utils.show_model(model)
val_losses = []
dc_weight = 1
for epoch in epochs:
if args.unfold_iterations:
if (epoch % int(params['num_epochs'] / 5) == 0):
n_iterations = int(params['num_epoch'] / epoch)
module.clusterer.n_iterations = min(params['num_clustering_iterations'], n_iterations)
if args.curriculum_learning:
if epoch >= int(params['num_epochs'] / 5):
# Lengthen sequences for learning
dataset.create_cache = True
dataset.length = 1.0
val_dataset.create_cache = True
val_dataset.length = 1.0
progress_bar = trange(num_iterations)
epoch_loss = []
for (spectrogram, magnitude_spectrogram, source_spectrograms, source_ibms, weights, one_hots) in dataloader:
spectrogram = spectrogram.to(device).requires_grad_()
magnitude_spectrogram = magnitude_spectrogram.to(device).unsqueeze(-1).requires_grad_()
source_spectrograms = source_spectrograms.float().to(device)
source_ibms = source_ibms.to(device).float()
one_hots = one_hots.float().to(device).requires_grad_()
if weights is not None:
weights = weights.float().to(device)
optimizer.zero_grad()
