How can I run build.py in pytorch1.9?

[BBBBchan]

When I run python build.py to build the native CUDA implementation of InPlace-ABN, I got this error message:

Traceback (most recent call last):
  File "build.py", line 3, in <module>
    from torch.utils.ffi import create_extension
  File "/home/bbchan/anaconda3/envs/pytorch19/lib/python3.7/site-packages/torch/utils/ffi/__init__.py", line 1, in <module>
    raise ImportError("torch.utils.ffi is deprecated. Please use cpp extensions instead.")

I consider that I got this error due to different pytorch version. I run this command in pytorch1,9 and torch.utils.ffi is deprecated. So I change the bulid.py file:

import os

# from torch.utils.ffi import create_extension
from torch.utils.cpp_extension import BuildExtension

sources = ['src/lib_cffi.cpp']
headers = ['src/lib_cffi.h']
extra_objects = ['src/bn.o']
with_cuda = True

this_file = os.path.dirname(os.path.realpath(__file__))
extra_objects = [os.path.join(this_file, fname) for fname in extra_objects]

ffi = BuildExtension(
    '_ext',
    headers=headers,
    sources=sources,
    relative_to=__file__,
    with_cuda=with_cuda,
    extra_objects=extra_objects,
    extra_compile_args=["-std=c++14"]
)

if __name__ == '__main__':
    ffi.build()

However, when I run python build.py command again, I got this error message:

Traceback (most recent call last):
  File "/home/disk/structure_knowledge_distillation/libs/build.py", line 21, in <module>
    extra_compile_args=["-std=c++14"]
  File "/home/anaconda3/envs/pytorch19/lib/python3.7/site-packages/torch/utils/cpp_extension.py", line 361, in __init__
    super(BuildExtension, self).__init__(*args, **kwargs)
  File "/home/anaconda3/envs/pytorch19/lib/python3.7/site-packages/setuptools/__init__.py", line 172, in __init__
    _Command.__init__(self, dist)
  File "/home/anaconda3/envs/pytorch19/lib/python3.7/distutils/cmd.py", line 57, in __init__
    raise TypeError("dist must be a Distribution instance")
TypeError: dist must be a Distribution instance

I consider this error is due to the higher pytorch version too. How can I modify the build.py file to get rid of this problem?

My environment: python 3.7 pytorch 1.9 cuda 11.1

[GeneralJing]

+1. same question.

[BBBBchan]

I have solved this. The build.py is used to build InplaceABN in kd_model.py, so I just replaced InplaceABN with official implement https://github.com/mapillary/inplace_abn. In this way, we don't need build.py and can run the project in pytorch1.9.1. In my condition, this works well and can get similar performance in paper.

[GeneralJing]

through pip install inplace-abn, i can use import inplace-abn. but how to import inplace_abn_sync? This doesn't seem to work. How do you do it?

[GeneralJing]

can you give me some specific operation?

[BBBBchan]

through pip install inplace-abn, i can use import inplace-abn. but how to import inplace_abn_sync? This doesn't seem to work. How do you do it?

Well, you can just from inplace_abn import InPlaceABNSync. BatchNorm2d = functools.partial(InPlaceABNSync, activation='identity')

[GeneralJing]

thank you. I will try that.

[Liuhao-128]

Could you give a more detail operation about how to deal with :TypeError: dist must be a Distribution instance, and how to put the inplace-abn? i don't understand, I have pip install inplace-abn, thanks!

[Liuhao-128]

I also has an error --ModuleNotFoundError: No module named 'libs._ext.__ext' )

[xuyuxiu83]

thank you. I will try that.

Hi, the website provided by the author for downloading the weight files is currently inaccessible, so I'm unable to download the weights. Could you please share the weight files with me? My email is xuyuxiu83@gmail.com.

[xuyuxiu83]

activation='identity'

hello，I cannot enter the website of the pretrained model now.Could you please share it with me?My email is xuyuxiu83@gmail.com.