xinference 1.0.1__py3-none-any.whl → 1.2.1__py3-none-any.whl

xinference/thirdparty/f5_tts/train/datasets/prepare_ljspeech.py

@@ -0,0 +1,65 @@
+import os
+import sys
+
+sys.path.append(os.getcwd())
+
+import json
+from importlib.resources import files
+from pathlib import Path
+from tqdm import tqdm
+import soundfile as sf
+from datasets.arrow_writer import ArrowWriter
+
+
+def main():
+    result = []
+    duration_list = []
+    text_vocab_set = set()
+
+    with open(meta_info, "r") as f:
+        lines = f.readlines()
+        for line in tqdm(lines):
+            uttr, text, norm_text = line.split("|")
+            norm_text = norm_text.strip()
+            wav_path = Path(dataset_dir) / "wavs" / f"{uttr}.wav"
+            duration = sf.info(wav_path).duration
+            if duration < 0.4 or duration > 30:
+                continue
+            result.append({"audio_path": str(wav_path), "text": norm_text, "duration": duration})
+            duration_list.append(duration)
+            text_vocab_set.update(list(norm_text))
+
+    # save preprocessed dataset to disk
+    if not os.path.exists(f"{save_dir}"):
+        os.makedirs(f"{save_dir}")
+    print(f"\nSaving to {save_dir} ...")
+
+    with ArrowWriter(path=f"{save_dir}/raw.arrow") as writer:
+        for line in tqdm(result, desc="Writing to raw.arrow ..."):
+            writer.write(line)
+
+    # dup a json separately saving duration in case for DynamicBatchSampler ease
+    with open(f"{save_dir}/duration.json", "w", encoding="utf-8") as f:
+        json.dump({"duration": duration_list}, f, ensure_ascii=False)
+
+    # vocab map, i.e. tokenizer
+    # add alphabets and symbols (optional, if plan to ft on de/fr etc.)
+    with open(f"{save_dir}/vocab.txt", "w") as f:
+        for vocab in sorted(text_vocab_set):
+            f.write(vocab + "\n")
+
+    print(f"\nFor {dataset_name}, sample count: {len(result)}")
+    print(f"For {dataset_name}, vocab size is: {len(text_vocab_set)}")
+    print(f"For {dataset_name}, total {sum(duration_list)/3600:.2f} hours")
+
+
+if __name__ == "__main__":
+    tokenizer = "char"  # "pinyin" | "char"
+
+    dataset_dir = "<SOME_PATH>/LJSpeech-1.1"
+    dataset_name = f"LJSpeech_{tokenizer}"
+    meta_info = os.path.join(dataset_dir, "metadata.csv")
+    save_dir = str(files("f5_tts").joinpath("../../")) + f"/data/{dataset_name}"
+    print(f"\nPrepare for {dataset_name}, will save to {save_dir}\n")
+
+    main()

xinference/thirdparty/f5_tts/train/datasets/prepare_wenetspeech4tts.py

@@ -0,0 +1,125 @@
+# generate audio text map for WenetSpeech4TTS
+# evaluate for vocab size
+
+import os
+import sys
+
+sys.path.append(os.getcwd())
+
+import json
+from concurrent.futures import ProcessPoolExecutor
+from importlib.resources import files
+from tqdm import tqdm
+
+import torchaudio
+from datasets import Dataset
+
+from f5_tts.model.utils import convert_char_to_pinyin
+
+
+def deal_with_sub_path_files(dataset_path, sub_path):
+    print(f"Dealing with: {sub_path}")
+
+    text_dir = os.path.join(dataset_path, sub_path, "txts")
+    audio_dir = os.path.join(dataset_path, sub_path, "wavs")
+    text_files = os.listdir(text_dir)
+
+    audio_paths, texts, durations = [], [], []
+    for text_file in tqdm(text_files):
+        with open(os.path.join(text_dir, text_file), "r", encoding="utf-8") as file:
+            first_line = file.readline().split("\t")
+        audio_nm = first_line[0]
+        audio_path = os.path.join(audio_dir, audio_nm + ".wav")
+        text = first_line[1].strip()
+
+        audio_paths.append(audio_path)
+
+        if tokenizer == "pinyin":
+            texts.extend(convert_char_to_pinyin([text], polyphone=polyphone))
+        elif tokenizer == "char":
+            texts.append(text)
+
+        audio, sample_rate = torchaudio.load(audio_path)
+        durations.append(audio.shape[-1] / sample_rate)
+
+    return audio_paths, texts, durations
+
+
+def main():
+    assert tokenizer in ["pinyin", "char"]
+
+    audio_path_list, text_list, duration_list = [], [], []
+
+    executor = ProcessPoolExecutor(max_workers=max_workers)
+    futures = []
+    for dataset_path in dataset_paths:
+        sub_items = os.listdir(dataset_path)
+        sub_paths = [item for item in sub_items if os.path.isdir(os.path.join(dataset_path, item))]
+        for sub_path in sub_paths:
+            futures.append(executor.submit(deal_with_sub_path_files, dataset_path, sub_path))
+    for future in tqdm(futures, total=len(futures)):
+        audio_paths, texts, durations = future.result()
+        audio_path_list.extend(audio_paths)
+        text_list.extend(texts)
+        duration_list.extend(durations)
+    executor.shutdown()
+
+    if not os.path.exists("data"):
+        os.makedirs("data")
+
+    print(f"\nSaving to {save_dir} ...")
+    dataset = Dataset.from_dict({"audio_path": audio_path_list, "text": text_list, "duration": duration_list})
+    dataset.save_to_disk(f"{save_dir}/raw", max_shard_size="2GB")  # arrow format
+
+    with open(f"{save_dir}/duration.json", "w", encoding="utf-8") as f:
+        json.dump(
+            {"duration": duration_list}, f, ensure_ascii=False
+        )  # dup a json separately saving duration in case for DynamicBatchSampler ease
+
+    print("\nEvaluating vocab size (all characters and symbols / all phonemes) ...")
+    text_vocab_set = set()
+    for text in tqdm(text_list):
+        text_vocab_set.update(list(text))
+
+    # add alphabets and symbols (optional, if plan to ft on de/fr etc.)
+    if tokenizer == "pinyin":
+        text_vocab_set.update([chr(i) for i in range(32, 127)] + [chr(i) for i in range(192, 256)])
+
+    with open(f"{save_dir}/vocab.txt", "w") as f:
+        for vocab in sorted(text_vocab_set):
+            f.write(vocab + "\n")
+    print(f"\nFor {dataset_name}, sample count: {len(text_list)}")
+    print(f"For {dataset_name}, vocab size is: {len(text_vocab_set)}\n")
+
+
+if __name__ == "__main__":
+    max_workers = 32
+
+    tokenizer = "pinyin"  # "pinyin" | "char"
+    polyphone = True
+    dataset_choice = 1  # 1: Premium, 2: Standard, 3: Basic
+
+    dataset_name = (
+        ["WenetSpeech4TTS_Premium", "WenetSpeech4TTS_Standard", "WenetSpeech4TTS_Basic"][dataset_choice - 1]
+        + "_"
+        + tokenizer
+    )
+    dataset_paths = [
+        "<SOME_PATH>/WenetSpeech4TTS/Basic",
+        "<SOME_PATH>/WenetSpeech4TTS/Standard",
+        "<SOME_PATH>/WenetSpeech4TTS/Premium",
+    ][-dataset_choice:]
+    save_dir = str(files("f5_tts").joinpath("../../")) + f"/data/{dataset_name}"
+    print(f"\nChoose Dataset: {dataset_name}, will save to {save_dir}\n")
+
+    main()
+
+    # Results (if adding alphabets with accents and symbols):
+    # WenetSpeech4TTS       Basic     Standard     Premium
+    # samples count       3932473      1941220      407494
+    # pinyin vocab size      1349         1348        1344   (no polyphone)
+    #                           -            -        1459   (polyphone)
+    # char   vocab size      5264         5219        5042
+
+    # vocab size may be slightly different due to jieba tokenizer and pypinyin (e.g. way of polyphoneme)
+    # please be careful if using pretrained model, make sure the vocab.txt is same

xinference/thirdparty/f5_tts/train/finetune_cli.py

@@ -0,0 +1,174 @@
+import argparse
+import os
+import shutil
+
+from cached_path import cached_path
+from f5_tts.model import CFM, UNetT, DiT, Trainer
+from f5_tts.model.utils import get_tokenizer
+from f5_tts.model.dataset import load_dataset
+from importlib.resources import files
+
+
+# -------------------------- Dataset Settings --------------------------- #
+target_sample_rate = 24000
+n_mel_channels = 100
+hop_length = 256
+win_length = 1024
+n_fft = 1024
+mel_spec_type = "vocos"  # 'vocos' or 'bigvgan'
+
+
+# -------------------------- Argument Parsing --------------------------- #
+def parse_args():
+    # batch_size_per_gpu = 1000 settting for gpu 8GB
+    # batch_size_per_gpu = 1600 settting for gpu 12GB
+    # batch_size_per_gpu = 2000 settting for gpu 16GB
+    # batch_size_per_gpu = 3200 settting for gpu 24GB
+
+    # num_warmup_updates = 300 for 5000 sample about 10 hours
+
+    # change save_per_updates , last_per_steps change this value what you need  ,
+
+    parser = argparse.ArgumentParser(description="Train CFM Model")
+
+    parser.add_argument(
+        "--exp_name", type=str, default="F5TTS_Base", choices=["F5TTS_Base", "E2TTS_Base"], help="Experiment name"
+    )
+    parser.add_argument("--dataset_name", type=str, default="Emilia_ZH_EN", help="Name of the dataset to use")
+    parser.add_argument("--learning_rate", type=float, default=1e-5, help="Learning rate for training")
+    parser.add_argument("--batch_size_per_gpu", type=int, default=3200, help="Batch size per GPU")
+    parser.add_argument(
+        "--batch_size_type", type=str, default="frame", choices=["frame", "sample"], help="Batch size type"
+    )
+    parser.add_argument("--max_samples", type=int, default=64, help="Max sequences per batch")
+    parser.add_argument("--grad_accumulation_steps", type=int, default=1, help="Gradient accumulation steps")
+    parser.add_argument("--max_grad_norm", type=float, default=1.0, help="Max gradient norm for clipping")
+    parser.add_argument("--epochs", type=int, default=100, help="Number of training epochs")
+    parser.add_argument("--num_warmup_updates", type=int, default=300, help="Warmup steps")
+    parser.add_argument("--save_per_updates", type=int, default=10000, help="Save checkpoint every X steps")
+    parser.add_argument("--last_per_steps", type=int, default=50000, help="Save last checkpoint every X steps")
+    parser.add_argument("--finetune", type=bool, default=True, help="Use Finetune")
+    parser.add_argument("--pretrain", type=str, default=None, help="the path to the checkpoint")
+    parser.add_argument(
+        "--tokenizer", type=str, default="pinyin", choices=["pinyin", "char", "custom"], help="Tokenizer type"
+    )
+    parser.add_argument(
+        "--tokenizer_path",
+        type=str,
+        default=None,
+        help="Path to custom tokenizer vocab file (only used if tokenizer = 'custom')",
+    )
+    parser.add_argument(
+        "--log_samples",
+        type=bool,
+        default=False,
+        help="Log inferenced samples per ckpt save steps",
+    )
+    parser.add_argument("--logger", type=str, default=None, choices=["wandb", "tensorboard"], help="logger")
+    parser.add_argument(
+        "--bnb_optimizer",
+        type=bool,
+        default=False,
+        help="Use 8-bit Adam optimizer from bitsandbytes",
+    )
+
+    return parser.parse_args()
+
+
+# -------------------------- Training Settings -------------------------- #
+
+
+def main():
+    args = parse_args()
+
+    checkpoint_path = str(files("f5_tts").joinpath(f"../../ckpts/{args.dataset_name}"))
+
+    # Model parameters based on experiment name
+    if args.exp_name == "F5TTS_Base":
+        wandb_resume_id = None
+        model_cls = DiT
+        model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
+        if args.finetune:
+            if args.pretrain is None:
+                ckpt_path = str(cached_path("hf://SWivid/F5-TTS/F5TTS_Base/model_1200000.pt"))
+            else:
+                ckpt_path = args.pretrain
+    elif args.exp_name == "E2TTS_Base":
+        wandb_resume_id = None
+        model_cls = UNetT
+        model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
+        if args.finetune:
+            if args.pretrain is None:
+                ckpt_path = str(cached_path("hf://SWivid/E2-TTS/E2TTS_Base/model_1200000.pt"))
+            else:
+                ckpt_path = args.pretrain
+
+    if args.finetune:
+        if not os.path.isdir(checkpoint_path):
+            os.makedirs(checkpoint_path, exist_ok=True)
+
+        file_checkpoint = os.path.join(checkpoint_path, os.path.basename(ckpt_path))
+        if not os.path.isfile(file_checkpoint):
+            shutil.copy2(ckpt_path, file_checkpoint)
+            print("copy checkpoint for finetune")
+
+    # Use the tokenizer and tokenizer_path provided in the command line arguments
+    tokenizer = args.tokenizer
+    if tokenizer == "custom":
+        if not args.tokenizer_path:
+            raise ValueError("Custom tokenizer selected, but no tokenizer_path provided.")
+        tokenizer_path = args.tokenizer_path
+    else:
+        tokenizer_path = args.dataset_name
+
+    vocab_char_map, vocab_size = get_tokenizer(tokenizer_path, tokenizer)
+
+    print("\nvocab : ", vocab_size)
+    print("\nvocoder : ", mel_spec_type)
+
+    mel_spec_kwargs = dict(
+        n_fft=n_fft,
+        hop_length=hop_length,
+        win_length=win_length,
+        n_mel_channels=n_mel_channels,
+        target_sample_rate=target_sample_rate,
+        mel_spec_type=mel_spec_type,
+    )
+
+    model = CFM(
+        transformer=model_cls(**model_cfg, text_num_embeds=vocab_size, mel_dim=n_mel_channels),
+        mel_spec_kwargs=mel_spec_kwargs,
+        vocab_char_map=vocab_char_map,
+    )
+
+    trainer = Trainer(
+        model,
+        args.epochs,
+        args.learning_rate,
+        num_warmup_updates=args.num_warmup_updates,
+        save_per_updates=args.save_per_updates,
+        checkpoint_path=checkpoint_path,
+        batch_size=args.batch_size_per_gpu,
+        batch_size_type=args.batch_size_type,
+        max_samples=args.max_samples,
+        grad_accumulation_steps=args.grad_accumulation_steps,
+        max_grad_norm=args.max_grad_norm,
+        logger=args.logger,
+        wandb_project=args.dataset_name,
+        wandb_run_name=args.exp_name,
+        wandb_resume_id=wandb_resume_id,
+        log_samples=args.log_samples,
+        last_per_steps=args.last_per_steps,
+        bnb_optimizer=args.bnb_optimizer,
+    )
+
+    train_dataset = load_dataset(args.dataset_name, tokenizer, mel_spec_kwargs=mel_spec_kwargs)
+
+    trainer.train(
+        train_dataset,
+        resumable_with_seed=666,  # seed for shuffling dataset
+    )
+
+
+if __name__ == "__main__":
+    main()