deviandice/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Issues Pull Requests Packages Projects Releases Wiki Activity

another fix

Browse Source
This commit is contained in:
James Betker 2022-05-22 21:32:43 -06:00
parent 41809a6330
commit 8f28404645
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
codes/models/audio/mel2vec.py
View File

@ -394,6 +394,7 @@ class Mel2Vec(nn.Module):
self.mask_time_length = mask_time_length
self.disable_custom_linear_init = disable_custom_linear_init
self.linear_init_scale = linear_init_scale
self.dim_reduction_mult = dim_reduction_multiplier
self.apply(self.init)
def init(self, module):
@ -630,7 +631,7 @@ class ContrastiveTrainingWrapper(nn.Module):
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]//self.m2v.dim_reduction_mult)
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)
mask_time_indices = torch.tensor(mask_time_indices, device=mel.device)