From ed35cfe39325ded6bfeff1b9c6c0b3666df50eee Mon Sep 17 00:00:00 2001 From: James Betker Date: Thu, 20 Jan 2022 11:28:50 -0700 Subject: [PATCH] Update inference scripts --- codes/scripts/audio/gen/use_diffuse_tts.py | 8 +++---- .../scripts/audio/gen/use_discrete_vocoder.py | 14 ++++++----- codes/scripts/audio/gen/use_gpt_tts.py | 24 +++++++++---------- 3 files changed, 23 insertions(+), 23 deletions(-) diff --git a/codes/scripts/audio/gen/use_diffuse_tts.py b/codes/scripts/audio/gen/use_diffuse_tts.py index 06c21eee..d6121441 100644 --- a/codes/scripts/audio/gen/use_diffuse_tts.py +++ b/codes/scripts/audio/gen/use_diffuse_tts.py @@ -22,10 +22,10 @@ if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, help='Path to options YAML file used to train the diffusion model', default='../options/train_diffusion_tts.yml') parser.add_argument('-diffusion_model_name', type=str, help='Name of the diffusion model in opt.', default='generator') - parser.add_argument('-diffusion_model_path', type=str, help='Name of the diffusion model in opt.', default='../experiments/train_diffusion_tts\\models\\13600_generator_ema.pth') + parser.add_argument('-diffusion_model_path', type=str, help='Path to saved model weights', default='../experiments/train_diffusion_tts_experimental_fp16\\models\\17800_generator_ema.pth') parser.add_argument('-aligned_codes', type=str, help='Comma-delimited list of integer codes that defines text & prosody. Get this by apply W2V to an existing audio clip or from a bespoke generator.', default='0,0,0,0,10,10,0,4,0,7,0,17,4,4,0,25,5,0,13,13,0,22,4,4,0,21,15,15,7,0,0,14,4,4,6,8,4,4,0,0,12,5,0,0,5,0,4,4,22,22,8,16,16,0,4,4,4,0,0,0,0,0,0,0') # Default: 'i am very glad to see you', libritts/train-clean-100/103/1241/103_1241_000017_000001.wav. - parser.add_argument('-cond', type=str, help='Path to the conditioning input audio file.', default='Y:\\clips\\books1\\3042_18_Holden__000000000\\00037.wav') + parser.add_argument('-cond', type=str, help='Path to the conditioning input audio file.', default='Y:\\clips\\books1\\754_Dan Simmons - The Rise Of Endymion 356 of 450\\00026.wav') parser.add_argument('-diffusion_steps', type=int, help='Number of diffusion steps to perform to create the generate. Lower steps reduces quality, but >40 is generally pretty good.', default=100) parser.add_argument('-output_path', type=str, help='Where to store outputs.', default='../results/use_diffuse_tts') args = parser.parse_args() @@ -48,11 +48,11 @@ if __name__ == '__main__': # Pad MEL to multiples of 4096//spectrogram_compression_factor msl = aligned_codes.shape[-1] - dsl = 4096 // aligned_codes_compression_factor + dsl = 2048 // aligned_codes_compression_factor gap = dsl - (msl % dsl) if gap > 0: aligned_codes = torch.nn.functional.pad(aligned_codes, (0, gap)) # This still isn't a perfect multiple, but it's close. - output_shape = (1, 1, ceil_multiple(aligned_codes.shape[-1]*aligned_codes_compression_factor, 4096)) + output_shape = (1, 1, ceil_multiple(aligned_codes.shape[-1]*aligned_codes_compression_factor, 2048)) output = diffuser.p_sample_loop(diffusion, output_shape, model_kwargs={'tokens': aligned_codes.unsqueeze(0), 'conditioning_input': cond.unsqueeze(0)}) diff --git a/codes/scripts/audio/gen/use_discrete_vocoder.py b/codes/scripts/audio/gen/use_discrete_vocoder.py index 29c240fc..725f0e02 100644 --- a/codes/scripts/audio/gen/use_discrete_vocoder.py +++ b/codes/scripts/audio/gen/use_discrete_vocoder.py @@ -19,17 +19,18 @@ def roundtrip_vocoding(dvae, vocoder, diffuser, clip, cond=None, plot_spec=False if plot_spec: plot_spectrogram(mel[0].cpu()) codes = convert_mel_to_codes(dvae, mel) - return do_spectrogram_diffusion(vocoder, dvae, diffuser, codes, cond, spectrogram_compression_factor=128, plt_spec=plot_spec) + return do_spectrogram_diffusion(vocoder, dvae, diffuser, codes, cond, spectrogram_compression_factor=256, plt_spec=plot_spec) if __name__ == '__main__': parser = argparse.ArgumentParser() - parser.add_argument('-opt', type=str, help='Path to options YAML file used to train the diffusion model', default='X:\\dlas\\experiments\\train_diffusion_vocoder_with_cond_new_dvae.yml') + parser.add_argument('-opt', type=str, help='Path to options YAML file used to train the diffusion model', + default='X:\\dlas\\experiments\\train_diffusion_vocoder_22k_level.yml') parser.add_argument('-diffusion_model_name', type=str, help='Name of the diffusion model in opt.', default='generator') - parser.add_argument('-diffusion_model_path', type=str, help='Name of the diffusion model in opt.', default='X:\\dlas\\experiments\\train_diffusion_vocoder_with_cond_new_dvae_full\\models\\6100_generator_ema.pth') + parser.add_argument('-diffusion_model_path', type=str, help='Diffusion model checkpoint to load.', default='X:\\dlas\\experiments\\train_diffusion_vocoder_22k_level\\models\\2500_generator.pth') parser.add_argument('-dvae_model_name', type=str, help='Name of the DVAE model in opt.', default='dvae') - parser.add_argument('-input_file', type=str, help='Path to the input audio file.', default='Z:\\clips\\books1\\3_dchha04 Romancing The Tribes\\00036.wav') - parser.add_argument('-cond', type=str, help='Path to the conditioning input audio file.', default='Z:\\clips\\books1\\3042_18_Holden__000000000\\00037.wav') + parser.add_argument('-input_file', type=str, help='Path to the input audio file.', default='Y:\\clips\\books1\\3_dchha04 Romancing The Tribes\\00036.wav') + parser.add_argument('-cond', type=str, help='Path to the conditioning input audio file.', default='Y:\\clips\\books1\\3042_18_Holden__000000000\\00037.wav') args = parser.parse_args() print("Loading DVAE..") @@ -43,7 +44,8 @@ if __name__ == '__main__': cond = inp if args.cond is None else load_audio(args.cond, 22050) if cond.shape[-1] > 44100+10000: cond = cond[:,10000:54100] + cond = cond.cuda() print("Performing inference..") roundtripped = roundtrip_vocoding(dvae, diffusion, diffuser, inp, cond).cpu() - torchaudio.save('roundtrip_vocoded_output.wav', roundtripped.squeeze(0), 11025) \ No newline at end of file + torchaudio.save('roundtrip_vocoded_output.wav', roundtripped.squeeze(0), 22050) \ No newline at end of file diff --git a/codes/scripts/audio/gen/use_gpt_tts.py b/codes/scripts/audio/gen/use_gpt_tts.py index 49ff364f..8bfff791 100644 --- a/codes/scripts/audio/gen/use_gpt_tts.py +++ b/codes/scripts/audio/gen/use_gpt_tts.py @@ -6,17 +6,13 @@ import torch import torch.nn.functional as F import torchaudio import yaml -from tokenizers import Tokenizer from tqdm import tqdm -from data.audio.paired_voice_audio_dataset import CharacterTokenizer from data.audio.unsupervised_audio_dataset import load_audio from data.audio.voice_tokenizer import VoiceBpeTokenizer from data.util import is_audio_file, find_files_of_type -from models.tacotron2.text import text_to_sequence from scripts.audio.gen.speech_synthesis_utils import do_spectrogram_diffusion, \ load_discrete_vocoder_diffuser, wav_to_mel -from trainer.injectors.base_injectors import TorchMelSpectrogramInjector from utils.options import Loader from utils.util import load_model_from_config @@ -81,22 +77,24 @@ def fix_autoregressive_output(codes, stop_token): if __name__ == '__main__': preselected_cond_voices = { 'trump': ['D:\\data\\audio\\sample_voices\\trump.wav'], + 'obama': ['D:\\data\\audio\\sample_voices\\obama1.mp3', 'D:\\data\\audio\\sample_voices\\obama2.wav'], 'ryan_reynolds': ['D:\\data\\audio\\sample_voices\\ryan_reynolds.wav'], 'ed_sheeran': ['D:\\data\\audio\\sample_voices\\ed_sheeran.wav'], 'simmons': ['Y:\\clips\\books1\\754_Dan Simmons - The Rise Of Endymion 356 of 450\\00026.wav'], 'news_girl': ['Y:\\clips\\podcasts-0\\8288_20210113-Is More Violence Coming_\\00022.wav', 'Y:\\clips\\podcasts-0\\8288_20210113-Is More Violence Coming_\\00016.wav'], 'dan_carlin': ['Y:\\clips\\books1\\5_dchha06 Shield of the West\\00476.wav'], - 'libri_test': ['Y:\\libritts\\test-clean\\672\\122797\\672_122797_000057_000002.wav'] + 'libri_test': ['Y:\\libritts\\test-clean\\672\\122797\\672_122797_000057_000002.wav'], + 'myself': ['D:\\data\\audio\\sample_voices\\myself1.wav', 'D:\\data\\audio\\sample_voices\\myself2.wav'], } parser = argparse.ArgumentParser() - parser.add_argument('-opt_diffuse', type=str, help='Path to options YAML file used to train the diffusion model', default='X:\\dlas\\experiments\\train_diffusion_vocoder_with_cond_new_dvae.yml') + parser.add_argument('-opt_diffuse', type=str, help='Path to options YAML file used to train the diffusion model', default='X:\\dlas\\experiments\\train_diffusion_vocoder_22k_level.yml') parser.add_argument('-diffusion_model_name', type=str, help='Name of the diffusion model in opt.', default='generator') - parser.add_argument('-diffusion_model_path', type=str, help='Diffusion model checkpoint to load.', default='X:\\dlas\\experiments\\train_diffusion_vocoder_with_cond_new_dvae_full\\models\\6100_generator_ema.pth') + parser.add_argument('-diffusion_model_path', type=str, help='Diffusion model checkpoint to load.', default='X:\\dlas\\experiments\\train_diffusion_vocoder_22k_level\\models\\12000_generator_ema.pth') parser.add_argument('-dvae_model_name', type=str, help='Name of the DVAE model in opt.', default='dvae') parser.add_argument('-opt_gpt_tts', type=str, help='Path to options YAML file used to train the GPT-TTS model', default='X:\\dlas\\experiments\\train_gpt_tts_unified.yml') parser.add_argument('-gpt_tts_model_name', type=str, help='Name of the GPT TTS model in opt.', default='gpt') - parser.add_argument('-gpt_tts_model_path', type=str, help='GPT TTS model checkpoint to load.', default='X:\\dlas\\experiments\\train_gpt_tts_unified_large\\models\\40000_gpt_ema.pth') + parser.add_argument('-gpt_tts_model_path', type=str, help='GPT TTS model checkpoint to load.', default='X:\\dlas\\experiments\\train_gpt_tts_unified_large\\models\\45000_gpt_ema.pth') parser.add_argument('-opt_clip', type=str, help='Path to options YAML file used to train the CLIP model', default='X:\\dlas\\experiments\\train_clip_text_to_voice.yml') parser.add_argument('-clip_model_name', type=str, help='Name of the CLIP model in opt.', default='clip') parser.add_argument('-clip_model_path', type=str, help='CLIP model checkpoint to load.', default='X:\\dlas\\experiments\\train_clip_text_to_voice_masking_bigger_batch\\models\\23500_clip_ema.pth') @@ -156,15 +154,15 @@ if __name__ == '__main__': del samples, clip print("Loading DVAE..") - dvae = load_model_from_config(args.opt_diffuse, args.dvae_model_name).eval() + dvae = load_model_from_config(args.opt_diffuse, args.dvae_model_name) print("Loading Diffusion Model..") - diffusion = load_model_from_config(args.opt_diffuse, args.diffusion_model_name, also_load_savepoint=False, load_path=args.diffusion_model_path).eval() - diffuser = load_discrete_vocoder_diffuser(desired_diffusion_steps=50) + diffusion = load_model_from_config(args.opt_diffuse, args.diffusion_model_name, also_load_savepoint=False, load_path=args.diffusion_model_path) + diffuser = load_discrete_vocoder_diffuser(desired_diffusion_steps=150) print("Performing vocoding..") # Perform vocoding on each batch element separately: Vocoding is very memory intensive. for b in range(best_results.shape[0]): code = best_results[b].unsqueeze(0) wav = do_spectrogram_diffusion(diffusion, dvae, diffuser, code, cond_wav, - spectrogram_compression_factor=128, plt_spec=False) - torchaudio.save(os.path.join(args.output_path, f'gpt_tts_output_{b}.wav'), wav.squeeze(0).cpu(), 11025) + spectrogram_compression_factor=256, plt_spec=False) + torchaudio.save(os.path.join(args.output_path, f'gpt_tts_output_{b}.wav'), wav.squeeze(0).cpu(), 22050)