This commit is contained in:
James Betker 2022-02-11 10:59:32 -07:00
parent 5175b7d91a
commit 46b97049dc
2 changed files with 6 additions and 3 deletions

View File

@ -32,7 +32,7 @@ if __name__ == '__main__':
'less_resblocks': {'networks': {'generator': {'kwargs': {'num_res_blocks': [1, 1, 1, 1, 1, 1, 1]}}}},
'wider': {'networks': {'generator': {'kwargs': {'channel_mult': [1,2,4,6,8,8,8]}}}},
'inject_every_layer': {'networks': {'generator': {'kwargs': {'token_conditioning_resolutions': [1,2,4,8,16,32,64]}}}},
'deep_conditioning': {'networks': {'generator': {'kwargs': {'cond_transformer_depth': 12}}}},
'cosine_diffusion': {'steps': {'generator': {'injectors': {'diffusion': {'beta_schedule': {'schedule_name': 'cosine'}}}}}},
}
opt = option.parse(base_opt, is_train=True)
all_opts = []

View File

@ -17,7 +17,7 @@ from data.audio.unsupervised_audio_dataset import load_audio
from scripts.audio.gen.speech_synthesis_utils import load_discrete_vocoder_diffuser
class StyleTransferEvaluator(evaluator.Evaluator):
class AudioDiffusionFid(evaluator.Evaluator):
"""
Evaluator produces generate from a diffusion model, then uses a pretrained wav2vec model to compute a frechet
distance between real and fake samples.
@ -31,7 +31,10 @@ class StyleTransferEvaluator(evaluator.Evaluator):
else:
self.skip = 1
diffusion_steps = opt_get(opt_eval, ['diffusion_steps'], 50)
diffusion_schedule = opt_get(opt_eval, ['diffusion_schedule'], 'cosine')
diffusion_schedule = opt_get(env['opt'], ['steps', 'generator', 'injectors', 'diffusion', 'beta_schedule', 'schedule_name'], None)
if diffusion_schedule is None:
print("Unable to infer diffusion schedule from master options. Getting it from eval (or guessing).")
diffusion_schedule = opt_get(opt_eval, ['diffusion_schedule'], 'cosine')
self.diffuser = load_discrete_vocoder_diffuser(desired_diffusion_steps=diffusion_steps, schedule=diffusion_schedule)
self.dev = self.env['device']