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

Update mdf spectral

Browse Source
This commit is contained in:
James Betker 2022-05-23 10:37:15 -06:00
parent 36dd4eb61f
commit 874de1775d
1 changed files with 5 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
codes/trainer/eval/music_diffusion_fid.py
View File

@ -2,6 +2,7 @@ import os
import os.path as osp import os.path as osp
from glob import glob from glob import glob
from random import shuffle from random import shuffle
from time import time
import numpy as np import numpy as np
import torch import torch
@ -67,7 +68,7 @@ class MusicDiffusionFid(evaluator.Evaluator):
elif 'from_codes' == mode: elif 'from_codes' == mode:
self.diffusion_fn = self.perform_diffusion_from_codes self.diffusion_fn = self.perform_diffusion_from_codes
self.local_modules['codegen'] = get_music_codegen() self.local_modules['codegen'] = get_music_codegen()
self.spec_fn = TorchMelSpectrogramInjector({'n_mel_channels': 256, 'mel_fmax': 22000, 'normalize': True, 'in': 'in', 'out': 'out'}, {}) self.spec_fn = TorchMelSpectrogramInjector({'n_mel_channels': 256, 'mel_fmax': 11000, 'filter_length': 16000, 'normalize': True, 'in': 'in', 'out': 'out'}, {})
def load_data(self, path): def load_data(self, path):
return list(glob(f'{path}/*.wav')) return list(glob(f'{path}/*.wav'))
@ -80,7 +81,7 @@ class MusicDiffusionFid(evaluator.Evaluator):
audio = audio.unsqueeze(0) audio = audio.unsqueeze(0)
output_shape = (1, 16, audio.shape[-1] // 16) output_shape = (1, 16, audio.shape[-1] // 16)
mel = self.spec_fn({'in': audio})['out'] 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), gen = self.diffuser.p_sample_loop(self.model, output_shape,
model_kwargs={'aligned_conditioning': mel}) model_kwargs={'aligned_conditioning': mel})
gen = pixel_shuffle_1d(gen, 16) gen = pixel_shuffle_1d(gen, 16)
@ -246,7 +247,8 @@ class MusicDiffusionFid(evaluator.Evaluator):
if __name__ == '__main__': if __name__ == '__main__':
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_diffusion_flat.yml', 'generator', diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_diffusion_flat.yml', 'generator',
also_load_savepoint=False, also_load_savepoint=False,
load_path='X:\\dlas\\experiments\\train_music_diffusion_flat\\models\\33000_generator_ema.pth').cuda() #load_path='X:\\dlas\\experiments\\train_music_diffusion_flat\\models\\33000_generator_ema.pth'
).cuda()
opt_eval = {'path': 'Y:\\split\\yt-music-eval', 'diffusion_steps': 100, opt_eval = {'path': 'Y:\\split\\yt-music-eval', 'diffusion_steps': 100,
'conditioning_free': False, 'conditioning_free_k': 1, 'conditioning_free': False, 'conditioning_free_k': 1,
'diffusion_schedule': 'linear', 'diffusion_type': 'from_codes'} 'diffusion_schedule': 'linear', 'diffusion_type': 'from_codes'}