optimize: some training optimizations (#95)

* optimzie(train&uvr5): rm sf & simp. AudioPre

* fix(audio): too many mallocs

* feat(audio): load_audio support stereo

* fix(audio): float32 wav saving

* fix(train): missing ckpt var

Author: fumiama

infer/lib/audio.py

@@ -1,11 +1,16 @@
 from io import BufferedWriter, BytesIO
 from pathlib import Path
-from typing import Dict, Tuple
+from typing import Dict, Tuple, Optional, Union, List
 import os
+import math
+import wave
 
 import numpy as np
+from numba import jit
 import av
 from av.audio.resampler import AudioResampler
+from av.audio.frame import AudioFrame
+import scipy.io.wavfile as wavfile
 
 video_format_dict: Dict[str, str] = {
     "m4a": "mp4",
@@ -17,6 +22,29 @@
 }
 
 
+@jit(nopython=True)
+def float_to_int16(audio: np.ndarray) -> np.ndarray:
+    am = int(math.ceil(float(np.abs(audio).max())) * 32768)
+    am = 32767 * 32768 // am
+    return np.multiply(audio, am).astype(np.int16)
+
+def float_np_array_to_wav_buf(wav: np.ndarray, sr: int, f32=False) -> BytesIO:
+    buf = BytesIO()
+    if f32:
+        wavfile.write(buf, sr, wav.astype(np.float32))
+    else:
+        with wave.open(buf, "wb") as wf:
+            wf.setnchannels(2 if len(wav.shape) > 1 else 1)
+            wf.setsampwidth(2)  # Sample width in bytes
+            wf.setframerate(sr)  # Sample rate in Hz
+            wf.writeframes(float_to_int16(wav.T if len(wav.shape) > 1 else wav))
+    buf.seek(0, 0)
+    return buf
+
+def save_audio(path: str, audio: np.ndarray, sr: int, f32=False):
+    with open(path, "wb") as f:
+        f.write(float_np_array_to_wav_buf(audio, sr, f32).getbuffer())
+
 def wav2(i: BytesIO, o: BufferedWriter, format: str):
     inp = av.open(i, "r")
     format = video_format_dict.get(format, format)
@@ -36,43 +64,72 @@ def wav2(i: BytesIO, o: BufferedWriter, format: str):
     inp.close()
 
 
-def load_audio(file: str, sr: int) -> np.ndarray:
-    if not Path(file).exists():
+def load_audio(
+        file: Union[str, BytesIO, Path],
+        sr: Optional[int]=None,
+        format: Optional[str]=None,
+        mono=True
+    ) -> Union[np.ndarray, Tuple[np.ndarray, int]]:
+    if (isinstance(file, str) and not Path(file).exists()) or (isinstance(file, Path) and not file.exists()):
         raise FileNotFoundError(f"File not found: {file}")
+    rate = 0
+
+    container = av.open(file, format=format)
+    audio_stream = next(s for s in container.streams if s.type == "audio")
+    channels = 1 if audio_stream.layout == "mono" else 2
+    container.seek(0)
+    resampler = AudioResampler(format="fltp", layout=audio_stream.layout, rate=sr) if sr is not None else None
+
+    # Estimated maximum total number of samples to pre-allocate the array
+    # AV stores length in microseconds by default
+    estimated_total_samples = int(container.duration * sr // 1_000_000) if sr is not None else 48000
+    decoded_audio = np.zeros(estimated_total_samples + 1 if channels == 1 else (channels, estimated_total_samples + 1), dtype=np.float32)
+
+    offset = 0
+
+    def process_packet(packet: List[AudioFrame]):
+        frames_data = []
+        rate = 0
+        for frame in packet:
+            frame.pts = None  # 清除时间戳，避免重新采样问题
+            resampled_frames = resampler.resample(frame) if resampler is not None else [frame]
+            for resampled_frame in resampled_frames:
+                frame_data = resampled_frame.to_ndarray()
+                rate = resampled_frame.rate
+                frames_data.append(frame_data)
+        return (rate, frames_data)
 
-    try:
-        container = av.open(file)
-        resampler = AudioResampler(format="fltp", layout="mono", rate=sr)
+    def frame_iter(container):
+        for p in container.demux(container.streams.audio[0]):
+            yield p.decode()
 
-        # Estimated maximum total number of samples to pre-allocate the array
-        # AV stores length in microseconds by default
-        estimated_total_samples = int(container.duration * sr // 1_000_000)
-        decoded_audio = np.zeros(estimated_total_samples + 1, dtype=np.float32)
+    for r, frames_data in map(process_packet, frame_iter(container)):
+        if not rate: rate = r
+        for frame_data in frames_data:
+            end_index = offset + len(frame_data[0])
 
-        offset = 0
-        for frame in container.decode(audio=0):
-            frame.pts = None  # Clear presentation timestamp to avoid resampling issues
-            resampled_frames = resampler.resample(frame)
-            for resampled_frame in resampled_frames:
-                frame_data = resampled_frame.to_ndarray()[0]
-                end_index = offset + len(frame_data)
+            # 检查 decoded_audio 是否有足够的空间，并在必要时调整大小
+            if end_index > decoded_audio.shape[1]:
+                decoded_audio = np.resize(decoded_audio, (decoded_audio.shape[0], end_index*4))
 
-                # Check if decoded_audio has enough space, and resize if necessary
-                if end_index > decoded_audio.shape[0]:
-                    decoded_audio = np.resize(decoded_audio, end_index + 1)
+            np.copyto(decoded_audio[..., offset:end_index], frame_data)
+            offset += len(frame_data[0])
 
-                decoded_audio[offset:end_index] = frame_data
-                offset += len(frame_data)
+    # Truncate the array to the actual size
+    decoded_audio = decoded_audio[..., :offset]
 
-        # Truncate the array to the actual size
-        decoded_audio = decoded_audio[:offset]
-    except Exception as e:
-        raise RuntimeError(f"Failed to load audio: {e}")
+    if mono and decoded_audio.shape[0] > 1:
+        decoded_audio = decoded_audio.mean(0)
 
-    return decoded_audio
+    if sr is not None:
+        return decoded_audio
+    return decoded_audio, rate
 
 
-def downsample_audio(input_path: str, output_path: str, format: str) -> None:
+def downsample_audio(input_path: str, output_path: str, format: str, br=128_000) -> None:
+    """
+    default to 128kb/s (equivalent to -q:a 2)
+    """
     if not os.path.exists(input_path):
         return
 
@@ -83,7 +140,7 @@ def downsample_audio(input_path: str, output_path: str, format: str) -> None:
     input_stream = input_container.streams.audio[0]
     output_stream = output_container.add_stream(format)
 
-    output_stream.bit_rate = 128_000  # 128kb/s (equivalent to -q:a 2)
+    output_stream.bit_rate = br
 
     # Copy packets from the input file to the output file
     for packet in input_container.demux(input_stream):
@@ -141,7 +198,7 @@ def resample_audio(
         print(f"Failed to remove the original file: {e}")
 
 
-def get_audio_properties(input_path: str) -> Tuple:
+def get_audio_properties(input_path: str) -> Tuple[int, int]:
     container = av.open(input_path)
     audio_stream = next(s for s in container.streams if s.type == "audio")
     channels = 1 if audio_stream.layout == "mono" else 2

infer/lib/slicer2.py

@@ -183,8 +183,7 @@ def main():
     import os.path
     from argparse import ArgumentParser
 
-    import librosa
-    import soundfile
+    from .audio import load_audio, save_audio
 
     parser = ArgumentParser()
     parser.add_argument("audio", type=str, help="The audio to be sliced")
@@ -230,7 +229,7 @@ def main():
     out = args.out
     if out is None:
         out = os.path.dirname(os.path.abspath(args.audio))
-    audio, sr = librosa.load(args.audio, sr=None, mono=False)
+    audio, sr = load_audio(args.audio, mono=False)
     slicer = Slicer(
         sr=sr,
         threshold=args.db_thresh,
@@ -245,15 +244,11 @@ def main():
     for i, chunk in enumerate(chunks):
         if len(chunk.shape) > 1:
             chunk = chunk.T
-        soundfile.write(
-            os.path.join(
-                out,
-                f"%s_%d.wav"
-                % (os.path.basename(args.audio).rsplit(".", maxsplit=1)[0], i),
-            ),
-            chunk,
-            sr,
-        )
+        save_audio(os.path.join(
+            out,
+            f"%s_%d.wav"
+            % (os.path.basename(args.audio).rsplit(".", maxsplit=1)[0], i),
+        ), chunk, sr)
 
 
 if __name__ == "__main__":

infer/lib/train/utils.py

@@ -16,62 +16,12 @@
 logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
 logger = logging
 
-"""
-def load_checkpoint_d(checkpoint_path, combd, sbd, optimizer=None, load_opt=1):
-    assert os.path.isfile(checkpoint_path)
-    checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
-
-    ##################
-    def go(model, bkey):
-        saved_state_dict = checkpoint_dict[bkey]
-        if hasattr(model, "module"):
-            state_dict = model.module.state_dict()
-        else:
-            state_dict = model.state_dict()
-        new_state_dict = {}
-        for k, v in state_dict.items():  # 模型需要的shape
-            try:
-                new_state_dict[k] = saved_state_dict[k]
-                if saved_state_dict[k].shape != state_dict[k].shape:
-                    logger.warning(
-                        "shape-%s-mismatch. need: %s, get: %s",
-                        k,
-                        state_dict[k].shape,
-                        saved_state_dict[k].shape,
-                    )  #
-                    raise KeyError
-            except:
-                # logger.info(traceback.format_exc())
-                logger.info("%s is not in the checkpoint", k)  # pretrain缺失的
-                new_state_dict[k] = v  # 模型自带的随机值
-        if hasattr(model, "module"):
-            model.module.load_state_dict(new_state_dict, strict=False)
-        else:
-            model.load_state_dict(new_state_dict, strict=False)
-        return model
-
-    go(combd, "combd")
-    model = go(sbd, "sbd")
-    #############
-    logger.info("Loaded model weights")
-
-    iteration = checkpoint_dict["iteration"]
-    learning_rate = checkpoint_dict["learning_rate"]
-    if (
-        optimizer is not None and load_opt == 1
-    ):  ###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
-        #   try:
-        optimizer.load_state_dict(checkpoint_dict["optimizer"])
-    #   except:
-    #     traceback.print_exc()
-    logger.info("Loaded checkpoint '{}' (epoch {})".format(checkpoint_path, iteration))
-    return model, optimizer, learning_rate, iteration
-"""
-
 
 def load_checkpoint(checkpoint_path, model, optimizer=None, load_opt=1):
     assert os.path.isfile(checkpoint_path)
-    saved_state_dict = torch.load(checkpoint_path, map_location="cpu", weights_only=True)["model"]
+    checkpoint_dict = torch.load(checkpoint_path, map_location="cpu", weights_only=True)
+
+    saved_state_dict = checkpoint_dict["model"]
     if hasattr(model, "module"):
         state_dict = model.module.state_dict()
     else:
@@ -132,34 +82,6 @@ def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path)
     )
 
 
-"""
-def save_checkpoint_d(combd, sbd, optimizer, learning_rate, iteration, checkpoint_path):
-    logger.info(
-        "Saving model and optimizer state at epoch {} to {}".format(
-            iteration, checkpoint_path
-        )
-    )
-    if hasattr(combd, "module"):
-        state_dict_combd = combd.module.state_dict()
-    else:
-        state_dict_combd = combd.state_dict()
-    if hasattr(sbd, "module"):
-        state_dict_sbd = sbd.module.state_dict()
-    else:
-        state_dict_sbd = sbd.state_dict()
-    torch.save(
-        {
-            "combd": state_dict_combd,
-            "sbd": state_dict_sbd,
-            "iteration": iteration,
-            "optimizer": optimizer.state_dict(),
-            "learning_rate": learning_rate,
-        },
-        checkpoint_path,
-    )
-"""
-
-
 def summarize(
     writer,
     global_step,
@@ -366,53 +288,6 @@ def get_hparams(init=True):
     return hparams
 
 
-"""
-def get_hparams_from_dir(model_dir):
-    config_save_path = os.path.join(model_dir, "config.json")
-    with open(config_save_path, "r") as f:
-        data = f.read()
-    config = json.loads(data)
-
-    hparams = HParams(**config)
-    hparams.model_dir = model_dir
-    return hparams
-
-
-def get_hparams_from_file(config_path):
-    with open(config_path, "r") as f:
-        data = f.read()
-    config = json.loads(data)
-
-    hparams = HParams(**config)
-    return hparams
-
-
-def check_git_hash(model_dir):
-    source_dir = os.path.dirname(os.path.realpath(__file__))
-    if not os.path.exists(os.path.join(source_dir, ".git")):
-        logger.warning(
-            "{} is not a git repository, therefore hash value comparison will be ignored.".format(
-                source_dir
-            )
-        )
-        return
-
-    cur_hash = subprocess.getoutput("git rev-parse HEAD")
-
-    path = os.path.join(model_dir, "githash")
-    if os.path.exists(path):
-        saved_hash = open(path).read()
-        if saved_hash != cur_hash:
-            logger.warning(
-                "git hash values are different. {}(saved) != {}(current)".format(
-                    saved_hash[:8], cur_hash[:8]
-                )
-            )
-    else:
-        open(path, "w").write(cur_hash)
-"""
-
-
 def get_logger(model_dir, filename="train.log"):
     global logger
     logger = logging.getLogger(os.path.basename(model_dir))

infer/modules/train/extract_feature_print.py

@@ -2,6 +2,11 @@
 import sys
 import traceback
 
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from infer.lib.audio import load_audio
+
 os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
 os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"
 
@@ -20,7 +25,6 @@
     is_half = sys.argv[7].lower() == "true"
 import fairseq
 import numpy as np
-import soundfile as sf
 import torch
 import torch.nn.functional as F
 
@@ -64,11 +68,9 @@ def printt(strr):
 
 # wave must be 16k, hop_size=320
 def readwave(wav_path, normalize=False):
-    wav, sr = sf.read(wav_path)
+    wav, sr = load_audio(wav_path)
     assert sr == 16000
     feats = torch.from_numpy(wav).float()
-    if feats.dim() == 2:  # double channels
-        feats = feats.mean(-1)
     assert feats.dim() == 1, feats.dim()
     if normalize:
         with torch.no_grad():