Upload 3 files

README.md

@@ -1,12 +1,12 @@
 ---
 title: YouTube To MT3
 emoji: 🎼
-colorFrom: green
-colorTo: gray
+colorFrom: purple
+colorTo: green
 sdk: gradio
 sdk_version: 3.4.1
 app_file: app.py
-pinned: false
+pinned: true
 duplicated_from: mdnestor/YouTube-to-MT3
 ---
 

app.py

@@ -1,18 +1,31 @@
 import os
+os.system("pip install gradio")
+
 import gradio as gr
+from pathlib import Path
+os.system("pip install gsutil")
 import glob
 
-os.system("apt-get update -qq && apt-get install -qq libfluidsynth2 build-essential libasound2-dev libjack-dev")
 
-# install mt3
+os.system("git clone --branch=main https://github.com/google-research/t5x")
+os.system("mv t5x t5x_tmp; mv t5x_tmp/* .; rm -r t5x_tmp")
+os.system("sed -i 's:jax\[tpu\]:jax:' setup.py")
+os.system("python3 -m pip install -e .")
+os.system("python3 -m pip install --upgrade pip")
+
+
+# 安装 mt3
 os.system("git clone --branch=main https://github.com/magenta/mt3")
 os.system("mv mt3 mt3_tmp; mv mt3_tmp/* .; rm -r mt3_tmp")
-os.system("python3 -m pip install nest-asyncio pyfluidsynth==1.3.0 -e .")
-
+os.system("python3 -m pip install -e .")
+os.system("pip install tensorflow_cpu")
+# 复制检查点
 os.system("gsutil -q -m cp -r gs://mt3/checkpoints .")
-os.system("gsutil -q -m cp gs://magentadata/soundfonts/SGM-v2.01-Sal-Guit-Bass-V1.3.sf2 .")
 
+# 复制 soundfont 文件（原始文件来自 https://sites.google.com/site/soundfonts4u）
+os.system("gsutil -q -m cp gs://magentadata/soundfonts/SGM-v2.01-Sal-Guit-Bass-V1.3.sf2 .")
 
+#@title 导入和定义
 import functools
 import os
 
@@ -24,6 +37,7 @@ import gin
 import jax
 import librosa
 import note_seq
+
 import seqio
 import t5
 import t5x
@@ -42,12 +56,16 @@ nest_asyncio.apply()
 SAMPLE_RATE = 16000
 SF2_PATH = 'SGM-v2.01-Sal-Guit-Bass-V1.3.sf2'
 
+def callbak_audio(audio, sample_rate):
+  return note_seq.audio_io.wav_data_to_samples_librosa(
+    audio, sample_rate=sample_rate)
+
 class InferenceModel(object):
-  """Wrapper of T5X model for music transcription."""
+  """音乐转录的 T5X 模型包装器。"""
 
   def __init__(self, checkpoint_path, model_type='mt3'):
 
-    # Model Constants.
+    # 模型常量。
     if model_type == 'ismir2021':
       num_velocity_bins = 127
       self.encoding_spec = note_sequences.NoteEncodingSpec
@@ -68,9 +86,9 @@ class InferenceModel(object):
                             'targets': self.outputs_length}
 
     self.partitioner = t5x.partitioning.PjitPartitioner(
-        model_parallel_submesh=(1, 1, 1, 1), num_partitions=1)
+        model_parallel_submesh=(1, 1, 1, 1))
 
-    # Build Codecs and Vocabularies.
+    # 构建编解码器和词汇表。
     self.spectrogram_config = spectrograms.SpectrogramConfig()
     self.codec = vocabularies.build_codec(
         vocab_config=vocabularies.VocabularyConfig(
@@ -81,11 +99,11 @@ class InferenceModel(object):
         'targets': seqio.Feature(vocabulary=self.vocabulary),
     }
 
-    # Create a T5X model.
+    # 创建 T5X 模型。
     self._parse_gin(gin_files)
     self.model = self._load_model()
 
-    # Restore from checkpoint.
+    # 从检查点中恢复。
     self.restore_from_checkpoint(checkpoint_path)
 
   @property
@@ -96,7 +114,7 @@ class InferenceModel(object):
     }
 
   def _parse_gin(self, gin_files):
-    """Parse gin files used to train the model."""
+    """解析用于训练模型的 gin 文件。"""
     gin_bindings = [
         'from __gin__ import dynamic_registration',
         'from mt3 import vocabularies',
@@ -108,7 +126,7 @@ class InferenceModel(object):
           gin_files, gin_bindings, finalize_config=False)
 
   def _load_model(self):
-    """Load up a T5X `Model` after parsing training gin config."""
+    """在解析训练 gin 配置后加载 T5X `Model`。"""
     model_config = gin.get_configurable(network.T5Config)()
     module = network.Transformer(config=model_config)
     return models.ContinuousInputsEncoderDecoderModel(
@@ -120,7 +138,7 @@ class InferenceModel(object):
 
 
   def restore_from_checkpoint(self, checkpoint_path):
-    """Restore training state from checkpoint, resets self._predict_fn()."""
+    """从检查点中恢复训练状态，重置 self._predict_fn()。"""
     train_state_initializer = t5x.utils.TrainStateInitializer(
       optimizer_def=self.model.optimizer_def,
       init_fn=self.model.get_initial_variables,
@@ -137,7 +155,7 @@ class InferenceModel(object):
 
   @functools.lru_cache()
   def _get_predict_fn(self, train_state_axes):
-    """Generate a partitioned prediction function for decoding."""
+    """生成一个分区的预测函数用于解码。"""
     def partial_predict_fn(params, batch, decode_rng):
       return self.model.predict_batch_with_aux(
           params, batch, decoder_params={'decode_rng': None})
@@ -150,18 +168,18 @@ class InferenceModel(object):
     )
 
   def predict_tokens(self, batch, seed=0):
-    """Predict tokens from preprocessed dataset batch."""
+    """从预处理的数据集批次中预测 tokens。"""
     prediction, _ = self._predict_fn(
         self._train_state.params, batch, jax.random.PRNGKey(seed))
     return self.vocabulary.decode_tf(prediction).numpy()
 
   def __call__(self, audio):
-    """Infer note sequence from audio samples.
+    """从音频样本推断出音符序列。
     
-    Args:
-      audio: 1-d numpy array of audio samples (16kHz) for a single example.
-    Returns:
-      A note_sequence of the transcribed audio.
+    参数：
+      audio：16kHz 的单个音频样本的 1 维 numpy 数组。
+    返回：
+      转录音频的音符序列。
     """
     ds = self.audio_to_dataset(audio)
     ds = self.preprocess(ds)
@@ -182,7 +200,7 @@ class InferenceModel(object):
     return result['est_ns']
 
   def audio_to_dataset(self, audio):
-    """Create a TF Dataset of spectrograms from input audio."""
+    """从输入音频创建一个包含频谱图的 TF Dataset。"""
     frames, frame_times = self._audio_to_frames(audio)
     return tf.data.Dataset.from_tensors({
         'inputs': frames,
@@ -190,7 +208,7 @@ class InferenceModel(object):
     })
 
   def _audio_to_frames(self, audio):
-    """Compute spectrogram frames from audio."""
+    """从音频计算频谱图帧。"""
     frame_size = self.spectrogram_config.hop_width
     padding = [0, frame_size - len(audio) % frame_size]
     audio = np.pad(audio, padding, mode='constant')
@@ -207,7 +225,7 @@ class InferenceModel(object):
             output_features=self.output_features,
             feature_key='inputs',
             additional_feature_keys=['input_times']),
-        # Cache occurs here during training.
+        # 在训练期间进行缓存。
         preprocessors.add_dummy_targets,
         functools.partial(
             preprocessors.compute_spectrograms,
@@ -220,12 +238,12 @@ class InferenceModel(object):
   def postprocess(self, tokens, example):
     tokens = self._trim_eos(tokens)
     start_time = example['input_times'][0]
-    # Round down to nearest symbolic token step.
+    # 向下取整到最接近的符号化时间步。
     start_time -= start_time % (1 / self.codec.steps_per_second)
     return {
         'est_tokens': tokens,
         'start_time': start_time,
-        # Internal MT3 code expects raw inputs, not used here.
+        # 内部 MT3 代码期望原始输入，这里不使用。
         'raw_inputs': []
     }
 
@@ -241,24 +259,23 @@ inference_model = InferenceModel('/home/user/app/checkpoints/mt3/', 'mt3')
 
 def inference(url):
   os.system(f"yt-dlp -x {url} -o 'audio.%(ext)s'")
-  audio_file = glob.glob('audio.*')[0]
-  with open(audio_file, 'rb') as f:
-    data = f.read()
-  audio = note_seq.audio_io.wav_data_to_samples_librosa(data, sample_rate=SAMPLE_RATE)
+  audio = glob.glob('audio.*')[0]
+  with open(audio, 'rb') as fd:
+      contents = fd.read()
+  audio = callbak_audio(contents,sample_rate=16000)
   est_ns = inference_model(audio)
-  midi_file = f"./transcribed.mid"
-  note_seq.sequence_proto_to_midi_file(est_ns, midi_file)
-  return midi_file
+  note_seq.sequence_proto_to_midi_file(est_ns, './transcribed.mid')
+  return './transcribed.mid'
   
 title = "YouTube-to-MT3"
-description = "Upload YouTube audio to MT3: Multi-Task Multitrack Music Transcription. Thanks to <a href=\"https://huggingface.co/spaces/akhaliq/MT3\">akhaliq</a> for the original <i>Spaces</i> implementation."
+description = "将YouTube音频上传到MT3：多任务多轨音乐转录。感谢 <a href=\"https://huggingface.co/spaces/akhaliq/MT3\">akhaliq</a> 的原始 <i>Spaces</i> 实现。"
 
-article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2111.03017' target='_blank'>MT3: Multi-Task Multitrack Music Transcription</a> | <a href='https://github.com/magenta/mt3' target='_blank'>Github Repo</a></p>"
+article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2111.03017' target='_blank'>MT3: 多任务多轨音乐转录</a> | <a href='https://github.com/magenta/mt3' target='_blank'>Github 仓库</a></p>"
 
 gr.Interface(
     inference, 
-    gr.Textbox(label="Audio URL"),
-    gr.outputs.File(label="Transcribed MIDI"),
+    gr.inputs.Textbox(label="URL"),
+    gr.outputs.File(label="输出"),
     title=title,
     description=description,
     article=article,

requirements.txt

@@ -1,6 +1,9 @@
-absl-py
+nest-asyncio
+pyfluidsynth
+absl-py 
 ddsp
-flax@git+https://github.com/google/flax#egg=flax
+flax
+glob
 gin-config
 immutabledict
 librosa
@@ -10,9 +13,8 @@ numpy
 pretty_midi
 scikit-learn
 scipy
-seqio @ git+https://github.com/google/seqio#egg=seqio
+seqio
 t5
-t5x@git+https://github.com/google-research/t5x#egg=t5x
-tensorflow
+tensorflow_cpu
 tensorflow-datasets
 yt-dlp