Open source URMP dataset pipeline

PiperOrigin-RevId: 460754854

ddsp/training/data_preparation/prepare_urmp_dataset.py

@@ -0,0 +1,81 @@
+# Copyright 2022 The DDSP Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+r"""Prepare URMP dataset DDSP and NoteSequence features.
+
+Usage:
+====================
+ddsp_prepare_urmp_dataset \
+  --input_filepath='/path/to/input.tfrecord-*' \
+  --output_filepath='/path/to/output.tfrecord' \
+  --instrument_key=vn \
+  --num_shards=10 \
+  --alsologtostderr
+
+"""
+
+from absl import app
+from absl import flags
+
+from ddsp.training.data_preparation.prepare_urmp_dataset_lib import prepare_urmp
+import tensorflow.compat.v2 as tf
+
+FLAGS = flags.FLAGS
+
+flags.DEFINE_string('input_filepath', '', 'Input filepath for dataset.')
+flags.DEFINE_string('output_filepath', '', 'Output filepath for dataset.')
+flags.DEFINE_multi_string(
+    'instrument_key', [], 'Instrument keys to extract. '
+    'If not set, extract all instruments. Possible keys '
+    'are vn, va, vc, db, fl, ob, cl, sax, bn, tpt, hn, '
+    'tbn, tba.')
+flags.DEFINE_integer(
+    'num_shards', None, 'Num shards for output dataset. If '
+    'None, this number will be determined automatically.')
+flags.DEFINE_bool('batch', True, 'Whether or not to batch the dataset.')
+flags.DEFINE_bool('force_monophonic', True, 'Fix URMP note labels such that '
+                  'note onsets and offsets do not overlap.')
+flags.DEFINE_list(
+    'pipeline_options', '--runner=DirectRunner',
+    'A comma-separated list of command line arguments to be used as options '
+    'for the Beam Pipeline.')
+flags.DEFINE_integer('ddsp_sample_rate', 250, 'Sample rate for dataset output.')
+flags.DEFINE_integer('audio_sample_rate', 16000, 'Sample rate for URMP audio.')
+
+
+def run():
+  prepare_urmp(
+      input_filepath=FLAGS.input_filepath,
+      output_filepath=FLAGS.output_filepath,
+      instrument_keys=FLAGS.instrument_key,
+      num_shards=FLAGS.num_shards,
+      batch=FLAGS.batch,
+      force_monophonic=FLAGS.force_monophonic,
+      pipeline_options=FLAGS.pipeline_options,
+      ddsp_sample_rate=FLAGS.ddsp_sample_rate,
+      audio_sample_rate=FLAGS.audio_sample_rate)
+
+
+def main(unused_argv):
+  """From command line."""
+  run()
+
+
+def console_entry_point():
+  """From pip installed script."""
+  app.run(main)
+
+
+if __name__ == '__main__':
+  console_entry_point()

ddsp/training/data_preparation/prepare_urmp_dataset_lib.py

@@ -0,0 +1,265 @@
+# Copyright 2022 The DDSP Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""URMP data import pipeline."""
+import apache_beam as beam
+import ddsp
+from ddsp.training import heuristics
+from mir_eval import melody
+from note_seq import audio_io
+from note_seq import constants
+from note_seq import sequences_lib
+from note_seq.protobuf import music_pb2
+import numpy as np
+import tensorflow as tf
+
+
+DDSP_SAMPLE_RATE = 250
+AUDIO_SAMPLE_RATE = 16000
+
+
+def parse_example(tfexample):
+  """Parse tf.Example protos to dict of numpy arrays."""
+  features = {
+      'id':
+          tf.io.FixedLenFeature([], dtype=tf.string),
+      'audio':
+          tf.io.FixedLenFeature([], dtype=tf.string),
+      'f0_hz':
+          tf.io.FixedLenSequenceFeature([],
+                                        dtype=tf.float32,
+                                        allow_missing=True),
+      'f0_time':
+          tf.io.FixedLenSequenceFeature([],
+                                        dtype=tf.float32,
+                                        allow_missing=True),
+      'sequence':
+          tf.io.FixedLenFeature([], dtype=tf.string)
+  }
+  ex = {
+      key: val.numpy()
+      for key, val in tf.io.parse_single_example(tfexample, features).items()
+  }
+  return ex
+
+
+def get_active_frame_indices(piano_roll):
+  """Create matrix of frame indices for active notes relative to onset."""
+  active_frame_indices = np.zeros_like(piano_roll.active_velocities)
+  for frame_i in range(1, active_frame_indices.shape[0]):
+    prev_indices = active_frame_indices[frame_i - 1, :]
+    active_notes = piano_roll.active[frame_i, :]
+    active_frame_indices[frame_i, :] = (prev_indices + 1) * active_notes
+  return active_frame_indices
+
+
+def attach_metadata(ex, ddsp_sample_rate, audio_sample_rate, force_monophonic):
+  """Parse and attach metadata from the dataset."""
+
+  def extract_recording_id(id_string):
+    id_string = id_string.split(b'/')[-1]
+    id_string = id_string.split(b'.')[0]
+    return id_string
+
+  def extract_instrument_id(id_string):
+    id_string = extract_recording_id(id_string).split(b'_')
+    return id_string[2]
+
+  def extract_notes(sequence_str, expected_seconds):
+    ns = music_pb2.NoteSequence.FromString(sequence_str)
+    # total time in dataset doesn't include silence at the end
+    if force_monophonic:
+      for i in range(1, len(ns.notes)):
+        note = ns.notes[i]
+        prev_note = ns.notes[i - 1]
+        onset_frame = int(note.start_time * ddsp_sample_rate)
+        prev_note_offset_frame = int(prev_note.end_time * ddsp_sample_rate)
+        if prev_note_offset_frame >= onset_frame:
+          frames_to_move = (prev_note_offset_frame - onset_frame) + 1
+          # move previous note's onset back by frames_to_move frames in seconds
+          prev_note.end_time -= float(frames_to_move) / ddsp_sample_rate
+
+    ns.total_time = expected_seconds
+    piano_roll = sequences_lib.sequence_to_pianoroll(
+        ns,
+        frames_per_second=ddsp_sample_rate,
+        min_pitch=constants.MIN_MIDI_PITCH,
+        max_pitch=constants.MAX_MIDI_PITCH,
+        onset_mode='length_ms')
+
+    note_dict = {
+        'note_active_velocities': piano_roll.active_velocities,
+        'note_active_frame_indices': get_active_frame_indices(piano_roll),
+        'note_onsets': piano_roll.onsets,
+        'note_offsets': piano_roll.offsets
+    }
+
+    return note_dict
+
+  ex['recording_id'] = extract_recording_id(ex['id'])
+  ex['instrument_id'] = extract_instrument_id(ex['id'])
+  ex['audio'] = audio_io.wav_data_to_samples_librosa(
+      ex['audio'], sample_rate=audio_sample_rate)
+  expected_seconds = ex['audio'].shape[0] / audio_sample_rate
+  ex.update(extract_notes(ex['sequence'], expected_seconds))
+  beam.metrics.Metrics.distribution('prepare-urmp',
+                                    'orig-audio-len').update(len(ex['audio']))
+  return ex
+
+
+def normalize_audio(ex, max_audio):
+  ex['audio'] /= max_audio
+  return ex
+
+
+def resample(ex, ddsp_sample_rate, audio_sample_rate):
+  """Resample features to standard DDSP sample rate."""
+  f0_times = ex['f0_time']
+  f0_orig = ex['f0_hz']
+  max_time = np.max(f0_times)
+  new_times = np.linspace(0, max_time, int(ddsp_sample_rate * max_time))
+  if f0_times[0] > 0:
+    f0_orig = np.insert(f0_orig, 0, f0_orig[0])
+    f0_times = np.insert(f0_times, 0, 0)
+  f0_interpolated, _ = melody.resample_melody_series(
+      f0_times, f0_orig,
+      melody.freq_to_voicing(f0_orig)[1], new_times)
+  ex['f0_hz'] = f0_interpolated
+  ex['f0_time'] = new_times
+  ex['orig_f0_hz'] = f0_orig
+  ex['orig_f0_time'] = f0_times
+
+  # Truncate audio to an integer multiple of f0_hz vector.
+  num_audio_samples = round(
+      len(ex['f0_hz']) * (audio_sample_rate / ddsp_sample_rate))
+  beam.metrics.Metrics.distribution(
+      'prepare-urmp',
+      'resampled-audio-diff').update(num_audio_samples - len(ex['audio']))
+
+  ex['audio'] = ex['audio'][:num_audio_samples]
+
+  # Truncate pianoroll features to length of f0_hz vector.
+  for key in [
+      'note_active_frame_indices', 'note_active_velocities', 'note_onsets',
+      'note_offsets'
+  ]:
+    ex[key] = ex[key][:len(ex['f0_hz']), :]
+
+  return ex
+
+
+def batch_dataset(ex, audio_sample_rate, ddsp_sample_rate):
+  """Split features and audio into 4 second sliding windows."""
+  batched = []
+  for key, vec in ex.items():
+    if isinstance(vec, np.ndarray):
+      if key == 'audio':
+        sampling_rate = audio_sample_rate
+      else:
+        sampling_rate = ddsp_sample_rate
+
+      frames = heuristics.window_array(vec, sampling_rate, 4.0, 0.25)
+      if not batched:
+        batched = [{} for _ in range(len(frames))]
+      for i, frame in enumerate(frames):
+        batched[i][key] = frame
+
+  # once batches are created, replicate ids and metadata over all elements.
+  for key, val in ex.items():
+    if not isinstance(val, np.ndarray):
+      for batch in batched:
+        batch[key] = val
+
+  beam.metrics.Metrics.counter('prepare-urmp',
+                               'batches-created').inc(len(batched))
+  return batched
+
+
+def attach_ddsp_features(ex):
+  ex['loudness_db'] = ddsp.spectral_ops.compute_loudness(ex['audio'])
+  ex['power_db'] = ddsp.spectral_ops.compute_power(ex['audio'], frame_size=256)
+  # ground truth annotations are set with confidence 1.0
+  ex['f0_confidence'] = np.ones_like(ex['f0_hz'])
+  beam.metrics.Metrics.counter('prepare-urmp', 'ddsp-features-attached').inc()
+  return ex
+
+
+def serialize_tfexample(ex):
+  """Creates a tf.Example message ready to be written to a file."""
+
+  def _feature(arr):
+    """Returns a feature from a numpy array or string."""
+    if isinstance(arr, (bytes, str)):
+      return tf.train.Feature(bytes_list=tf.train.BytesList(value=[arr]))
+    else:
+      arr = np.asarray(arr).reshape(-1)
+      return tf.train.Feature(float_list=tf.train.FloatList(value=arr))
+
+  # Create a dictionary mapping the feature name to the tf.Example-compatible
+  # data type.
+  feature = {k: _feature(v) for k, v in ex.items()}
+
+  # Create a Features message using tf.train.Example.
+  example_proto = tf.train.Example(features=tf.train.Features(feature=feature))
+  return example_proto
+
+
+def prepare_urmp(input_filepath,
+                 output_filepath,
+                 instrument_keys,
+                 num_shards,
+                 batch,
+                 force_monophonic,
+                 pipeline_options,
+                 ddsp_sample_rate=DDSP_SAMPLE_RATE,
+                 audio_sample_rate=AUDIO_SAMPLE_RATE):
+  """Pipeline for parsing URMP dataset to a usable format for DDSP."""
+  pipeline_options = beam.options.pipeline_options.PipelineOptions(
+      pipeline_options)
+  with beam.Pipeline(options=pipeline_options) as pipeline:
+    examples = (
+        pipeline
+        |
+        'read_tfrecords' >> beam.io.tfrecordio.ReadFromTFRecord(input_filepath)
+        | 'parse_example' >> beam.Map(parse_example)
+        | 'attach_metadata' >> beam.Map(
+            attach_metadata,
+            ddsp_sample_rate=ddsp_sample_rate,
+            audio_sample_rate=audio_sample_rate,
+            force_monophonic=force_monophonic))
+
+    if instrument_keys:
+      examples |= 'filter_instruments' >> beam.Filter(
+          lambda ex: ex['instrument_id'].decode() in instrument_keys)
+
+    examples |= 'resample' >> beam.Map(
+        resample,
+        ddsp_sample_rate=ddsp_sample_rate,
+        audio_sample_rate=audio_sample_rate)
+    if batch:
+      examples |= 'batch' >> beam.FlatMap(
+          batch_dataset,
+          audio_sample_rate=audio_sample_rate,
+          ddsp_sample_rate=ddsp_sample_rate)
+    _ = (
+        examples
+        | 'attach_ddsp_features' >> beam.Map(attach_ddsp_features)
+        | 'filter_silence' >>
+        beam.Filter(lambda ex: np.any(ex['loudness_db'] > -70))
+        | 'serialize_tfexamples' >> beam.Map(serialize_tfexample)
+        | 'shuffle' >> beam.Reshuffle()
+        | beam.io.tfrecordio.WriteToTFRecord(
+            output_filepath,
+            num_shards=num_shards,
+            coder=beam.coders.ProtoCoder(tf.train.Example)))