Please add a requirements.txt with pinned versions (pip freeze)

[yepremyana]

Required Python Packages pytorch >= 1.6 numpy scipy tqdm easydict == 1.9 PyYAML PIL opencv pydensecrf

[twke18]

Thanks for the suggestion! The requirements.txt is uploaded correspondingly.

[yepremyana]

I think for this project it would be best to do pip freeze > requirements.txt to capture the current version and achieve repeatable installs. This will be especially useful months/years down the line if someone comes across this repo wanting to reproduce the results.

[twke18]

Great suggestion! Since pip freeze > requirements.txt lists many packages which are not used for the project, we instead create another file our_requirements.txt for listing our complete python environment to avoid confusion.

[yepremyana]

thank you for adding this information. One thing I noticed is that on our_requirements the torch version is listed as torch==1.5.0+cu101, but in the readme and the requirement.txt pytorch >= 1.6? Also can you please confirm what version of python and cuda you are using, I am having trouble replicating your results.

[twke18]

In our experience, we do not observe performance gap among different versions of PyTorch. Which experiments are you currently trying to reproduce? Do you have exactly the same number of batch_size and crop_size as mentioned in the paper (and in the bashscript)?