All strategies were implemented! With #18 it looks like we won't really even need these as we're not forcing the network to memorize target signal amplitudes.

WaveUNet overfits... How to prevent overfitting.

Data augmentation strategy

3 tasks, 3 dataloaders

1/ mix: premixed signal from the challenge.
2/ remix = on the fly mix while training, frozen mix at validation :warning: cannot find the right mixing rule
3/ denoise : create own mixing rule, signal stays constant, noise is modulated

1/ Mix (=premixed)

[x] augment by scaling all signals (target, noise, mix)
[x] augment by trimming mini batchs
[x] augment by adding a bit of Gaussian Noise

:bulb: other ideas for augmentation:

add clicks , harmonics, clamp the signal ...
mask out (including target and mix)

2/ Remix :lock:

= mix data on the fly, frozen mix at validation & test time :x: :lock: Blocked until we understand how to remix

[x] sample much bigger random coefficients range (~ 3-10) :x: not enough (exp 6)

mixed_audio_signal = clean_audio_signal + \
        (3.+7.*torch.rand((1,), device=clean_audio_signal.device)) * noise_audio_signal

3/ Denoise:

Creates a whole new problem

Dropout

[x] implement Dropout
[x] p=0.2 :test_tube: exp 306 :x: overfit
[x] p=0.5 :test_tube: exp_307 ??? :x: overfit
[x] Deal with legacy wavunet double non linearity :face_with_head_bandage:
[x] Retrain baseline WaveUNet without augmentation

balthazarneveu / gyraudio