mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity

degrade gumbel softmax over time

Browse Source
This commit is contained in:
James Betker 2022-05-17 16:23:04 -06:00
parent 3853f37257
commit c1bdb4f9a1
1 changed files with 13 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
codes/models/audio/mel2vec.py
View File

@ -533,7 +533,8 @@ class Wav2Vec2GumbelVectorQuantizer(nn.Module):
class ContrastiveTrainingWrapper(nn.Module): class ContrastiveTrainingWrapper(nn.Module):
def __init__(self, inner_dim=1024, dropout=.1, mask_time_prob=.65, mask_time_length=4, num_negatives=100, **kwargs): def __init__(self, inner_dim=1024, dropout=.1, mask_time_prob=.65, mask_time_length=4, num_negatives=100,
max_gumbel_temperature=2.0, min_gumbel_temperature=.5, gumbel_temperature_decay=.999995, **kwargs):
super().__init__() super().__init__()
self.m2v = Mel2Vec(inner_dim=inner_dim, dropout=dropout, mask_time_prob=mask_time_prob, self.m2v = Mel2Vec(inner_dim=inner_dim, dropout=dropout, mask_time_prob=mask_time_prob,
mask_time_length=mask_time_length, **kwargs) mask_time_length=mask_time_length, **kwargs)
@ -541,6 +542,9 @@ class ContrastiveTrainingWrapper(nn.Module):
self.num_negatives = num_negatives self.num_negatives = num_negatives
self.mask_time_prob = mask_time_prob self.mask_time_prob = mask_time_prob
self.mask_time_length = mask_time_length self.mask_time_length = mask_time_length
self.max_gumbel_temperature = max_gumbel_temperature
self.min_gumbel_temperature = min_gumbel_temperature
self.gumbel_temperature_decay = gumbel_temperature_decay
self.quantizer = Wav2Vec2GumbelVectorQuantizer(inner_dim) self.quantizer = Wav2Vec2GumbelVectorQuantizer(inner_dim)
@ -569,7 +573,15 @@ class ContrastiveTrainingWrapper(nn.Module):
logits = logits / temperature logits = logits / temperature
return logits return logits
def update_for_step(self, step, *args):
self.quantizer.temperature = max(
self.max_gumbel_temperature * self.gumbel_temperature_decay**step,
self.min_gumbel_temperature,
)
def forward(self, mel): def forward(self, mel):
mel = mel[:, :, :-1] # The MEL computation always pads with 1, throwing off optimal tensor math.
features_shape = (mel.shape[0], mel.shape[-1]//4) features_shape = (mel.shape[0], mel.shape[-1]//4)
mask_time_indices = _compute_mask_indices(features_shape, self.mask_time_prob, self.mask_time_length) mask_time_indices = _compute_mask_indices(features_shape, self.mask_time_prob, self.mask_time_length)
sampled_negative_indices = torch.tensor(_sample_negative_indices(features_shape, self.num_negatives, mask_time_indices=mask_time_indices), device=mel.device) sampled_negative_indices = torch.tensor(_sample_negative_indices(features_shape, self.num_negatives, mask_time_indices=mask_time_indices), device=mel.device)