From d26909a1d6b3cecf2a8c7f2680ab2cb3d17aff8f Mon Sep 17 00:00:00 2001 From: erew123 <35898566+erew123@users.noreply.github.com> Date: Sun, 24 Dec 2023 12:32:10 +0000 Subject: [PATCH] All the updates for Finetuning Additional files to download Updated documentation Updated requirements files --- finetune.py | 1017 +++++++++++++++++++++++++++++++++++++ requirements_finetune.txt | 2 +- requirements_nvidia.txt | 2 +- requirements_other.txt | 2 +- tts_server.py | 77 +++ 5 files changed, 1097 insertions(+), 3 deletions(-) create mode 100644 finetune.py diff --git a/finetune.py b/finetune.py new file mode 100644 index 0000000..ac658a2 --- /dev/null +++ b/finetune.py @@ -0,0 +1,1017 @@ +import argparse +import os +import sys +import tempfile +import signal +import gradio as gr +import torch +import torchaudio +import traceback +from TTS.tts.configs.xtts_config import XttsConfig +from TTS.tts.models.xtts import Xtts +import random +import gc +import time +import shutil +import pandas +import json +from pathlib import Path +from tqdm import tqdm +from faster_whisper import WhisperModel +from TTS.tts.layers.xtts.tokenizer import multilingual_cleaners + + +# STARTUP VARIABLES +this_dir = Path(__file__).parent.resolve() +audio_folder = this_dir / "finetune" / "put-voice-samples-in-here" +out_path = this_dir / "finetune" / "tmp-trn" +progress = 0 +theme = gr.themes.Default() +refresh_symbol = '🔄' + +# Define the path to the modeldownload config file file +modeldownload_config_file_path = this_dir / "modeldownload.json" + +# Check if the JSON file exists +if modeldownload_config_file_path.exists(): + with open(modeldownload_config_file_path, "r") as config_file: + settings = json.load(config_file) + # Extract settings from the loaded JSON + base_path = Path(settings.get("base_path", "")) + model_path = Path(settings.get("model_path", "")) + files_to_download = settings.get("files_to_download", {}) +else: + # Default settings if the JSON file doesn't exist or is empty + print("[FINETUNE] \033[91mWarning\033[0m modeldownload.json is missing. Please run this script in the /alltalk_tts/ folder") + sys.exit(1) + +##################### +#### STEP 1 BITS #### +##################### + +def create_temp_folder(): + temp_folder = os.path.join(os.path.dirname(__file__), 'temp_files') + os.makedirs(temp_folder, exist_ok=True) + return temp_folder + +def create_temporary_file(folder, suffix=".wav"): + unique_filename = f"custom_tempfile_{int(time.time())}_{random.randint(1, 1000)}{suffix}" + return os.path.join(folder, unique_filename) + +def format_audio_list(target_language, whisper_model, out_path, gradio_progress=progress): + audio_files = [os.path.join(audio_folder, file) for file in os.listdir(audio_folder) if file.endswith('.wav')] + buffer=0.2 + eval_percentage=0.15 + speaker_name="coqui" + audio_total_size = 0 + os.makedirs(out_path, exist_ok=True) + temp_folder = os.path.join(out_path, "temp") # Update with your folder name + os.makedirs(temp_folder, exist_ok=True) + print("[FINETUNE] \033[94mPart of AllTalk\033[0m https://github.com/erew123/alltalk_tts/") + print("[FINETUNE] \033[94mCoqui Public Model License\033[0m") + print("[FINETUNE] \033[94mhttps://coqui.ai/cpml.txt\033[0m") + print("[FINETUNE] \033[94mStarting Step 1\033[0m - Preparing Audio/Generating the dataset") + # Write the target language to lang.txt in the output directory + lang_file_path = os.path.join(out_path, "lang.txt") + + # Check if lang.txt already exists and contains a different language + current_language = None + if os.path.exists(lang_file_path): + with open(lang_file_path, 'r', encoding='utf-8') as existing_lang_file: + current_language = existing_lang_file.read().strip() + + if current_language != target_language: + # Only update lang.txt if target language is different from the current language + with open(lang_file_path, 'w', encoding='utf-8') as lang_file: + lang_file.write(target_language + '\n') + print("[FINETUNE] Updated lang.txt with the target language.") + else: + print("[FINETUNE] The existing language matches the target language") + + # Loading Whisper + device = "cuda" if torch.cuda.is_available() else "cpu" + + print("[FINETUNE] Loading Whisper Model:", whisper_model) + asr_model = WhisperModel(whisper_model, device=device, compute_type="float16") + + metadata = {"audio_file": [], "text": [], "speaker_name": []} + + existing_metadata = {'train': None, 'eval': None} + train_metadata_path = os.path.join(out_path, "metadata_train.csv") + eval_metadata_path = os.path.join(out_path, "metadata_eval.csv") + + if os.path.exists(train_metadata_path): + existing_metadata['train'] = pandas.read_csv(train_metadata_path, sep="|") + print("[FINETUNE] Existing training metadata found and loaded.") + + if os.path.exists(eval_metadata_path): + existing_metadata['eval'] = pandas.read_csv(eval_metadata_path, sep="|") + print("[FINETUNE] Existing evaluation metadata found and loaded.") + + for idx, audio_path in tqdm(enumerate(audio_files)): + if isinstance(audio_path, str): + audio_file_name_without_ext, _ = os.path.splitext(os.path.basename(audio_path)) + # If it's a string, it's already the path to the file + audio_path_name = audio_path + elif hasattr(audio_path, 'read'): + # If it has a 'read' attribute, treat it as a file-like object + # and use a temporary file to save its content + audio_file_name_without_ext, _ = os.path.splitext(os.path.basename(audio_path.name)) + audio_path_name = create_temporary_file(temp_folder) + with open(audio_path, 'rb') as original_file: + file_content = original_file.read() + with open(audio_path_name, 'wb') as temp_file: + temp_file.write(file_content) + + # Create a temporary file path within the new folder + temp_audio_path = create_temporary_file(temp_folder) + + try: + if isinstance(audio_path, str): + audio_path_name = audio_path + elif hasattr(audio_path, 'name'): + audio_path_name = audio_path.name + else: + raise ValueError(f"Unsupported audio_path type: {type(audio_path)}") + except Exception as e: + print("[FINETUNE] Error reading original file: {e}") + # Handle the error or raise it if needed + print("[FINETUNE] Current working file:", audio_path_name) + try: + # Copy the audio content + time.sleep(0.5) # Introduce a small delay + shutil.copy2(audio_path_name, temp_audio_path) + except Exception as e: + print("[FINETUNE] Error copying file: {e}") + # Handle the error or raise it if needed + + # Load the temporary audio file + wav, sr = torchaudio.load(temp_audio_path, format="wav") + wav = torch.as_tensor(wav).clone().detach().t().to(torch.float32), sr + + prefix_check = f"wavs/{audio_file_name_without_ext}_" + + # Check both training and evaluation metadata for an entry that starts with the file name. + skip_processing = False + + for key in ['train', 'eval']: + if existing_metadata[key] is not None: + mask = existing_metadata[key]['audio_file'].str.startswith(prefix_check) + + if mask.any(): + print(f"[FINETUNE] Segments from {audio_file_name_without_ext} have been previously processed; skipping...") + skip_processing = True + break + + # If we found that we've already processed this file before, continue to the next iteration. + if skip_processing: + continue + + wav, sr = torchaudio.load(audio_path) + # stereo to mono if needed + if wav.size(0) != 1: + wav = torch.mean(wav, dim=0, keepdim=True) + + wav = wav.squeeze() + audio_total_size += (wav.size(-1) / sr) + + segments, _ = asr_model.transcribe(audio_path, vad_filter=True, word_timestamps=True, language=target_language) + segments = list(segments) + i = 0 + sentence = "" + sentence_start = None + first_word = True + # added all segments words in a unique list + words_list = [] + for _, segment in enumerate(segments): + words = list(segment.words) + words_list.extend(words) + + # process each word + for word_idx, word in enumerate(words_list): + if first_word: + sentence_start = word.start + # If it is the first sentence, add buffer or get the beginning of the file + if word_idx == 0: + sentence_start = max(sentence_start - buffer, 0) # Add buffer to the sentence start + else: + # get the previous sentence end + previous_word_end = words_list[word_idx - 1].end + # add buffer or get the silence middle between the previous sentence and the current one + sentence_start = max(sentence_start - buffer, (previous_word_end + sentence_start) / 2) + + sentence = word.word + first_word = False + else: + sentence += word.word + + if word.word[-1] in ["!", ".", "?"]: + sentence = sentence[1:] + # Expand number and abbreviations plus normalization + sentence = multilingual_cleaners(sentence, target_language) + audio_file_name, _ = os.path.splitext(os.path.basename(audio_path)) + + audio_file = f"wavs/{audio_file_name}_{str(i).zfill(8)}.wav" + + # Check for the next word's existence + if word_idx + 1 < len(words_list): + next_word_start = words_list[word_idx + 1].start + else: + # If don't have more words it means that it is the last sentence then use the audio len as next word start + next_word_start = (wav.shape[0] - 1) / sr + + # Average the current word end and next word start + word_end = min((word.end + next_word_start) / 2, word.end + buffer) + + absolute_path = os.path.join(out_path, audio_file) + os.makedirs(os.path.dirname(absolute_path), exist_ok=True) + i += 1 + first_word = True + + audio = wav[int(sr * sentence_start):int(sr * word_end)].unsqueeze(0) + # if the audio is too short, ignore it (i.e., < 0.33 seconds) + if audio.size(-1) >= sr / 3: + torchaudio.save( + absolute_path, + audio, + sr + ) + else: + continue + + metadata["audio_file"].append(audio_file) + metadata["text"].append(sentence) + metadata["speaker_name"].append(speaker_name) + + os.remove(temp_audio_path) + + if os.path.exists(train_metadata_path) and os.path.exists(eval_metadata_path): + existing_train_df = existing_metadata['train'] + existing_eval_df = existing_metadata['eval'] + audio_total_size = 121 + else: + existing_train_df = pandas.DataFrame(columns=["audio_file", "text", "speaker_name"]) + existing_eval_df = pandas.DataFrame(columns=["audio_file", "text", "speaker_name"]) + + new_data_df = pandas.DataFrame(metadata) + + combined_train_df = pandas.concat([existing_train_df, new_data_df], ignore_index=True).drop_duplicates().reset_index(drop=True) + + combined_train_df_shuffled = combined_train_df.sample(frac=1) + num_val_samples = int(len(combined_train_df_shuffled) * eval_percentage) + + final_eval_set = combined_train_df_shuffled[:num_val_samples] + final_training_set = combined_train_df_shuffled[num_val_samples:] + + final_training_set.sort_values('audio_file').to_csv(train_metadata_path, sep='|', index=False) + final_eval_set.sort_values('audio_file').to_csv(eval_metadata_path, sep='|', index=False) + + # deallocate VRAM and RAM + del asr_model, final_eval_set, final_training_set, new_data_df, existing_metadata + gc.collect() + print("[FINETUNE] Train CSV:", train_metadata_path) + print("[FINETUNE] Eval CSV:", eval_metadata_path) + print("[FINETUNE] Audio Total:", audio_total_size) + return train_metadata_path, eval_metadata_path, audio_total_size + +###################### +#### STEP 2 BITS ##### +###################### + +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 + + +def train_gpt(language, num_epochs, batch_size, grad_acumm, train_csv, eval_csv, output_path, max_audio_length=255995): + # Logging parameters + RUN_NAME = "XTTS_FT" + PROJECT_NAME = "XTTS_trainer" + DASHBOARD_LOGGER = "tensorboard" + LOGGER_URI = None + + # Set here the path that the checkpoints will be saved. Default: ./training/ + OUT_PATH = os.path.join(output_path, "training") + print("[FINETUNE] \033[94mStarting Step 2\033[0m - Fine-tuning the XTTS Encoder") + # Create the directory + os.makedirs(OUT_PATH, exist_ok=True) + + # 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 = batch_size # set here the batch size + GRAD_ACUMM_STEPS = grad_acumm # set here the grad accumulation steps + + # Define here the dataset that you want to use for the fine-tuning on. + config_dataset = BaseDatasetConfig( + formatter="coqui", + dataset_name="ft_dataset", + path=os.path.dirname(train_csv), + meta_file_train=train_csv, + meta_file_val=eval_csv, + language=language, + ) + + # Add here the configs of the datasets + DATASETS_CONFIG_LIST = [config_dataset] + + # XTTS model checkpoints for fine-tuning. + TOKENIZER_FILE = str(this_dir / base_path / model_path / "vocab.json") # vocab.json file + XTTS_CHECKPOINT = str(this_dir / base_path / model_path / "model.pth") # model.pth file + XTTS_CONFIG_FILE = str(this_dir / base_path / model_path / "config.json") # config.json file + DVAE_CHECKPOINT = str(this_dir / base_path / model_path / "dvae.pth") + MEL_NORM_FILE = str(this_dir / base_path / model_path / "mel_stats.pth") + + # 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=max_audio_length, # ~11.6 seconds + max_text_length=200, + 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( + epochs=num_epochs, + 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=48, + eval_batch_size=BATCH_SIZE, + num_loader_workers=8, + eval_split_max_size=256, + print_step=50, + plot_step=100, + log_model_step=100, + save_step=1000, + save_n_checkpoints=1, + save_checkpoints=True, + # target_loss="loss", + print_eval=False, + # 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=[], + ) + + # 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 + 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() + + # get the longest text audio file to use as speaker reference + samples_len = [len(item["text"].split(" ")) for item in train_samples] + longest_text_idx = samples_len.index(max(samples_len)) + speaker_ref = train_samples[longest_text_idx]["audio_file"] + + trainer_out_path = trainer.output_path + + # deallocate VRAM and RAM + del model, trainer, train_samples, eval_samples + gc.collect() + + return XTTS_CONFIG_FILE, XTTS_CHECKPOINT, TOKENIZER_FILE, trainer_out_path, speaker_ref + +########################## +#### STEP 3 AND OTHER #### +########################## + +def clear_gpu_cache(): + # clear the GPU cache + if torch.cuda.is_available(): + torch.cuda.empty_cache() + +XTTS_MODEL = None +def load_model(xtts_checkpoint, xtts_config, xtts_vocab): + global XTTS_MODEL + clear_gpu_cache() + if not xtts_checkpoint or not xtts_config or not xtts_vocab: + return "You need to run the previous steps or manually set the `XTTS checkpoint path`, `XTTS config path`, and `XTTS vocab path` fields !!" + config = XttsConfig() + config.load_json(xtts_config) + XTTS_MODEL = Xtts.init_from_config(config) + print("[FINETUNE] \033[94mStarting Step 3\033[0m Loading XTTS model!") + XTTS_MODEL.load_checkpoint(config, checkpoint_path=xtts_checkpoint, vocab_path=xtts_vocab, use_deepspeed=False) + if torch.cuda.is_available(): + XTTS_MODEL.cuda() + + print("[FINETUNE] Model Loaded!") + return "Model Loaded!" + +def run_tts(lang, tts_text, speaker_audio_file): + if XTTS_MODEL is None or not speaker_audio_file: + return "You need to run the previous step to load the model !!", None, None + speaker_audio_file = str(speaker_audio_file) + gpt_cond_latent, speaker_embedding = XTTS_MODEL.get_conditioning_latents(audio_path=speaker_audio_file, gpt_cond_len=XTTS_MODEL.config.gpt_cond_len, max_ref_length=XTTS_MODEL.config.max_ref_len, sound_norm_refs=XTTS_MODEL.config.sound_norm_refs) + out = XTTS_MODEL.inference( + text=tts_text, + language=lang, + gpt_cond_latent=gpt_cond_latent, + speaker_embedding=speaker_embedding, + temperature=XTTS_MODEL.config.temperature, # Add custom parameters here + length_penalty=XTTS_MODEL.config.length_penalty, + repetition_penalty=XTTS_MODEL.config.repetition_penalty, + top_k=XTTS_MODEL.config.top_k, + top_p=XTTS_MODEL.config.top_p, + ) + + with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as fp: + out["wav"] = torch.tensor(out["wav"]).unsqueeze(0) + out_path = fp.name + torchaudio.save(out_path, out["wav"], 24000) + + return "Speech generated !", out_path, speaker_audio_file + + +def get_available_voices(minimum_size_kb=1300): + voice_files = [ + voice for voice in Path(f"{this_dir}/finetune/tmp-trn/wavs").glob("*.wav") + if voice.stat().st_size > minimum_size_kb * 1300 # Convert KB to bytes + ] + return sorted([str(file) for file in voice_files]) # Return full path as string + +def find_models(directory, extension): + """Find files with a specific extension in the given directory.""" + return [file for file in Path(directory).rglob(f"*.{extension}")] + +def find_jsons(directory, filename): + """Find files with a specific filename in the given directory.""" + return list(Path(directory).rglob(filename)) + +# Your main directory +main_directory = Path(this_dir) / "finetune" / "tmp-trn" +# XTTS checkpoint files (*.pth) +xtts_checkpoint_files = find_models(main_directory, "pth") +# XTTS config files (config.json) +xtts_config_files = find_jsons(main_directory, "config.json") +# XTTS vocab files (vocab.json) +xtts_vocab_files = find_jsons(main_directory, "vocab.json") + + +# define a logger to redirect +class Logger: + def __init__(self, filename="finetune.log"): + self.log_file = filename + self.terminal = sys.stdout + self.log = open(self.log_file, "w") + + def write(self, message): + self.terminal.write(message) + self.log.write(message) + + def flush(self): + self.terminal.flush() + self.log.flush() + + def isatty(self): + return False + +# redirect stdout and stderr to a file +sys.stdout = Logger() +sys.stderr = sys.stdout + + +# logging.basicConfig(stream=sys.stdout, level=logging.INFO) +import logging +logging.basicConfig( + level=logging.INFO, + format="[FINETUNE] %(message)s", + handlers=[ + logging.StreamHandler(sys.stdout) + ] +) + + +def read_logs(): + sys.stdout.flush() + with open(sys.stdout.log_file, "r") as f: + return f.read() + + +def cleanup_before_exit(signum, frame): + print("[FINETUNE] Received interrupt signal. Cleaning up and exiting...") + # Perform cleanup operations here if necessary + sys.exit(0) + + +def create_refresh_button(refresh_component, refresh_method, refreshed_args, elem_class, interactive=True): + """ + Copied from https://github.com/AUTOMATIC1111/stable-diffusion-webui + """ + def refresh(): + refresh_method() + args = refreshed_args() if callable(refreshed_args) else refreshed_args + + for k, v in args.items(): + setattr(refresh_component, k, v) + + return gr.update(**(args or {})) + + refresh_button = gr.Button(refresh_symbol, elem_classes=elem_class, interactive=interactive, scale=0) + refresh_button.click( + fn=refresh, + inputs=[], + outputs=[refresh_component] + ) + + return refresh_button + + +if __name__ == "__main__": + # Register the signal handler + signal.signal(signal.SIGINT, cleanup_before_exit) + + +################ +#### GRADIO #### +################ + + parser = argparse.ArgumentParser( + description="""XTTS fine-tuning demo\n\n""" + """ + Example runs: + python3 TTS/demos/xtts_ft_demo/xtts_demo.py --port + """, + formatter_class=argparse.RawTextHelpFormatter, + ) + parser.add_argument( + "--num_epochs", + type=int, + help="Number of epochs to train. Default: 10", + default=10, + ) + parser.add_argument( + "--batch_size", + type=int, + help="Batch size. Default: 4", + default=4, + ) + parser.add_argument( + "--grad_acumm", + type=int, + help="Grad accumulation steps. Default: 1", + default=1, + ) + parser.add_argument( + "--max_audio_length", + type=int, + help="Max permitted audio size in seconds. Default: 11", + default=11, + ) + + args = parser.parse_args() + +##################### +#### GRADIO INFO #### +##################### + + with gr.Blocks(theme=theme) as demo: + with gr.Tab("Information/Guide"): + + gr.Markdown( + f""" + ## Finetuning Information
+ ### 🟥 Important Note + ####      🟥 finetune.py needs to be run from the /alltalk_tts/ folder. Don't move the location of this script. + ####      🟥 Have you run AllTalk at least once? It needs to have downloaded the voice model, before we can finetune it. + ### 🟦 What you need to run finetuning + ####      🟦 An Nvidia GPU. Tested on Windows with extended shared VRAM and training used about 16GB's total (which worked on a 12GB card). + ####      🟦 I have not been able to test this on a GPU with less than 12GB of VRAM, so cannot say if it will work or how that would affect performance. + ####      🟦 Version 11.8 of Nvidia cuBLAS and cuDNN (guide below). Only 11.8 of cuBLAS and cuDNN work for this process currently. + ####      🟦 Minimum 18GB free disk space (most of it is used temporarily). + ####      🟦 Some decent quality audio, multiple files if you like. Minimum of 2 minutes and Ive tested up to 20 minutes of audio. + ####      🟦 There is no major need to chop down your audio files into small slices as Step 1 will do that for you automatically and prepare the training set. Ive been testing with 5 minute long clips. + ####      🟦 This process will need access to all your GPU and VRAM, so close any other software that's using your GPU currently. + ### 🟨 Setting up cuBLAS and cuDNN Version 11.8 + ####      🟨 If you have the Nvidia CUDA Toolkit Version 11.8 installed and can type nvcc --version at the command prompt/terminal and it reports Cuda compilation tools, release 11.8 you should be good to go. + ####      🟨 If you dont have the toolkit installed, the idea is just to install the smallest bit possible and this will not affect or impact other things on your system. + ####      🟨 You will need to download the Nvidia Cuda Toolkit 11.8 network install from [here](https://developer.nvidia.com/cuda-11-8-0-download-archive) + ####          1) Run the installer and select Custom Advanced Uncheck everything at the top then expand CUDA, Development > Compiler > and select nvcc then expand Libraries and select CUBLAS. + ####          2) Back at the top of CUDA, expand Runtime > Libraries and select CUBLAS. Click Next, accept the default path (taking a note of its location) and let the install run. + ####          3) You should be able to drop to your terminal or command prompt and type nvcc --version and have it report Cuda compilation tools, release 11.8. If it does you are good to go. If it doesn't > Step 4. + ####          4) Linux users, you can temporarily add these paths on your current terminal window with (you may need to confirm these are correct for your flavour of Linux): + ####              export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64:${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}} (Add it to your ~/.bashrc if you want this to be permanent) + ####              export LD_LIBRARY_PATH=/usr/local/cuda-11.8/bin + ####          Windows users need the add the following to the PATH environment variable. Start menu and search for "Environment Variables" or "Edit the system environment variables.". + ####          Find and select the "Path" variable, then click on the "Edit...". Click on the "New" button and add: + ####              C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin. + ####          5) Once you have these set correctly, you should be able to open a new command prompt/terminal and nvcc --version at the command prompt/terminal, resulting in Cuda compilation tools, release 11.8. + ####          6) If the nvcc command doesn't work OR it reports a version different from 11.8, finetuning wont work, so you will to double check your environment variables and get them working correctly. + """ + ) + +####################### +#### GRADIO STEP 1 #### +####################### + with gr.Tab("Step 1 - Preparing Audio/Generating the dataset"): + + gr.Markdown( + f""" + ## STEP 1 - Preparing Audio/Generating the dataset
+ ### 🟦 What you need to do + ####      🟦 Please read Coqui's guide on what makes a good dataset [here](https://docs.coqui.ai/en/latest/what_makes_a_good_dataset.html#what-makes-a-good-dataset) + ####      🟦 Place your audio files in {str(audio_folder)} + ####      🟦 Your audio samples can be in the format mp3, wav, or flac. + ####      🟦 You will need a minimum of 2 minutes of audio in either one or multiple audio files. + ####      🟦 When you have completed Steps 1, 2, and 3, you are welcome to delete your samples from "put-voice-samples-in-here".
+ ### 🟨 What this step is doing + ####      🟨 With step one, we are going to be stripping your audio file(s) into smaller files, using Whisper to find spoken words/sentences, compile that into excel sheets of training data, ready for finetuning the model on Step 2. + ####      🟨 Whilst you can choose multiple Whisper models, its best only to use the 1 model as each one is about 3GB in size and will download to your local huggingface cache on first-time use. If and when you have completed training, you wish to delete this 3GB model from your system, you are welcome to do so. + ### 🟩 How long will this take? + ####      🟩 First time, it needs to download the Whisper model which is 3GB. After that a few minutes on average 3-4 year old system. + """ + ) + + out_path = gr.Textbox( + label="Output path (where data and checkpoints will be saved):", + value=out_path, + visible=False, + ) + + whisper_model = gr.Dropdown( + label="Whisper Model", + value="large-v3", + choices=[ + "large-v3", + "large-v2", + "large", + "medium", + "small" + ], + ) + + lang = gr.Dropdown( + label="Dataset Language", + value="en", + choices=[ + "en", + "es", + "fr", + "de", + "it", + "pt", + "pl", + "tr", + "ru", + "nl", + "cs", + "ar", + "zh", + "hu", + "ko", + "ja" + ], + ) + progress_data = gr.Label( + label="Progress:" + ) + logs = gr.Textbox( + label="Logs:", + interactive=False, + ) + demo.load(read_logs, None, logs, every=1) + + prompt_compute_btn = gr.Button(value="Step 1 - Create dataset") + + def preprocess_dataset(language, whisper_model, out_path, progress=gr.Progress(track_tqdm=True)): + clear_gpu_cache() + test_for_audio_files = [file for file in os.listdir(audio_folder) if any(file.lower().endswith(ext) for ext in ['.wav', '.mp3', '.flac'])] + if not test_for_audio_files: + return "I cannot find any mp3, wav or flac files in the folder called 'put-voice-samples-in-here'", "", "" + else: + try: + train_meta, eval_meta, audio_total_size = format_audio_list(target_language=language, whisper_model=whisper_model ,out_path=out_path, gradio_progress=progress) + except: + traceback.print_exc() + error = traceback.format_exc() + return f"The data processing was interrupted due an error !! Please check the console to verify the full error message! \n Error summary: {error}", "", "" + + clear_gpu_cache() + + # if audio total len is less than 2 minutes raise an error + if audio_total_size < 120: + message = "The total duration of the audio file or files you provided, was less than 2 minutes in length. Please add more audio samples." + print("[FINETUNE] ", message) + return message, "", "" + + print("[FINETUNE] Dataset Generated. Move to Step 2") + return "Dataset Generated. Move to Step 2", train_meta, eval_meta +####################### +#### GRADIO STEP 2 #### +####################### + with gr.Tab("Step 2 - Fine-tuning the XTTS Encoder"): + + gr.Markdown( + f""" + ## STEP 2 - Fine-tuning the XTTS Encoder
+ ### 🟦 What you need to do + ####      🟦 The Train CSV and Eval CSV should already be populated. If not, just go back to Step 1 and click the button again. + ####      🟦 If for any reason it crashed, close the app and re-start the finetuning and kick it off again, clicking through setp 1 (it doesnt need to regenerate the data) + ####      🟦 The settings below are factory and should be ok.
+ ### 🟨 What this step is doing + ####      🟨 Very simply put, it's taking all our wav files generated in Step 1, along with our recorded speech in out excel documents and its training the model on that voice. + ### 🟩 How long will this take? + ####      🟩 On a RTX 4070 with factory settings (as below) and 20 minutes of audio, this took 20 minutes. Again, it will vary by system and how much audio you are throwing at it. You have time to go make a coffee. + ####      🟩 Look at your command prompt/terminal window if you want to see what it is doing. + """ + ) + + train_csv = gr.Textbox( + label="Train CSV:", + ) + eval_csv = gr.Textbox( + label="Eval CSV:", + ) + num_epochs = gr.Slider( + label="Number of epochs:", + minimum=1, + maximum=100, + step=1, + value=args.num_epochs, + ) + batch_size = gr.Slider( + label="Batch size:", + minimum=2, + maximum=512, + step=1, + value=args.batch_size, + ) + grad_acumm = gr.Slider( + label="Grad accumulation steps:", + minimum=2, + maximum=128, + step=1, + value=args.grad_acumm, + ) + max_audio_length = gr.Slider( + label="Max permitted audio size in seconds:", + minimum=2, + maximum=20, + step=1, + value=args.max_audio_length, + ) + progress_train = gr.Label( + label="Progress:" + ) + logs_tts_train = gr.Textbox( + label="Logs:", + interactive=False, + ) + demo.load(read_logs, None, logs_tts_train, every=1) + train_btn = gr.Button(value="Step 2 - Run the training") + + def train_model(language, train_csv, eval_csv, num_epochs, batch_size, grad_acumm, output_path, max_audio_length): + clear_gpu_cache() + if not train_csv or not eval_csv: + return "You need to run the data processing step or manually set `Train CSV` and `Eval CSV` fields !", "", "", "", "" + try: + # convert seconds to waveform frames + max_audio_length = int(max_audio_length * 22050) + config_path, original_xtts_checkpoint, vocab_file, exp_path, speaker_wav = train_gpt(language, num_epochs, batch_size, grad_acumm, train_csv, eval_csv, output_path=str(output_path), max_audio_length=max_audio_length) + except: + traceback.print_exc() + error = traceback.format_exc() + return f"The training was interrupted due an error !! Please check the console to check the full error message! \n Error summary: {error}", "", "", "", "" + + # copy original files to avoid parameters changes issues + os.system(f"cp {config_path} {exp_path}") + os.system(f"cp {vocab_file} {exp_path}") + + ft_xtts_checkpoint = os.path.join(exp_path, "best_model.pth") + print("[FINETUNE] Model training done. Move to Step 3") + clear_gpu_cache() + return "Model training done. Move to Step 3", config_path, vocab_file, ft_xtts_checkpoint, speaker_wav + +####################### +#### GRADIO STEP 3 #### +####################### + + with gr.Tab("Step 3 - Inference/Testing"): + gr.Markdown( + f""" + ## STEP 3 - Inference/Testing
+ ### 🟥 Important Note + ####      🟥 This step will error if you are using TTS version 0.22.0. Please re-run pip install -r requirements_finetune.txt if it errors loading the model. + ### 🟦 What you need to do + ####      🟦 The model is now trained and you are at the testing stage. Hopefully all the dropdowns should be pre-populated now. + ####      🟦 You need to Load Fine-tuned XTTS model and then select your Speaker Reference Audio. You can choose various Speaker Reference Audios to see which works best. + ####      🟦 All the Speaker Reference Audios in the dropdown are ones that are 8 seconds long or more. You can use one of these later for your voice sample in All Talk, so remember the one you like. + ####      🟦 Some Speaker Reference Audios are a bit long and may cause issues generating the TTS, so try different ones. + ####      🟦 Once you are happy, move to the next tab for instruction on cleaning out the training data and how to copy your newly trained model over your current model. + ### 🟨 What this step is doing + ####      🟨 Its loading the finetuned model into memory, loading a voice sample and generating TTS, so that you can test out how well fine tuning worked. + """ + ) + + with gr.Row(): + with gr.Column() as col1: + xtts_checkpoint = gr.Dropdown( + [str(file) for file in xtts_checkpoint_files], + label="XTTS checkpoint path:", + value="", + allow_custom_value=True, + ) + + xtts_config = gr.Dropdown( + [str(file) for file in xtts_config_files], + label="XTTS config path:", + value="", + allow_custom_value=True, + ) + + xtts_vocab = gr.Dropdown( + [str(file) for file in xtts_vocab_files], + label="XTTS vocab path:", + value="", + allow_custom_value=True, + ) + progress_load = gr.Label( + label="Progress:" + ) + load_btn = gr.Button(value="Step 3 - Load Fine-tuned XTTS model") + + with gr.Column() as col2: + with gr.Row(): + # Gather the voice files + available_speaker_audios = get_available_voices() + + # Create Dropdown for speaker reference audio + speaker_reference_audio = gr.Dropdown( + available_speaker_audios, + label="Speaker reference audio:", + value="", # Set the default value if needed + allow_custom_value=True, # Allow custom values + scale=1, + ) + # Create refresh button + refresh_button = create_refresh_button( + speaker_reference_audio, + lambda: None, # Specify your refresh logic if needed + lambda: {"choices": get_available_voices(), "value": ""}, # Update choices and value + "refresh-button", + ) + + tts_language = gr.Dropdown( + label="Language", + value="en", + choices=[ + "en", + "es", + "fr", + "de", + "it", + "pt", + "pl", + "tr", + "ru", + "nl", + "cs", + "ar", + "zh", + "hu", + "ko", + "ja", + ] + ) + tts_text = gr.Textbox( + label="Input Text:", + value="I've just fine tuned a text to speech language model and this is how it sounds. If it doesn't sound right, I will try a different Speaker Reference Audio file.", + lines=6, + ) + tts_btn = gr.Button(value="Step 4 - Inference (Generate TTS)") + + with gr.Column() as col3: + progress_gen = gr.Label( + label="Progress:" + ) + tts_output_audio = gr.Audio(label="TTS Generated Audio.") + reference_audio = gr.Audio(label="Speaker Reference Audio Used.") + + with gr.Tab("What to do next"): + gr.Markdown( + f""" + ## What to do next
+ ### 🟦 What you need to do + ####      🟦 The wav(s) you liked from Speaker Reference Audios are in /finetune/tmp-trn/wavs/ so copy the one(s) you selected to your /alltalk_tts/voices/ samples folder. + ####      🟦 You can delete your original samples from inside of the /finetune/put-voice-samples-in-here/ folder. + ####      🟦 In the folder /finetune/tmp-trn/training/XTTS_FT(date here)/ you will find your best_model.pth, config.json and voice.json (also the last epoch training best_model_xxx.pth) + ####      🟦 You can delete the last epoch training model file best_model_xxx.pth (the one with the numbers). + ####      🟦 You will need to rename best_model.pth to model.pth. + ####      🟦 You will close the Finetuning application and copy your 3 files listed above over the top of the existing model files inside your /alltalk_tts/models/xttsv2_2.0.2/ folder. + ####      🟦 Start up Alltalk and you should be off and running. Preferably use the XTTS Local method for best results. + ### 🟨 Clearing up disk space + ####      🟨 When you are happy with everything, you can delete the /alltalk_tts/finetune/tmp-trn/ folder. This will delete all training data that was generated. + ####      🟨 If you are not going to train anything again, you can delete the whisper model from inside of your huggingface cache (3GB approx) + ####          Linux ~/.cache/huggingface/hub/(folder-here) + ####          Windows C:\users(your-username)\.cache\huggingface\hub\(folder-here). + ####      🟨 You can also uninstall the Nvidia CUDA 11.8 Toolkit if you wish and remove your environment variable entries. + """ + ) + + + prompt_compute_btn.click( + fn=preprocess_dataset, + inputs=[ + lang, + whisper_model, + out_path, + ], + outputs=[ + progress_data, + train_csv, + eval_csv, + ], + ) + + + train_btn.click( + fn=train_model, + inputs=[ + lang, + train_csv, + eval_csv, + num_epochs, + batch_size, + grad_acumm, + out_path, + max_audio_length, + ], + outputs=[progress_train, xtts_config, xtts_vocab, xtts_checkpoint, speaker_reference_audio], + ) + + load_btn.click( + fn=load_model, + inputs=[ + xtts_checkpoint, + xtts_config, + xtts_vocab + ], + outputs=[progress_load], + ) + + tts_btn.click( + fn=run_tts, + inputs=[ + tts_language, + tts_text, + speaker_reference_audio, + ], + outputs=[progress_gen, tts_output_audio, reference_audio], + ) + + demo.queue().launch( + show_api=False, + inbrowser=True, + share=False, + debug=False, + server_port=7052, + server_name="127.0.0.1", + ) + diff --git a/requirements_finetune.txt b/requirements_finetune.txt index 8d44517..7cd0a2d 100644 --- a/requirements_finetune.txt +++ b/requirements_finetune.txt @@ -1,7 +1,7 @@ gradio>=3.50.2 torch>=2.1.0+cu118 torchaudio>=2.1.0+cu118 -TTS>=0.22.0 +TTS==0.21.3 faster-whisper>=0.10.0 tqdm>=4.66.0 pandas>=1.5.0 \ No newline at end of file diff --git a/requirements_nvidia.txt b/requirements_nvidia.txt index 08244e4..64bdcfe 100644 --- a/requirements_nvidia.txt +++ b/requirements_nvidia.txt @@ -1,7 +1,7 @@ numpy>=1.22.0 soundfile>=0.12.1 uvicorn>=0.24.0.post1 -TTS>=0.22.0 +TTS>=0.21.3 torch>=2.1.0+cu118 torchaudio>=2.1.0+cu118 fastapi>=0.104.1 diff --git a/requirements_other.txt b/requirements_other.txt index 2754be7..667d885 100644 --- a/requirements_other.txt +++ b/requirements_other.txt @@ -1,7 +1,7 @@ numpy>=1.22.0 soundfile>=0.12.1 uvicorn>=0.24.0.post1 -TTS>=0.22.0 +TTS>=0.21.3 torch>=2.1.0 torchaudio>=2.1.0 fastapi>=0.104.1 diff --git a/tts_server.py b/tts_server.py index c4e9899..938bbc6 100644 --- a/tts_server.py +++ b/tts_server.py @@ -1033,6 +1033,7 @@ simple_webpage = """
  • Using Voice Samples
  • Text Not inside
  • Low VRAM
    • +
  • Finetuning (Training the model)
  • DeepSpeed