Which torch and torchvision version are appropriate?

[ninellekam]

I can't do nothing, awlays some kind of problems about inplace_abn compiling. Which torch and torchvision versions are needed here, please, could you help?

[springbayy]

The versions are in the .yaml - file.

I suspect you get the error because you use a newer version of pytorch.
However, there is a problem I can't get around:

I use a nivida A100- Running the environment as is in the yaml as provided does not work on the sm80-architecture. Versions of pytorch / CUDA are too old to run on the A100 and need to be upgraded. But installing the newest versions of everything gives me the following error:

RuntimeError: Error building extension 'inplace_abn': [1/3] /usr/bin/nvcc --generate-dependencies-with-compile --dependency-output inplace_abn_cuda_half.cuda.o.d -DTORCH_EXTENSION_NAME=inplace_abn -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE="gcc" -DPYBIND11_STDLIB="libstdcpp" -DPYBIND11_BUILD_ABI="cxxabi1011" -isystem /outfit_anyone/.conda/lib/python3.8/site-packages/torch/include -isystem /outfit_anyone/.conda/lib/python3.8/site-packages/torch/include/torch/csrc/api/include -isystem /outfit_anyone/.conda/lib/python3.8/site-packages/torch/include/TH -isystem /outfit_anyone/.conda/lib/python3.8/site-packages/torch/include/THC -isystem /outfit_anyone/.conda/include/python3.8 -D_GLIBCXX_USE_CXX11_ABI=0 -D__CUDA_NO_HALF_OPERATORS -D__CUDA_NO_HALF_CONVERSIONS_ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_80,code=sm_80 --compiler-options '-fPIC' --expt-extended-lambda -std=c++17 -c /outfit_anyone/segmentation/Self-Correction-Human-Parsing/modules/src/inplace_abn_cuda_half.cu -o inplace_abn_cuda_half.cuda.o
FAILED: inplace_abn_cuda_half.cuda.o

...etc...
ninja: build stopped: subcommand failed.

I have tried to find intermediate versions of pytorch that may support the GPU as well as the rest of the environment but without success.

To my knowledge this is the best and most accessible pre-trained model on clothing that gives nice labels on the different parts. One could train a newer model on the same dataset I suppose, but I'd rather not.

Would be very appreciated if the github could be updated to support the newer versions of GPUs, otherwise the model is obsolete :)

[Hooray-Lee]

The versions are in the .yaml - file.

I suspect you get the error because you use a newer version of pytorch. However, there is a problem I can't get around:

I use a nivida A100- Running the environment as is in the yaml as provided does not work on the sm80-architecture. Versions of pytorch / CUDA are too old to run on the A100 and need to be upgraded. But installing the newest versions of everything gives me the following error:

RuntimeError: Error building extension 'inplace_abn': [1/3] /usr/bin/nvcc --generate-dependencies-with-compile --dependency-output inplace_abn_cuda_half.cuda.o.d -DTORCH_EXTENSION_NAME=inplace_abn -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE="gcc" -DPYBIND11_STDLIB="libstdcpp" -DPYBIND11_BUILD_ABI="cxxabi1011" -isystem /outfit_anyone/.conda/lib/python3.8/site-packages/torch/include -isystem /outfit_anyone/.conda/lib/python3.8/site-packages/torch/include/torch/csrc/api/include -isystem /outfit_anyone/.conda/lib/python3.8/site-packages/torch/include/TH -isystem /outfit_anyone/.conda/lib/python3.8/site-packages/torch/include/THC -isystem /outfit_anyone/.conda/include/python3.8 -D_GLIBCXX_USE_CXX11_ABI=0 -D__CUDA_NO_HALF_OPERATORS -D__CUDA_NO_HALF_CONVERSIONS_ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_80,code=sm_80 --compiler-options '-fPIC' --expt-extended-lambda -std=c++17 -c /outfit_anyone/segmentation/Self-Correction-Human-Parsing/modules/src/inplace_abn_cuda_half.cu -o inplace_abn_cuda_half.cuda.o FAILED: inplace_abn_cuda_half.cuda.o

...etc... ninja: build stopped: subcommand failed.

I have tried to find intermediate versions of pytorch that may support the GPU as well as the rest of the environment but without success.

To my knowledge this is the best and most accessible pre-trained model on clothing that gives nice labels on the different parts. One could train a newer model on the same dataset I suppose, but I'd rather not.

Would be very appreciated if the github could be updated to support the newer versions of GPUs, otherwise the model is obsolete :)

Hi, for me, pip install inplace-abn and from inplace_abn import InPlaceABNSync instead of using the origin one helps. Maybe you can try it. :)

[springbayyy]

Thanks!
Not working on it anymore. However, i later used segment anything with GrindingDINO i think. Actually, it works even better!
(Or at least the models are more sophisticated)

So this is a useful model but as of 2024 i think the newer option is superior!

Best

[Hooray-Lee]

Thanks! Not working on it anymore. However, i later used segment anything with GrindingDINO i think. Actually, it works even better! (Or at least the models are more sophisticated)

So this is a useful model but as of 2024 i think the newer option is superior!

Best

Thanks a lot! Your idea of using Grounded SAM works so well to me! (not this project but another important one) I think I was constrained by the methods in existing papers before. Thank you!

[frequiem11]

Any luck working with recent torch versions?