kxngluca/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

Ffix tfdpc_v5 conditioning

Browse Source
This commit is contained in:
James Betker 2022-06-27 14:06:09 -06:00
parent 0278576f37
commit ee3b426dae
3 changed files with 11 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
codes/models/audio/music/tfdpc_v3.py
View File

@ -215,8 +215,8 @@ class TransformerDiffusionWithConditioningEncoder(nn.Module):
self.diff = TransformerDiffusionWithPointConditioning(**kwargs) self.diff = TransformerDiffusionWithPointConditioning(**kwargs)
self.conditioning_encoder = ConditioningEncoder(256, kwargs['model_channels']) self.conditioning_encoder = ConditioningEncoder(256, kwargs['model_channels'])
def forward(self, x, timesteps, true_cheater, conditioning_input=None, disable_diversity=False, conditioning_free=False): def forward(self, x, timesteps, conditioning_input=None, disable_diversity=False, conditioning_free=False):
cond = self.conditioning_encoder(true_cheater) cond = self.conditioning_encoder(conditioning_input)
diff = self.diff(x, timesteps, conditioning_input=cond, conditioning_free=conditioning_free) diff = self.diff(x, timesteps, conditioning_input=cond, conditioning_free=conditioning_free)
return diff return diff

12
codes/models/audio/music/tfdpc_v5.py
View File

@ -196,16 +196,16 @@ class TransformerDiffusionWithPointConditioning(nn.Module):
unused_params = [] unused_params = []
time_emb = self.time_embed(timestep_embedding(timesteps, self.time_embed_dim)) time_emb = self.time_embed(timestep_embedding(timesteps, self.time_embed_dim))
if conditioning_free:
cond = self.unconditioned_embedding
cond = cond.repeat(1,x.shape[-1],1)
else:
cond_enc = self.conditioning_encoder(conditioning_input, time_emb) cond_enc = self.conditioning_encoder(conditioning_input, time_emb)
cs = cond_enc[:,:,cond_start] cs = cond_enc[:,:,cond_start]
ce = cond_enc[:,:,x.shape[-1]+cond_start] ce = cond_enc[:,:,x.shape[-1]+cond_start]
cond_enc = torch.cat([cs.unsqueeze(-1), ce.unsqueeze(-1)], dim=-1) cond_enc = torch.cat([cs.unsqueeze(-1), ce.unsqueeze(-1)], dim=-1)
cond_enc = F.interpolate(cond_enc, size=(x.shape[-1],), mode='linear').permute(0,2,1) cond = F.interpolate(cond_enc, size=(x.shape[-1],), mode='linear').permute(0,2,1)
if conditioning_free:
cond = self.unconditioned_embedding
else:
cond = cond_enc
# Mask out the conditioning branch for whole batch elements, implementing something similar to classifier-free guidance. # Mask out the conditioning branch for whole batch elements, implementing something similar to classifier-free guidance.
if self.training and self.unconditioned_percentage > 0: if self.training and self.unconditioned_percentage > 0:
unconditioned_batches = torch.rand((cond.shape[0], 1, 1), unconditioned_batches = torch.rand((cond.shape[0], 1, 1),

6
codes/trainer/eval/music_diffusion_fid.py
View File

@ -301,14 +301,14 @@ class MusicDiffusionFid(evaluator.Evaluator):
if __name__ == '__main__': if __name__ == '__main__':
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_cheater_gen.yml', 'generator', diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_cheater_gen_r8.yml', 'generator',
also_load_savepoint=False, also_load_savepoint=False,
load_path='X:\\dlas\\experiments\\train_music_cheater_gen_v4\\models\\28000_generator_ema.pth' load_path='X:\\dlas\\experiments\\train_music_cheater_gen_v5\\models\\18000_generator_ema.pth'
).cuda() ).cuda()
opt_eval = {'path': 'Y:\\split\\yt-music-eval', # eval music, mostly electronica. :) opt_eval = {'path': 'Y:\\split\\yt-music-eval', # eval music, mostly electronica. :)
#'path': 'E:\\music_eval', # this is music from the training dataset, including a lot more variety. #'path': 'E:\\music_eval', # this is music from the training dataset, including a lot more variety.
'diffusion_steps': 32, 'diffusion_steps': 32,
'conditioning_free': False, 'conditioning_free_k': 1, 'clip_audio': False, 'use_ddim': True, 'conditioning_free': True, 'conditioning_free_k': 1, 'clip_audio': False, 'use_ddim': True,
'diffusion_schedule': 'linear', 'diffusion_type': 'cheater_gen', 'diffusion_schedule': 'linear', 'diffusion_type': 'cheater_gen',
#'partial_low': 128, 'partial_high': 192 #'partial_low': 128, 'partial_high': 192
} }