Update README.md

Ubuntu 20.04.1 CUDA 11.3 缺少依赖，还有依赖冲突

Co-authored-by: Vega <babysor00@gmail.com>

README-CN.md

@@ -29,6 +29,7 @@
 > 如果在用 pip 方式安装的时候出现 `ERROR: Could not find a version that satisfies the requirement torch==1.9.0+cu102 (from versions: 0.1.2, 0.1.2.post1, 0.1.2.post2)` 这个错误可能是 python 版本过低，3.9 可以安装成功
 * 安装 [ffmpeg](https://ffmpeg.org/download.html#get-packages)。
 * 运行`pip install -r requirements.txt` 来安装剩余的必要包。
+> 这里的环境建议使用 `Repo Tag 0.0.1` `Pytorch1.9.0 with Torchvision0.10.0 and cudatoolkit10.2` `requirements.txt` `webrtcvad-wheels` 因为 `requiremants.txt` 是在几个月前导出的，所以不适配新版本
 * 安装 webrtcvad `pip install webrtcvad-wheels`。
 
 或者

README.md

@@ -1,3 +1,5 @@
+> 🚧 While I no longer actively update this repo, you can find me continuously pushing this tech forward to good side and open-source. Join me at [MaskGCT](https://github.com/open-mmlab/Amphion/tree/main/models/tts/maskgct). I'm also building an optimized and cloud hosted version: https://noiz.ai/
+>
 ![mockingbird](https://user-images.githubusercontent.com/12797292/131216767-6eb251d6-14fc-4951-8324-2722f0cd4c63.jpg)
 
 
@@ -29,6 +31,7 @@
 > If you get an `ERROR: Could not find a version that satisfies the requirement torch==1.9.0+cu102 (from versions: 0.1.2, 0.1.2.post1, 0.1.2.post2 )` This error is probably due to a low version of python, try using 3.9 and it will install successfully
 * Install [ffmpeg](https://ffmpeg.org/download.html#get-packages).
 * Run `pip install -r requirements.txt` to install the remaining necessary packages.
+> The recommended environment here is `Repo Tag 0.0.1` `Pytorch1.9.0 with Torchvision0.10.0 and cudatoolkit10.2` `requirements.txt` `webrtcvad-wheels` because `requirements. txt` was exported a few months ago, so it doesn't work with newer versions
 * Install webrtcvad `pip install webrtcvad-wheels`(If you need)
 
 or

models/synthesizer/preprocess.py

@@ -39,6 +39,9 @@ data_info = {
     }
 }
 
+def should_skip(fpath: Path, skip_existing: bool) -> bool:
+    return skip_existing and fpath.exists()
+
 def preprocess_dataset(datasets_root: Path, out_dir: Path, n_processes: int,
                            skip_existing: bool, hparams, no_alignments: bool, 
                            dataset: str, emotion_extract = False, encoder_model_fpath=None):
@@ -99,7 +102,7 @@ def preprocess_dataset(datasets_root: Path, out_dir: Path, n_processes: int,
     print("Max mel frames length: %d" % max(int(m[4]) for m in metadata))
     print("Max audio timesteps length: %d" % max(int(m[3]) for m in metadata))
 
-def embed_utterance(fpaths, encoder_model_fpath):
+def _embed_utterance(fpaths: str, encoder_model_fpath: str):
     if not encoder.is_loaded():
         encoder.load_model(encoder_model_fpath)
 
@@ -110,15 +113,13 @@ def embed_utterance(fpaths, encoder_model_fpath):
     embed = encoder.embed_utterance(wav)
     np.save(embed_fpath, embed, allow_pickle=False)
     
-def _emo_extract_from_utterance(fpaths, hparams, skip_existing=False):
-    if skip_existing and fpaths.exists():
-        return
+def _emo_extract_from_utterance(fpaths, hparams):
     wav_fpath, emo_fpath = fpaths
     wav = np.load(wav_fpath)
     emo = extract_emo(np.expand_dims(wav, 0), hparams.sample_rate, True)
     np.save(emo_fpath, emo.squeeze(0), allow_pickle=False)
  
-def create_embeddings(synthesizer_root: Path, encoder_model_fpath: Path, n_processes: int):
+def create_embeddings(synthesizer_root: Path, encoder_model_fpath: Path, n_processes: int, skip_existing: bool):
     wav_dir = synthesizer_root.joinpath("audio")
     metadata_fpath = synthesizer_root.joinpath("train.txt")
     assert wav_dir.exists() and metadata_fpath.exists()
@@ -128,11 +129,11 @@ def create_embeddings(synthesizer_root: Path, encoder_model_fpath: Path, n_proce
     # Gather the input wave filepath and the target output embed filepath
     with metadata_fpath.open("r", encoding="utf-8") as metadata_file:
         metadata = [line.split("|") for line in metadata_file]
-        fpaths = [(wav_dir.joinpath(m[0]), embed_dir.joinpath(m[2])) for m in metadata]
-        
+        fpaths = [(wav_dir.joinpath(m[0]), embed_dir.joinpath(m[2])) for m in metadata if not should_skip(embed_dir.joinpath(m[2]), skip_existing)]
+
     # TODO: improve on the multiprocessing, it's terrible. Disk I/O is the bottleneck here.
     # Embed the utterances in separate threads
-    func = partial(embed_utterance, encoder_model_fpath=encoder_model_fpath)
+    func = partial(_embed_utterance, encoder_model_fpath=encoder_model_fpath)
     job = Pool(n_processes).imap(func, fpaths)
     tuple(tqdm(job, "Embedding", len(fpaths), unit="utterances"))
 
@@ -142,14 +143,14 @@ def create_emo(synthesizer_root: Path, n_processes: int, skip_existing: bool, hp
     assert wav_dir.exists() and metadata_fpath.exists()
     emo_dir = synthesizer_root.joinpath("emo")
     emo_dir.mkdir(exist_ok=True)
-    
+
     # Gather the input wave filepath and the target output embed filepath
     with metadata_fpath.open("r", encoding="utf-8") as metadata_file:
         metadata = [line.split("|") for line in metadata_file]
-        fpaths = [(wav_dir.joinpath(m[0]), emo_dir.joinpath(m[0].replace("audio-", "emo-"))) for m in metadata]
+        fpaths = [(wav_dir.joinpath(m[0]), emo_dir.joinpath(m[0].replace("audio-", "emo-"))) for m in metadata if not should_skip(emo_dir.joinpath(m[0].replace("audio-", "emo-")), skip_existing)]
         
     # TODO: improve on the multiprocessing, it's terrible. Disk I/O is the bottleneck here.
     # Embed the utterances in separate threads
-    func = partial(_emo_extract_from_utterance, hparams=hparams, skip_existing=skip_existing)
+    func = partial(_emo_extract_from_utterance, hparams=hparams)
     job = Pool(n_processes).imap(func, fpaths)
     tuple(tqdm(job, "Emo", len(fpaths), unit="utterances"))

models/synthesizer/preprocess_audio.py

@@ -45,7 +45,7 @@ def extract_emo(
     return y
 
 def _process_utterance(wav: np.ndarray, text: str, out_dir: Path, basename: str, 
-                      skip_existing: bool, hparams, encoder_model_fpath):
+                      mel_fpath: str, wav_fpath: str, hparams, encoder_model_fpath):
     ## FOR REFERENCE:
     # For you not to lose your head if you ever wish to change things here or implement your own
     # synthesizer.
@@ -58,13 +58,6 @@ def _process_utterance(wav: np.ndarray, text: str, out_dir: Path, basename: str,
     #   without extra padding. This means that you won't have an exact relation between the length
     #   of the wav and of the mel spectrogram. See the vocoder data loader.
         
-    # Skip existing utterances if needed
-    mel_fpath = out_dir.joinpath("mels", "mel-%s.npy" % basename)
-    wav_fpath = out_dir.joinpath("audio", "audio-%s.npy" % basename)
-    
-    if skip_existing and mel_fpath.exists() and wav_fpath.exists():
-        return None
-
     # Trim silence
     if hparams.trim_silence:
         if not encoder.is_loaded():
@@ -112,50 +105,27 @@ def _split_on_silences(wav_fpath, words, hparams):
 def preprocess_general(speaker_dir, out_dir: Path, skip_existing: bool, hparams, dict_info, no_alignments: bool, encoder_model_fpath: Path):
     metadata = []
     extensions = ("*.wav", "*.flac", "*.mp3")
-    if skip_existing:
-        for extension in extensions:
-            wav_fpath_list = speaker_dir.glob(extension)
-            # Iterate over each wav
-            for wav_fpath in wav_fpath_list:
-                words = dict_info.get(wav_fpath.name.split(".")[0])
+    for extension in extensions:
+        wav_fpath_list = speaker_dir.glob(extension)
+        # Iterate over each wav
+        for wav_fpath in wav_fpath_list:
+            words = dict_info.get(wav_fpath.name.split(".")[0])
+            if not words:
+                words = dict_info.get(wav_fpath.name) # try with extension 
                 if not words:
-                    words = dict_info.get(wav_fpath.name) # try with extension 
-                    if not words:
-                        print("no wordS")
-                        continue
-                sub_basename = "%s_%02d" % (wav_fpath.name, 0)
-                
-                mel_fpath = out_dir.joinpath("mels", f"mel-{sub_basename}.npy")
-                wav_fpath_ = out_dir.joinpath("audio", f"audio-{sub_basename}.npy")
-                
-                if mel_fpath.exists() and wav_fpath_.exists():
+                    print(f"No word found in dict_info for {wav_fpath.name}, skip it")
                     continue
+            sub_basename = "%s_%02d" % (wav_fpath.name, 0)
+            mel_fpath_out = out_dir.joinpath("mels", f"mel-{sub_basename}.npy")
+            wav_fpath_out = out_dir.joinpath("audio", f"audio-{sub_basename}.npy")
+            
+            if skip_existing and mel_fpath_out.exists() and wav_fpath_out.exists():
+                continue
+            wav, text = _split_on_silences(wav_fpath, words, hparams)
+            result = _process_utterance(wav, text, out_dir, sub_basename, mel_fpath_out, wav_fpath_out, hparams, encoder_model_fpath)
+            if result is None:
+                continue
+            wav_fpath_name, mel_fpath_name, embed_fpath_name, wav, mel_frames, text = result
+            metadata.append ((wav_fpath_name, mel_fpath_name, embed_fpath_name, len(wav), mel_frames, text))
 
-                wav, text = _split_on_silences(wav_fpath, words, hparams)
-                result = _process_utterance(wav, text, out_dir, sub_basename, 
-                                                    False, hparams, encoder_model_fpath) # accelarate
-                if result is None:
-                    continue
-                wav_fpath_name, mel_fpath_name, embed_fpath_name, wav, mel_frames, text = result
-                metadata.append ((wav_fpath_name, mel_fpath_name, embed_fpath_name, len(wav), mel_frames, text))
-    else:
-        for extension in extensions:
-            wav_fpath_list = speaker_dir.glob(extension)
-            # Iterate over each wav
-            for wav_fpath in wav_fpath_list:
-                words = dict_info.get(wav_fpath.name.split(".")[0])
-                if not words:
-                    words = dict_info.get(wav_fpath.name) # try with extension 
-                    if not words:
-                        print("no wordS")
-                        continue
-                sub_basename = "%s_%02d" % (wav_fpath.name, 0)
-
-                wav, text = _split_on_silences(wav_fpath, words, hparams)
-                result = _process_utterance(wav, text, out_dir, sub_basename, 
-                                                    False, hparams, encoder_model_fpath)
-                if result is None:
-                    continue
-                wav_fpath_name, mel_fpath_name, embed_fpath_name, wav, mel_frames, text = result
-                metadata.append ((wav_fpath_name, mel_fpath_name, embed_fpath_name, len(wav), mel_frames, text))
     return metadata

pre.py

@@ -71,7 +71,7 @@ if __name__ == "__main__":
     del args.n_processes_embed
     preprocess_dataset(**vars(args))
     
-    create_embeddings(synthesizer_root=args.out_dir, n_processes=n_processes_embed, encoder_model_fpath=encoder_model_fpath)
+    create_embeddings(synthesizer_root=args.out_dir, n_processes=n_processes_embed, encoder_model_fpath=encoder_model_fpath, skip_existing=args.skip_existing)
     
     if args.emotion_extract:
         create_emo(synthesizer_root=args.out_dir, n_processes=n_processes_embed, skip_existing=args.skip_existing, hparams=args.hparams)

requirements.txt

@@ -2,7 +2,8 @@ umap-learn
 visdom
 librosa
 matplotlib>=3.3.0
-numpy
+numpy==1.19.3; platform_system == "Windows"
+numpy==1.20.3; platform_system != "Windows"
 scipy>=1.0.0
 tqdm
 sounddevice
@@ -12,8 +13,8 @@ inflect
 PyQt5
 multiprocess
 numba
-webrtcvad
-pypinyin
+webrtcvad; platform_system != "Windows"
+pypinyin==0.44.0
 flask
 flask_wtf
 flask_cors
@@ -25,9 +26,9 @@ PyYAML
 torch_complex
 espnet
 PyWavelets
-monotonic-align==0.0.3
-transformers
 fastapi
 loguru
-typer[all]
-click
+click==8.0.4
+typer
+monotonic-align==1.0.0
+transformers