gpastal24 / vitpose-pytorch Goto Github PK

Are the bounding boxes when extracting the image patches to be inputs into vit in the form x1,y1,x2,y2 or aare they converted to tlwh format?

Also what is the purpose of the correction factor in the pose_utils.py file?

Hi, thank you for your work!
Do you know if the current implementation would work with the large wholdebody_model from Vit_pose?

Hello thank you for this amazing work. Please I would like to know if there if a way to generate 2D keypoints intead of 3D ones.

Hi，I am trying to use your code to convert the vitpose model. Andf I meet some errors below

File "convert_to_trt.py", line 8, in
net_trt = torch2trt(pose, [x],max_batch_size=10, fp16_mode=True)
File "/usr/local/python/lib/python3.8/site-packages/torch2trt-0.4.0-py3.8.egg/torch2trt/torch2trt.py", line 736, in torch2trt
outputs = module(*inputs)
File "/usr/local/python/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1120, in _call_impl
result = forward_call(*input, **kwargs)
File "/data/ViTPose-Pytorch/builder/model_builder.py", line 49, in forward
x = self.backbone(x)
File "/usr/local/python/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1120, in _call_impl
result = forward_call(*input, **kwargs)
File "/data/ViTPose-Pytorch/backbones/vit.py", line 302, in forward
x = self.forward_features(x)
File "/data/ViTPose-Pytorch/backbones/vit.py", line 287, in forward_features
x = x + self.pos_embed[:, 1:] + self.pos_embed[:, :1]
File "/usr/local/python/lib/python3.8/site-packages/torch2trt-0.4.0-py3.8.egg/torch2trt/torch2trt.py", line 307, in wrapper
converter"converter"
File "/usr/local/python/lib/python3.8/site-packages/torch2trt-0.4.0-py3.8.egg/torch2trt/converters/getitem.py", line 30, in convert_tensor_getitem
input_trt = input._trt
AttributeError: 'Parameter' object has no attribute '_trt'

Have you meet this error before? Any suggestion? Looking forward to your reply ￣ω￣=

Thank you for nice work.
I found the issue when I try to install required library.

Environment

OS: MacOS (Intel CPU)
python: 3.8.10 (on virtualenv)
(also happened with python3.11.0)

Script

cd ViTPose-Pytorch
pip install -r requirements.txt

Error message

enevis/dev/ai/ViTPose/ViTPose-Pytorch  
Collecting ffmpeg-python
  Using cached ffmpeg_python-0.2.0-py3-none-any.whl (25 kB)
Collecting munkres
  Using cached munkres-1.1.4-py2.py3-none-any.whl (7.0 kB)
Collecting numpy<=1.23.0
  Downloading numpy-1.23.0-cp38-cp38-macosx_10_9_x86_64.whl (18.1 MB)
     |████████████████████████████████| 18.1 MB 730 kB/s 
ERROR: Could not find a version that satisfies the requirement nvidia-tensorrt==8.4.1.5 (from versions: 0.0.1.dev4, 0.0.1.dev5)
ERROR: No matching distribution found for nvidia-tensorrt==8.4.1.5
WARNING: You are using pip version 21.1.1; however, version 23.2.1 is available.
You should consider upgrading via the '/Users/xxx/ViTPose/vitpose-env/bin/python -m pip install --upgrade pip' command.

Thank you

Hi, is there any difference between the original vitpose architecture? I get different result for same frame when I run on https://huggingface.co/spaces/Gradio-Blocks/ViTPose . Do you thing it might be about image correction part?

Do you have any plans to release a train script for this?