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

some random crap

Browse Source
This commit is contained in:
James Betker 2022-05-04 20:29:23 -06:00
parent 6655f7845a
commit 47662b9ec5
3 changed files with 8 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
codes/models/arch_util.py
View File

@ -378,7 +378,6 @@ class ResBlock(nn.Module):
dropout,
out_channels=None,
use_conv=False,
use_scale_shift_norm=False,
dims=2,
up=False,
down=False,
@ -389,7 +388,6 @@ class ResBlock(nn.Module):
self.dropout = dropout
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.use_scale_shift_norm = use_scale_shift_norm
padding = 1 if kernel_size == 3 else 2
self.in_layers = nn.Sequential(
@ -427,7 +425,7 @@ class ResBlock(nn.Module):
else:
self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
def forward(self, x, emb):
def forward(self, x):
"""
Apply the block to a Tensor, conditioned on a timestep embedding.
@ -435,7 +433,7 @@ class ResBlock(nn.Module):
:return: an [N x C x ...] Tensor of outputs.
"""
return checkpoint(
self._forward, x, emb
self._forward, x
)
def _forward(self, x):

2
codes/train.py
View File

@ -327,7 +327,7 @@ class Trainer:
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_tortoise_random_latent_gen_diffuser.yml')
parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../experiments/train_mel_upsampler.yml')
parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher')
args = parser.parse_args()
opt = option.parse(args.opt, is_train=True)

9
codes/trainer/eval/music_diffusion_fid.py
View File

@ -22,7 +22,7 @@ from models.diffusion.gaussian_diffusion import get_named_beta_schedule
from models.diffusion.respace import space_timesteps, SpacedDiffusion
from scripts.audio.gen.speech_synthesis_utils import load_discrete_vocoder_diffuser, wav_to_mel, load_speech_dvae, \
convert_mel_to_codes, load_univnet_vocoder, wav_to_univnet_mel
from trainer.injectors.audio_injectors import denormalize_tacotron_mel, TorchMelSpectrogramInjector
from trainer.injectors.audio_injectors import denormalize_tacotron_mel, TorchMelSpectrogramInjector, pixel_shuffle_1d
from utils.util import ceil_multiple, opt_get, load_model_from_config, pad_or_truncate
@ -64,10 +64,11 @@ class MusicDiffusionFid(evaluator.Evaluator):
else:
real_resampled = audio
audio = audio.unsqueeze(0)
output_shape = audio.shape
output_shape = (1, 16, audio.shape[-1] // 16)
mel = self.spec_fn({'in': audio})['out']
gen = self.diffuser.p_sample_loop(self.model, output_shape, noise=torch.zeros(*output_shape, device=audio.device),
model_kwargs={'aligned_conditioning': mel})
gen = pixel_shuffle_1d(gen, 16)
real_resampled = real_resampled + torch.FloatTensor(real_resampled.shape).uniform_(0.0, 1e-5).to(real_resampled.device)
return gen, real_resampled, sample_rate
@ -149,8 +150,8 @@ class MusicDiffusionFid(evaluator.Evaluator):
if __name__ == '__main__':
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_waveform_gen3.yml', 'generator',
also_load_savepoint=False,
load_path='X:\\dlas\\experiments\\train_music_waveform_gen3_r0\\models\\15400_generator_ema.pth').cuda()
opt_eval = {'path': 'Y:\\split\\yt-music-eval', 'diffusion_steps': 100,
load_path='X:\\dlas\\experiments\\train_music_waveform_gen3_r1\\models\\10000_generator_ema.pth').cuda()
opt_eval = {'path': 'Y:\\split\\yt-music-eval', 'diffusion_steps': 50,
'conditioning_free': False, 'conditioning_free_k': 1,
'diffusion_schedule': 'linear', 'diffusion_type': 'standard'}
env = {'rank': 0, 'base_path': 'D:\\tmp\\test_eval_music', 'step': 1, 'device': 'cuda', 'opt': {}}