click opencv-python inflect nltk scipy scikit-learn pycocotools")
-# os.system("pip install transformers")
 os.system("python setup.py build develop --user")
 
 from maskrcnn_benchmark.config import cfg
 from maskrcnn_benchmark.engine.predictor_glip import GLIPDemo
 
-# Use this command for evaluate the GLIP-T model
-#config_file = "configs/pretrain/glip_Swin_T_O365_GoldG.yaml"
-#weight_file = "MODEL/glip_tiny_model_o365_goldg_cc_sbu.pth"
-
 config_file = "configs/pretrain_new/desco_glip.yaml"
 weight_file = "MODEL/desco_glip_tiny.pth"
 
-
-# Use this command if you want to try the GLIP-L model
-# ! wget https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/models/glip_large_model.pth -O MODEL/glip_large_model.pth
-# config_file = "configs/pretrain/glip_Swin_L.yaml"
-# weight_file = "MODEL/glip_large_model.pth"
-
 # update the config options with the config file
 # manual override some options
 cfg.local_rank = 0
@@ -77,11 +49,8 @@ fiber_demo = GLIPDemo(
     show_mask_heatmaps=False
 )
 
-def predict(image, text, specified_tokens=""):
-    if specified_tokens.strip() == "":
-        specified_tokens = None
-    else:
-        specified_tokens = specified_tokens.strip().split(";")
+def predict(image, text, ground_tokens=""):
+    ground_tokens = None if ground_tokens.strip() == "" else ground_tokens.strip().split(";")
     result, _ = glip_demo.run_on_web_image(deepcopy(image[:, :, [2, 1, 0]]), text, 0.5, specified_tokens)
     fiber_result, _ = fiber_demo.run_on_web_image(deepcopy(image[:, :, [2, 1, 0]]), text, 0.5, specified_tokens)
     return result[:, :, [2, 1, 0]], fiber_result[:, :, [2, 1, 0]]
@@ -90,27 +59,24 @@ def predict(image, text, specified_tokens=""):
 image = gr.inputs.Image()
 
 gr.Interface(
-    description="Object Recognition with DesCo",
+    description="Object Recognition with DesCo (https://github.com/liunian-harold-li/DesCo)",
     fn=predict,
     inputs=["image", "text", "text"],
     outputs=[
         gr.outputs.Image(
             type="pil",
-            # label="grounding results"
+            label="DesCo-GLIP"
         ),
         gr.outputs.Image(
             type="pil",
-            # label="grounding results"
+            label="DesCo-FIBER"
         ),
     ],
     examples=[
-        #["./flickr_9472793441.jpg", "bobble heads on top of the shelf ."],
-        #["./flickr_9472793441.jpg", "sofa . remote . dog . person . car . sky . plane ."],
         ["./coco_000000281759.jpg", "A green umbrella. A pink striped umbrella. A plain white umbrella.", ""],
         ["./coco_000000281759.jpg", "a flowery top. A blue dress. An orange shirt .", ""],
         ["./coco_000000281759.jpg", "a car . An electricity box .", ""],
         ["./1.jpg", "a train besides sidewalk", "train;sidewalk"],
-        #["./flickr_7520721.jpg", "A woman figure skater in a blue costume holds her leg by the blade of her skate ."]
     ],
     article=Path("docs/intro.md").read_text()
 ).launch()

docs/intro.md

@@ -1 +1,6 @@
-DesCo
+# DesCo
+This is the HuggingFace Gradio Demo for the paper [DesCo: Learning Object Recognition with Rich Language Descriptions (NeurIPS 2023)](https://arxiv.org/pdf/2306.14060.pdf).
+
+`ground_tokens` specifies which tokens we wish the model to ground to, separated by `;`; if it is not specified, the script will use NLTK to extract noun phrases automatically.
+
+Code: https://github.com/liunian-harold-li/DesCo

main.py

@@ -1,105 +0,0 @@
-import gradio as gr
-import sys
-import torch
-import requests
-import socket
-from torchvision import transforms
-import os
-import asyncio
-
-torch.hub._validate_not_a_forked_repo=lambda a,b,c: True
-
-model = torch.hub.load("pytorch/vision:v0.6.0", "resnet18", pretrained=True).eval()
-response = requests.get("https://git.io/JJkYN")
-labels = response.text.split("\n")
-
-
-INITIAL_PORT_VALUE = int(os.getenv("GRADIO_SERVER_PORT", "7860"))
-TRY_NUM_PORTS = int(os.getenv("GRADIO_NUM_PORTS", "100"))
-LOCALHOST_NAME = os.getenv("GRADIO_SERVER_NAME", "127.0.0.1")
-GRADIO_API_SERVER = "https://api.gradio.app/v1/tunnel-request"
-
-
-if sys.platform == "win32" and hasattr(asyncio, "WindowsSelectorEventLoopPolicy"):
-    # "Any thread" and "selector" should be orthogonal, but there's not a clean
-    # interface for composing policies so pick the right base.
-    _BasePolicy = asyncio.WindowsSelectorEventLoopPolicy  # type: ignore
-else:
-    _BasePolicy = asyncio.DefaultEventLoopPolicy
-
-
-class AnyThreadEventLoopPolicy(_BasePolicy):  # type: ignore
-    """Event loop policy that allows loop creation on any thread.
-    The default `asyncio` event loop policy only automatically creates
-    event loops in the main threads. Other threads must create event
-    loops explicitly or `asyncio.get_event_loop` (and therefore
-    `.IOLoop.current`) will fail. Installing this policy allows event
-    loops to be created automatically on any thread, matching the
-    behavior of Tornado versions prior to 5.0 (or 5.0 on Python 2).
-    Usage::
-        asyncio.set_event_loop_policy(AnyThreadEventLoopPolicy())
-    .. versionadded:: 5.0
-    """
-
-    def get_event_loop(self) -> asyncio.AbstractEventLoop:
-        try:
-            return super().get_event_loop()
-        except (RuntimeError, AssertionError):
-            # This was an AssertionError in python 3.4.2 (which ships with debian jessie)
-            # and changed to a RuntimeError in 3.4.3.
-            # "There is no current event loop in thread %r"
-            loop = self.new_event_loop()
-            self.set_event_loop(loop)
-            return loop
-
-
-asyncio.set_event_loop_policy(AnyThreadEventLoopPolicy())
-
-def get_first_available_port(initial: int, final: int) -> int:
-    """
-    Gets the first open port in a specified range of port numbers
-    Parameters:
-    initial: the initial value in the range of port numbers
-    final: final (exclusive) value in the range of port numbers, should be greater than `initial`
-    Returns:
-    port: the first open port in the range
-    """
-    for port in range(initial, final):
-        try:
-            s = socket.socket()  # create a socket object
-            s.bind((LOCALHOST_NAME, port))  # Bind to the port
-            s.close()
-            return port
-        except OSError:
-            pass
-    raise OSError(
-        "All ports from {} to {} are in use. Please close a port.".format(
-            initial, final
-        )
-    )
-
-def predict(inp):
-    inp = transforms.ToTensor()(inp).unsqueeze(0)
-    with torch.no_grad():
-        prediction = torch.nn.functional.softmax(model(inp)[0], dim=0)
-        confidences = {labels[i]: float(prediction[i]) for i in range(1000)}
-    return confidences
-
-
-def run():
-    demo = gr.Interface(
-        fn=predict,
-        inputs=gr.inputs.Image(type="pil"),
-        outputs=gr.outputs.Label(num_top_classes=3),
-    )
-
-    demo.launch() #server_name="0.0.0.0", server_port=7861)
-    #demo.launch(server_name=LOCALHOST_NAME, server_port=get_first_available_port(
-    #        INITIAL_PORT_VALUE, INITIAL_PORT_VALUE + TRY_NUM_PORTS
-    #    ), share=True)
-    #demo.launch(server_name="0.0.0.0", server_port=7861)
-
-
-if __name__ == "__main__":
-    run()
-

maskrcnn_benchmark/modeling/rpn/inference.py

@@ -726,7 +726,6 @@ class ATSSPostProcessor(torch.nn.Module):
             if number_of_detections > self.fpn_post_nms_top_n > 0:
                 cls_scores = result.get_field("scores")
                 image_thresh, _ = torch.kthvalue(
-                    # TODO: confirm with Pengchuan and Xiyang, torch.kthvalue is not implemented for 'Half'
                     # cls_scores.cpu(),
                     cls_scores.cpu().float(),
                     number_of_detections - self.fpn_post_nms_top_n + 1,