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

expand codes before the code converters for cheater latents

Browse Source
This commit is contained in:
James Betker 2022-06-17 09:20:29 -06:00
parent f70b16214d
commit 3efd64ed7a
1 changed files with 9 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
codes/models/audio/music/transformer_diffusion12.py
View File

@ -99,7 +99,7 @@ class TransformerDiffusion(nn.Module):
dropout=0,
use_fp16=False,
ar_prior=False,
code_expansion_mode='nearest',
new_code_expansion=False,
# Parameters for regularization.
unconditioned_percentage=.1, # This implements a mechanism similar to what is used in classifier-free training.
# Parameters for re-training head
@ -115,7 +115,7 @@ class TransformerDiffusion(nn.Module):
self.dropout = dropout
self.unconditioned_percentage = unconditioned_percentage
self.enable_fp16 = use_fp16
self.code_expansion_mode = code_expansion_mode
self.new_code_expansion = new_code_expansion
self.inp_block = conv_nd(1, in_channels, prenet_channels, 3, 1, 1)
@ -209,7 +209,9 @@ class TransformerDiffusion(nn.Module):
return groups
def timestep_independent(self, prior, expected_seq_len):
code_emb = self.ar_input(prior) if self.ar_prior else self.input_converter(prior)
if self.new_code_expansion:
code_emb = F.interpolate(prior.permute(0,2,1), size=expected_seq_len, mode='linear').permute(0,2,1)
code_emb = self.ar_input(code_emb) if self.ar_prior else self.input_converter(code_emb)
code_emb = self.ar_prior_intg(code_emb) if self.ar_prior else self.code_converter(code_emb)
# Mask out the conditioning branch for whole batch elements, implementing something similar to classifier-free guidance.
@ -219,8 +221,9 @@ class TransformerDiffusion(nn.Module):
code_emb = torch.where(unconditioned_batches, self.unconditioned_embedding.repeat(prior.shape[0], 1, 1),
code_emb)
expanded_code_emb = F.interpolate(code_emb.permute(0,2,1), size=expected_seq_len, mode=self.code_expansion_mode).permute(0,2,1)
return expanded_code_emb
if not self.new_code_expansion:
code_emb = F.interpolate(code_emb.permute(0,2,1), size=expected_seq_len, mode='nearest').permute(0,2,1)
return code_emb
def forward(self, x, timesteps, codes=None, conditioning_input=None, precomputed_code_embeddings=None, conditioning_free=False):
if precomputed_code_embeddings is not None:
@ -722,7 +725,7 @@ def test_cheater_model():
model_channels=1024, contraction_dim=512,
prenet_channels=1024, num_heads=8,
input_vec_dim=256, num_layers=12, prenet_layers=6,
dropout=.1,
dropout=.1, new_code_expansion=True,
)
diff_weights = torch.load('extracted_diff.pth')
model.diff.load_state_dict(diff_weights, strict=False)