Separate channel configs for 2 G's and 2 D's for CycleGAN

[cnmy-ro]

No description provided.

[cnmy-ro]

Made some very minor but important changes in this latest commit, other than my HX4 stuff. For example:

1. midaGAN/engines/validator_tester.py: First, in the denormalize block starting from line 63, now we give clones of tensors to the denormalize() function because otherwise, it changes the visuals dict data and messes up the wandb visualization. Second, change in visuals dictionary key names to match those in GAN classes. For example, visuals['A'] is changed to visuals['real_A'].
2. midaGAN/utils/trackers/utils.py: A minor change in the _split_multimodal_visuals() function to also account for the masks
3. midaGAN/nn/discriminators/patchgan/patchgan3d.py: Set the output channel to 1 in the last `Conv3d' layer
4. midaGAN/nn/generators/unet/unet3d.py: In the config class, the datatype of use_dropout setting was missing. Set it to bool

EDIT: Bugs (3) and (4) were also fixed by @surajpaib in PR #90 and merged into master by @ibro45

[ibro45]

great stuff, i would merge it right away, but won't atm because of the separate in_channels config, need to figure out how to make it better as we discussed a while ago

[ibro45]

could we make both the unpaired and paired patchsampler part of the framework? Maybe have a base patch sampler that has the main functionalities and then write Unpaired and Paired on top of them? Or smth like that

[cnmy-ro]

Sounds good! I'm in favour as well of having a base sampler.

And what do you think about this probability map-based system? The default one is the uniform-random sampling scheme and I added some specialized stuff like fdg-pet-weighted sampling. We could give each of them an integer code like 1, 2 and so on. And maybe also an option for the user to supply their custom sampling map.
EDIT: @surajpaib did create a separate issue for this one #89

[ibro45]

ah true, yes, i agree with everything. once we merge this pull request we could do that if you agree. we'd just have to give more general names than fdg-pet-weighted, if that makes sense of course, rn i don't know what it is. let's keep the discussion in that issue then.

[ibro45]

@cnmy-ro i redesigned how the channels are separated.

For generator:

generator:  
            name: "Resnet2D"
            n_residual_blocks: 9
            in_out_channels:
                AB: [3, 3]
                BA: [3, 3] # Optional, it's interpolated from AB by default

Similar thing for discriminator.

Right now it's defined like that, as a dict, but in the future, when OmegaConf allows Union, we'll allow either like that, as a dist with keys for direction/domain, or just a single tuple (generator) or int (discriminator) when there is only one network or when both networks should have the same value. Added TODOs in the code explaining that. OmegaConf should enable it in the next version, 2.2, but that's probably months away.

Do you agree with what I did? Could you give it a run? We should also update all the yamls then.

It's not ideal the way it is now, i don't find it super beginner-friendly that it exposes channel separation so much, coz you only need it in rare situations. But when OmegaConf will allow Unions then it will be good.

[cnmy-ro]

Yeah, I agree with this. It's really much cleaner this way compared to how I defined these before. I'll run some tests then.

[ibro45]

thanks! merging. if you update yamls or something feel free to push directly to master