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

get rid of encoder checkpointing

Browse Source
This commit is contained in:
James Betker 2022-06-10 10:50:34 -06:00
parent 97b32dd39d
commit 84469f3538
2 changed files with 16 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
codes/models/audio/music/music_quantizer2.py
View File

@ -45,7 +45,7 @@ class Upsample(nn.Module):
class ResBlock(nn.Module): class ResBlock(nn.Module):
def __init__(self, chan): def __init__(self, chan, checkpoint=True):
super().__init__() super().__init__()
self.net = nn.Sequential( self.net = nn.Sequential(
nn.Conv1d(chan, chan, 3, padding = 1), nn.Conv1d(chan, chan, 3, padding = 1),
@ -56,9 +56,13 @@ class ResBlock(nn.Module):
nn.SiLU(), nn.SiLU(),
zero_module(nn.Conv1d(chan, chan, 3, padding = 1)), zero_module(nn.Conv1d(chan, chan, 3, padding = 1)),
) )
self.checkpoint = checkpoint
def forward(self, x): def forward(self, x):
return checkpoint(self._forward, x) + x if self.checkpoint:
return checkpoint(self._forward, x) + x
else:
return self._forward(x) + x
def _forward(self, x): def _forward(self, x):
return self.net(x) return self.net(x)
@ -165,7 +169,7 @@ class Wav2Vec2GumbelVectorQuantizer(nn.Module):
class MusicQuantizer2(nn.Module): class MusicQuantizer2(nn.Module):
def __init__(self, inp_channels=256, inner_dim=1024, codevector_dim=1024, down_steps=2, def __init__(self, inp_channels=256, inner_dim=1024, codevector_dim=1024, down_steps=2,
max_gumbel_temperature=2.0, min_gumbel_temperature=.5, gumbel_temperature_decay=.999995, max_gumbel_temperature=2.0, min_gumbel_temperature=.5, gumbel_temperature_decay=.999995,
codebook_size=16, codebook_groups=4, codebook_size=16, codebook_groups=4, checkpoint=True,
# Downsample args: # Downsample args:
expressive_downsamples=False): expressive_downsamples=False):
super().__init__() super().__init__()
@ -191,14 +195,14 @@ class MusicQuantizer2(nn.Module):
self.up = nn.Sequential(*[Upsample(inner_dim[i], inner_dim[i+1]) for i in range(len(inner_dim)-1)] + self.up = nn.Sequential(*[Upsample(inner_dim[i], inner_dim[i+1]) for i in range(len(inner_dim)-1)] +
[nn.Conv1d(inner_dim[-1], inp_channels, kernel_size=3, padding=1)]) [nn.Conv1d(inner_dim[-1], inp_channels, kernel_size=3, padding=1)])
self.encoder = nn.Sequential(ResBlock(inner_dim[0]), self.encoder = nn.Sequential(ResBlock(inner_dim[0], checkpoint=checkpoint),
ResBlock(inner_dim[0]), ResBlock(inner_dim[0], checkpoint=checkpoint),
ResBlock(inner_dim[0])) ResBlock(inner_dim[0], checkpoint=checkpoint))
self.enc_norm = nn.LayerNorm(inner_dim[0], eps=1e-5) self.enc_norm = nn.LayerNorm(inner_dim[0], eps=1e-5)
self.decoder = nn.Sequential(nn.Conv1d(codevector_dim, inner_dim[0], kernel_size=3, padding=1), self.decoder = nn.Sequential(nn.Conv1d(codevector_dim, inner_dim[0], kernel_size=3, padding=1),
ResBlock(inner_dim[0]), ResBlock(inner_dim[0], checkpoint=checkpoint),
ResBlock(inner_dim[0]), ResBlock(inner_dim[0], checkpoint=checkpoint),
ResBlock(inner_dim[0])) ResBlock(inner_dim[0], checkpoint=checkpoint))
self.codes = torch.zeros((3000000,), dtype=torch.long) self.codes = torch.zeros((3000000,), dtype=torch.long)
self.internal_step = 0 self.internal_step = 0

7
codes/models/audio/music/transformer_diffusion8.py
View File

@ -203,7 +203,7 @@ class TransformerDiffusionWithQuantizer(nn.Module):
self.freeze_quantizer_until = freeze_quantizer_until self.freeze_quantizer_until = freeze_quantizer_until
self.diff = TransformerDiffusion(**kwargs) self.diff = TransformerDiffusion(**kwargs)
self.quantizer = MusicQuantizer2(inp_channels=kwargs['in_channels'], inner_dim=quantizer_dims, self.quantizer = MusicQuantizer2(inp_channels=kwargs['in_channels'], inner_dim=quantizer_dims,
codevector_dim=quantizer_dims[0], codevector_dim=quantizer_dims[0], checkpoint=False,
codebook_size=256, codebook_groups=2, codebook_size=256, codebook_groups=2,
max_gumbel_temperature=4, min_gumbel_temperature=.5) max_gumbel_temperature=4, min_gumbel_temperature=.5)
self.quantizer.quantizer.temperature = self.quantizer.min_gumbel_temperature self.quantizer.quantizer.temperature = self.quantizer.min_gumbel_temperature
@ -219,14 +219,13 @@ class TransformerDiffusionWithQuantizer(nn.Module):
) )
def forward(self, x, timesteps, truth_mel, conditioning_input=None, disable_diversity=False, conditioning_free=False): def forward(self, x, timesteps, truth_mel, conditioning_input=None, disable_diversity=False, conditioning_free=False):
quant_grad_enabled = self.internal_step > self.freeze_quantizer_until
mse, diversity_loss, proj = self.quantizer(truth_mel, return_decoder_latent=True) mse, diversity_loss, proj = self.quantizer(truth_mel, return_decoder_latent=True)
proj = proj.permute(0,2,1) proj = proj.permute(0,2,1)
# Make sure this does not cause issues in DDP by explicitly using the parameters for nothing. quant_grad_enabled = self.internal_step > self.freeze_quantizer_until
if not quant_grad_enabled: if not quant_grad_enabled:
proj = proj.detach() proj = proj.detach()
# Make sure this does not cause issues in DDP by explicitly using the parameters for nothing.
unused = 0 unused = 0
for p in self.quantizer.parameters(): for p in self.quantizer.parameters():
unused = unused + p.mean() * 0 unused = unused + p.mean() * 0