diff --git a/TTS/tts/datasets/formatters.py b/TTS/tts/datasets/formatters.py index 8358524bba..bdccc8386f 100644 --- a/TTS/tts/datasets/formatters.py +++ b/TTS/tts/datasets/formatters.py @@ -657,10 +657,13 @@ def bel_tts_formatter(root_path, meta_file, **kwargs): # pylint: disable=unused def afrotts(root_path, meta_file, **kwargs): csv_path = os.path.join(root_path, meta_file) csv_file = pd.read_csv(csv_path) + csv_file["char_count"] = csv_file.transcript.apply(lambda x: len(list(x))) + csv_file = csv_file[csv_file.char_count < 400].copy() csv_file["audio_paths"] = csv_file["audio_paths"].apply( lambda x: x.replace("/AfriSpeech-TTS-D/", root_path) ) csv_file = csv_file.rename(columns={"transcript":"text", "audio_paths":"audio_file", "user_ids":"speaker_name"}) + csv_file = csv_file[["audio_file", "text", "speaker_name"]] csv_file['root_path'] = root_path items = csv_file.to_dict('records') return items diff --git a/recipes/ljspeech/xtts_v2/README.md b/recipes/ljspeech/xtts_v2/README.md index d4f9af0cf0..15b6aa1a75 100644 --- a/recipes/ljspeech/xtts_v2/README.md +++ b/recipes/ljspeech/xtts_v2/README.md @@ -16,15 +16,21 @@ This repository contains the code for training a text-to-speech (TTS) model usin ```bash cd coqui-ai-TTS ``` +3. **Create a virtual environment with python version 3.10** -3. **Install system dependencies and the code:** + ```bash + conda create -n xtts python==3.10 + conda activate xtts + ``` + +4. **Install system dependencies and the code:** ```bash make system-deps # Intended to be used on Ubuntu (Debian). Let us know if you have a different OS. make install ``` -4. **Open the following file and redefine the specified variables:** +5. **Open the following file and redefine the specified variables:** File: `recipes/ljspeech/xtts_v2/train_gpt_xtts.py` @@ -50,4 +56,12 @@ From the repository root, run the following command: python3 recipes/ljspeech/xtts_v2/train_gpt_xtts.py ``` -You are now ready to train your TTS model using Coqui AI's framework. Enjoy! \ No newline at end of file +You are now ready to train your TTS model using Coqui AI's framework. Enjoy! + +## Optional: Resuming from a checkpoint + +File: `recipes/ljspeech/xtts_v2/train_gpt_xtts_resume.py` + +Update the parameters in the file for the models + + diff --git a/recipes/ljspeech/xtts_v2/train_gpt_xtts.py b/recipes/ljspeech/xtts_v2/train_gpt_xtts.py index d800bc2d65..a456cbe9bd 100644 --- a/recipes/ljspeech/xtts_v2/train_gpt_xtts.py +++ b/recipes/ljspeech/xtts_v2/train_gpt_xtts.py @@ -1,4 +1,6 @@ import os +os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" +os.environ["CUDA_VISIBLE_DEVICES"] = "0" from trainer import Trainer, TrainerArgs diff --git a/recipes/ljspeech/xtts_v2/train_gpt_xtts_resume_ft.py b/recipes/ljspeech/xtts_v2/train_gpt_xtts_resume_ft.py new file mode 100644 index 0000000000..3324bc2ec8 --- /dev/null +++ b/recipes/ljspeech/xtts_v2/train_gpt_xtts_resume_ft.py @@ -0,0 +1,171 @@ +import os +os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" +os.environ["CUDA_VISIBLE_DEVICES"] = "0" + +from trainer import Trainer, TrainerArgs + +from TTS.config.shared_configs import BaseDatasetConfig +from TTS.tts.datasets import load_tts_samples +from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig, XttsAudioConfig +from TTS.utils.manage import ModelManager + +# Logging parameters +RUN_NAME = "GPT_XTTS_v2.0_AfroTTS_FT" +PROJECT_NAME = "XTTS_trainer" +DASHBOARD_LOGGER = "tensorboard" +LOGGER_URI = None + +# Set here the path that the checkpoints will be saved. Default: ./run/training/ +OUT_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), "run", "training") + +# Training Parameters +OPTIMIZER_WD_ONLY_ON_WEIGHTS = True # for multi-gpu training please make it False +START_WITH_EVAL = False # if True it will star with evaluation +BATCH_SIZE = 2 # set here the batch size +GRAD_ACUMM_STEPS = 126 # set here the grad accumulation steps +# Note: we recommend that BATCH_SIZE * GRAD_ACUMM_STEPS need to be at least 252 for more efficient training. You can increase/decrease BATCH_SIZE but then set GRAD_ACUMM_STEPS accordingly. + +afrotts_dir = "AfriSpeech-TTS-D" # add path to afrotts data here + +# Define here the dataset that you want to use for the fine-tuning on. +config_dataset = BaseDatasetConfig( + formatter="afrotts", + dataset_name="afrotts", + path=afrotts_dir, + meta_file_train=os.path.join(afrotts_dir, "data/afritts-train-clean-upsamp.csv") #afritts-train-clean-upsamp.csv + meta_file_val=os.path.join(afrotts_dir,"data/afritts-dev-clean.csv"), + language="en", +) + +# Add here the configs of the datasets +DATASETS_CONFIG_LIST = [config_dataset] + +# Define the path where XTTS v2.0.1 files will be downloaded +CHECKPOINTS_OUT_PATH = os.path.join(OUT_PATH, "XTTS_v2.0_original_model_files/") +os.makedirs(CHECKPOINTS_OUT_PATH, exist_ok=True) + +# Set the path to the downloaded files +DVAE_CHECKPOINT = "coqui-ai-TTS/recipes/ljspeech/xtts_v2/run/training/XTTS_v2.0_original_model_files/dvae.pth" +MEL_NORM_FILE = "coqui-ai-TTS/recipes/ljspeech/xtts_v2/run/training/XTTS_v2.0_original_model_files/mel_stats.pth" + +# download DVAE files if needed +if not os.path.isfile(DVAE_CHECKPOINT) or not os.path.isfile(MEL_NORM_FILE): + print(" > Downloading DVAE files!") + ModelManager._download_model_files([MEL_NORM_LINK, DVAE_CHECKPOINT_LINK], CHECKPOINTS_OUT_PATH, progress_bar=True) + + +# XTTS transfer learning parameters: You we need to provide the paths of XTTS model checkpoint that you want to do the fine tuning. +TOKENIZER_FILE = "coqui-ai-TTS/recipes/ljspeech/xtts_v2/run/training/XTTS_v2.0_original_model_files/vocab.json" +XTTS_CHECKPOINT = "coqui-ai-TTS/recipes/ljspeech/xtts_v2/run/training/GPT_XTTS_v2.0_AfroTTS_FT-March-06-2024_06+36AM-581cf506/checkpoint_135000.pth" + + +# download XTTS v2.0 files if needed +if not os.path.isfile(TOKENIZER_FILE) or not os.path.isfile(XTTS_CHECKPOINT): + print(" > Downloading XTTS v2.0 files!") + ModelManager._download_model_files( + [TOKENIZER_FILE_LINK, XTTS_CHECKPOINT_LINK], CHECKPOINTS_OUT_PATH, progress_bar=True + ) + + +# Training sentences generations +SPEAKER_REFERENCE = [ + "/AfriSpeech-TTS/train/defc5e03-926c-4e0b-a639-c821e5e7db89/14f64f13c57f9a64a2a1521253934a0b_KYA8MaKS.wav" # speaker reference to be used in training test sentences +] +LANGUAGE = config_dataset.language + +def main(): + # init args and config + model_args = GPTArgs( + max_conditioning_length=132300, # 6 secs + min_conditioning_length=66150, # 3 secs + debug_loading_failures=False, + max_wav_length=255995, # ~11.6 seconds 661500, #~ 30 seconds # + max_text_length=300, + mel_norm_file=MEL_NORM_FILE, + dvae_checkpoint=DVAE_CHECKPOINT, + xtts_checkpoint=XTTS_CHECKPOINT, # checkpoint path of the model that you want to fine-tune + tokenizer_file=TOKENIZER_FILE, + gpt_num_audio_tokens=1026, + gpt_start_audio_token=1024, + gpt_stop_audio_token=1025, + gpt_use_masking_gt_prompt_approach=True, + gpt_use_perceiver_resampler=True, + ) + # define audio config + audio_config = XttsAudioConfig(sample_rate=22050, dvae_sample_rate=22050, output_sample_rate=24000) + # training parameters config + config = GPTTrainerConfig( + output_path=OUT_PATH, + model_args=model_args, + run_name=RUN_NAME, + project_name=PROJECT_NAME, + run_description=""" + GPT XTTS training + """, + dashboard_logger=DASHBOARD_LOGGER, + logger_uri=LOGGER_URI, + audio=audio_config, + batch_size=BATCH_SIZE, + batch_group_size=64, + eval_batch_size=BATCH_SIZE, + num_loader_workers=8, + eval_split_max_size=256, + print_step=1000, + plot_step=1000, + log_model_step=1000, + save_step=1000, + save_n_checkpoints=3, + save_checkpoints=True, + # target_loss="loss", + print_eval=True, + # Optimizer values like tortoise, pytorch implementation with modifications to not apply WD to non-weight parameters. + optimizer="AdamW", + optimizer_wd_only_on_weights=OPTIMIZER_WD_ONLY_ON_WEIGHTS, + optimizer_params={"betas": [0.9, 0.96], "eps": 1e-8, "weight_decay": 1e-2}, + lr=5e-06, # learning rate + lr_scheduler="MultiStepLR", + # it was adjusted accordly for the new step scheme + lr_scheduler_params={"milestones": [50000 * 18, 150000 * 18, 300000 * 18], "gamma": 0.5, "last_epoch": -1}, + test_sentences=[ + { + "text": "It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.", + "speaker_wav": SPEAKER_REFERENCE, + "language": LANGUAGE, + }, + { + "text": "This cake is great. It's so delicious and moist.", + "speaker_wav": SPEAKER_REFERENCE, + "language": LANGUAGE, + }, + ], + ) + + # init the model from config + model = GPTTrainer.init_from_config(config) + + # load training samples + train_samples, eval_samples = load_tts_samples( + DATASETS_CONFIG_LIST, + eval_split=True, + eval_split_max_size=config.eval_split_max_size, + eval_split_size=config.eval_split_size, + ) + # init the trainer and 🚀 + trainer = Trainer( + TrainerArgs( + restore_path=None, # xtts checkpoint is restored via xtts_checkpoint key so no need of restore it using Trainer restore_path parameter + skip_train_epoch=False, + start_with_eval=START_WITH_EVAL, + grad_accum_steps=GRAD_ACUMM_STEPS, + ), + config, + output_path=OUT_PATH, + model=model, + train_samples=train_samples, + eval_samples=eval_samples, + ) + trainer.fit() + + +if __name__ == "__main__": + main()