Upload 2 files

.gitattributes

@@ -29,3 +29,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zstandard filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+canon.flac filter=lfs diff=lfs merge=lfs -text

app.py

@@ -13,22 +13,18 @@ os.system("python3 -m pip install -e .")
 os.system("python3 -m pip install --upgrade pip")
 
 
-
-# install mt3
+# 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 -e .")
 os.system("pip install tensorflow_cpu")
-# copy checkpoints
+# 复制检查点
 os.system("gsutil -q -m cp -r gs://mt3/checkpoints .")
 
-# copy soundfont (originally from https://sites.google.com/site/soundfonts4u)
+# 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 Imports and Definitions
-
-
-
+#@title 导入和定义
 
 
 import functools
@@ -45,7 +41,6 @@ import librosa
 import note_seq
 
 
-
 import seqio
 import t5
 import t5x
@@ -72,11 +67,11 @@ def upload_audio(audio, 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
@@ -99,7 +94,7 @@ class InferenceModel(object):
     self.partitioner = t5x.partitioning.PjitPartitioner(
         model_parallel_submesh=None, num_partitions=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(
@@ -110,11 +105,11 @@ 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
@@ -125,7 +120,7 @@ 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',
@@ -137,7 +132,7 @@ 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(
@@ -149,7 +144,7 @@ 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,
@@ -166,7 +161,7 @@ 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})
@@ -179,18 +174,18 @@ 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))
+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)
@@ -211,7 +206,7 @@ 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,
@@ -219,7 +214,7 @@ 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')
@@ -236,7 +231,7 @@ 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,
@@ -249,12 +244,12 @@ 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': []
     }
 
@@ -285,16 +280,16 @@ def inference(audio):
   return './transcribed.mid'
   
 title = "MT3"
-description = "Gradio demo for MT3: Multi-Task Multitrack Music Transcription. To use it, simply upload your audio file, or click one of the examples to load them. Read more at the links below."
+description = "MT3：多任务多音轨音乐转录的 Gradio 演示。要使用它，只需上传音频文件，或点击示例以加载它们。更多信息请参阅下面的链接。"
 
-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: 多任务多音轨音乐转录</a> | <a href='https://github.com/magenta/mt3' target='_blank'>Github 仓库</a></p>"
 
-examples=[['download.wav']]
+examples=[['canon.flac'], ['download.wav']]
 
 gr.Interface(
     inference, 
-    gr.inputs.Audio(type="filepath", label="Input"), 
-    [gr.outputs.File(label="Output")],
+    gr.inputs.Audio(type="filepath", label="输入"), 
+    [gr.outputs.File(label="输出")],
     title=title,
     description=description,
     article=article,

canon.flac

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d13c270188979b6840a736cbd85f5e1bdb12b1bdab3e35af8a4ae7eb2c1c80ac
+size 6229211