Hard-Coded Scaling is removed for ProgressiveGANs, GanTrainer and ProgressiveAE Training and Model files

[Aakanksha-Rana]

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)

Summary

Checklist

• I have added tests to cover my changes
• I have updated documentation (if necessary)

Acknowledgment

• I acknowledge that this contribution will be available under the Apache 2 license.
  Fixes Problems creating a TFrecord dataset #201

[Aakanksha-Rana]

We should use MinMaxIntensityScaling for [0,1] or CustomIntensityScaling for [-1,1] in "augment" option or set "normalizer = True" while loading the data. This also removes any need of post-processing in the generate function of ProgressiveGANs.

[satra]

it's ok to remove instead of commenting. the details will remain in git history.

[Aakanksha-Rana]

@satra @Hoda1394 any idea why this update could fail the checks? I simply removed comments.

[Hoda1394]

It seems to be a pytest issue. have you run pytest locally?

[Aakanksha-Rana]

It seems to be a pytest issue. have you run pytest locally?

yes and that is fine.

[wazeerzulfikar]

I think that with these changes (if the real images are not scaled), during the training of the discriminator, the generated images will be in range [-1, 1] while real images will not be in that range.
The final activation function should probably be removed then to not constrain the output of the generator.

Also, the generate function will not convert the tanh final outputs. Hence, that function cannot be used to visualize/save images.
Has the effect of not scaling the data on training also been verified? or upon reading the above comments, it seems like normalization will be a different PR?

[satra]

@wazeerzulfikar - normalization will be part of the get dataset step, in the sense that this tool can expect it's inputs to have been normalized in a [1, -1] range, and thus it's outputs should also be in that range. any mapping is thus external to this code. you can see an example through the change in the example notebook.

@Aakanksha-Rana - i do think this would be a good class for the new API and it would be good to create an example. also in the notebook, does the generation part of the notebook still work with these changes?

[satra]

@wanderine - can you try this branch in this PR for testing?

[satra]

i'm going to merge this in and we can check it in relation to the new generation api.

[wanderine]

I pulled the latest docker container from docker hub and ran a training using the same training script as before, but the problem is still there. Do I need to change something more, redo the creation of TFrecords ? Replace some python file in the container?

https://github.com/wanderine/ASSIST/blob/main/3DGANsegmentation/train_multigpu.py

[Aakanksha-Rana]

https://github.com/wanderine/ASSIST/blob/main/3DGANsegmentation/train_multigpu.py
Have you tired using "normalizer=True" in line 71? It still looks like set as None...

[wanderine]

I will try that, thanks

[wanderine]

TypeError: in user code:

File "/usr/local/lib/python3.8/dist-packages/nobrainer/dataset.py", line 149, in None  *
    lambda x, y: (normalizer(x), y)

TypeError: 'bool' object is not callable

[satra]

@wanderine - are you using the enh/api branch as noted in the other issue? the docker containers will not have all these changes.

[satra]

this generation example also includes scaling the data: https://github.com/neuronets/nobrainer/blob/enh/api/guide/api_train_generation_progressive.ipynb (but it will require installing that branch of nobrainer - pip install https://github.com/neuronets/nobrainer/archive/refs/heads/enh/api.zip

[Aakanksha-Rana]

TypeError: in user code:

File "/usr/local/lib/python3.8/dist-packages/nobrainer/dataset.py", line 149, in None  *
    lambda x, y: (normalizer(x), y)

TypeError: 'bool' object is not callable

Ah! by default normalizer is set to standardize the data in get_dataset. Can you just remove the "normalizer=None" in line 71 ?

[wanderine]

OK let me know when the Docker containers have the changes, because it seems too complicated to run it otherwise.

[satra]

it seems too complicated to run it otherwise.

@wanderine - could you please let us know as to why it is complicated to run otherwise? does the system you are working on not support conda or some such environment ? if this is something that can be improved with documentation, we would like to do so.

[wanderine]

There are too many abstraction layers, I can barely keep track of everything. For the old code I used anaconda, then I switched to singularity, now I should use anaconda again?

[satra]

@wanderine - thanks for letting us know. i don't know what made the switch to singularity. all of nobrainer can be run in a conda-forge environment (not anaconda - as most packages are not systematically updated there). the docker is only provided as a convenience function for those who are using it as a command line tool. i'm assuming you are running a python script for your training, hence the api refactor should really help you. also it will help us to know if our abstractions are useful.