File size: 14,610 Bytes
b100e1c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
# Copyright 2022 The MT3 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.

"""Transcription task definitions."""

import functools
from typing import Optional, Sequence

from mt3 import datasets
from mt3 import event_codec
from mt3 import metrics
from mt3 import mixing
from mt3 import preprocessors
from mt3 import run_length_encoding
from mt3 import spectrograms
from mt3 import vocabularies

import note_seq
import numpy as np
import seqio
import t5
import tensorflow as tf

# Split audio frame sequences into this length before the cache placeholder.
MAX_NUM_CACHED_FRAMES = 2000

seqio.add_global_cache_dirs(['gs://mt3/data/cache_tasks/'])


def construct_task_name(
    task_prefix: str,
    spectrogram_config=spectrograms.SpectrogramConfig(),
    vocab_config=vocabularies.VocabularyConfig(),
    task_suffix: Optional[str] = None
) -> str:
  """Construct task name from prefix, config, and optional suffix."""
  fields = [task_prefix]
  if spectrogram_config.abbrev_str:
    fields.append(spectrogram_config.abbrev_str)
  if vocab_config.abbrev_str:
    fields.append(vocab_config.abbrev_str)
  if task_suffix:
    fields.append(task_suffix)
  return '_'.join(fields)


def trim_eos(tokens: Sequence[int]) -> np.ndarray:
  """If EOS is present, remove it and everything after."""
  tokens = np.array(tokens, np.int32)
  if vocabularies.DECODED_EOS_ID in tokens:
    tokens = tokens[:np.argmax(tokens == vocabularies.DECODED_EOS_ID)]
  return tokens


def postprocess(tokens, example, is_target, codec):
  """Transcription postprocessing function."""
  tokens = trim_eos(tokens)

  if is_target:
    return {
        'unique_id': example['unique_id'][0],
        'ref_ns': (note_seq.NoteSequence.FromString(example['sequence'][0])
                   if example['sequence'][0] else None),
        'ref_tokens': tokens,
    }

  start_time = example['input_times'][0]
  # Round down to nearest symbolic token step.
  start_time -= start_time % (1 / codec.steps_per_second)

  return {
      'unique_id': example['unique_id'][0],
      'raw_inputs': example['raw_inputs'],
      'est_tokens': tokens,
      'start_time': start_time
  }


def add_transcription_task_to_registry(
    dataset_config: datasets.DatasetConfig,
    spectrogram_config: spectrograms.SpectrogramConfig,
    vocab_config: vocabularies.VocabularyConfig,
    tokenize_fn,  # TODO(iansimon): add type signature
    onsets_only: bool,
    include_ties: bool,
    skip_too_long: bool = False
) -> None:
  """Add note transcription task to seqio.TaskRegistry."""
  codec = vocabularies.build_codec(vocab_config)
  vocabulary = vocabularies.vocabulary_from_codec(codec)

  output_features = {
      'targets': seqio.Feature(vocabulary=vocabulary),
      'inputs': seqio.ContinuousFeature(dtype=tf.float32, rank=2)
  }

  task_name = 'onsets' if onsets_only else 'notes'
  if include_ties:
    task_name += '_ties'
  task_prefix = f'{dataset_config.name}_{task_name}'

  train_task_name = construct_task_name(
      task_prefix=task_prefix,
      spectrogram_config=spectrogram_config,
      vocab_config=vocab_config,
      task_suffix='train')

  mixture_task_names = []

  tie_token = codec.encode_event(event_codec.Event('tie', 0))
  track_specs = (dataset_config.track_specs
                 if dataset_config.track_specs else None)

  # Add transcription training task.
  seqio.TaskRegistry.add(
      train_task_name,
      source=seqio.TFExampleDataSource(
          split_to_filepattern={
              'train': dataset_config.paths[dataset_config.train_split],
              'eval': dataset_config.paths[dataset_config.train_eval_split]
          },
          feature_description=dataset_config.features),
      output_features=output_features,
      preprocessors=[
          functools.partial(
              tokenize_fn,
              spectrogram_config=spectrogram_config, codec=codec,
              is_training_data=True, onsets_only=onsets_only,
              include_ties=include_ties),
          functools.partial(
              t5.data.preprocessors.split_tokens,
              max_tokens_per_segment=MAX_NUM_CACHED_FRAMES,
              feature_key='inputs',
              additional_feature_keys=[
                  'input_event_start_indices', 'input_event_end_indices',
                  'input_state_event_indices'
              ],
              passthrough_feature_keys=['targets', 'state_events']),
          seqio.CacheDatasetPlaceholder(),
          functools.partial(
              t5.data.preprocessors.select_random_chunk,
              feature_key='inputs',
              additional_feature_keys=[
                  'input_event_start_indices', 'input_event_end_indices',
                  'input_state_event_indices'
              ],
              passthrough_feature_keys=['targets', 'state_events'],
              uniform_random_start=True),
          functools.partial(
              run_length_encoding.extract_target_sequence_with_indices,
              state_events_end_token=tie_token if include_ties else None),
          functools.partial(preprocessors.map_midi_programs, codec=codec),
          run_length_encoding.run_length_encode_shifts_fn(
              codec,
              feature_key='targets'),
          functools.partial(
              mixing.mix_transcription_examples,
              codec=codec,
              targets_feature_keys=['targets']),
          run_length_encoding.remove_redundant_state_changes_fn(
              feature_key='targets', codec=codec,
              state_change_event_types=['velocity', 'program']),
          functools.partial(
              preprocessors.compute_spectrograms,
              spectrogram_config=spectrogram_config),
          functools.partial(preprocessors.handle_too_long, skip=skip_too_long),
          functools.partial(
              seqio.preprocessors.tokenize_and_append_eos,
              copy_pretokenized=False)
      ],
      postprocess_fn=None,
      metric_fns=[],
  )

  # Add transcription eval tasks.
  for split in dataset_config.infer_eval_splits:
    eval_task_name = construct_task_name(
        task_prefix=task_prefix,
        spectrogram_config=spectrogram_config,
        vocab_config=vocab_config,
        task_suffix=split.suffix)

    if split.include_in_mixture:
      mixture_task_names.append(eval_task_name)

    seqio.TaskRegistry.add(
        eval_task_name,
        source=seqio.TFExampleDataSource(
            split_to_filepattern={'eval': dataset_config.paths[split.name]},
            feature_description=dataset_config.features),
        output_features=output_features,
        preprocessors=[
            functools.partial(
                tokenize_fn,
                spectrogram_config=spectrogram_config, codec=codec,
                is_training_data='train' in split.name, onsets_only=onsets_only,
                include_ties=include_ties),
            seqio.CacheDatasetPlaceholder(),
            preprocessors.add_unique_id,
            preprocessors.pad_notesequence_array,
            functools.partial(
                t5.data.preprocessors.split_tokens_to_inputs_length,
                feature_key='inputs',
                additional_feature_keys=['input_times', 'sequence'],
                passthrough_feature_keys=['unique_id']),
            # Add dummy targets as they are dropped during the above split to
            # avoid memory blowups, but expected to be present by seqio; the
            # evaluation metrics currently only use the target NoteSequence.
            preprocessors.add_dummy_targets,
            functools.partial(
                preprocessors.compute_spectrograms,
                spectrogram_config=spectrogram_config),
            functools.partial(preprocessors.handle_too_long, skip=False),
            functools.partial(
                seqio.preprocessors.tokenize_and_append_eos,
                copy_pretokenized=False)
        ],
        postprocess_fn=functools.partial(postprocess, codec=codec),
        metric_fns=[
            functools.partial(
                metrics.transcription_metrics,
                codec=codec,
                spectrogram_config=spectrogram_config,
                onsets_only=onsets_only,
                use_ties=include_ties,
                track_specs=track_specs)
        ],
    )

  seqio.MixtureRegistry.add(
      construct_task_name(
          task_prefix=task_prefix, spectrogram_config=spectrogram_config,
          vocab_config=vocab_config, task_suffix='eval'),
      mixture_task_names,
      default_rate=1)


# Just use default spectrogram config.
SPECTROGRAM_CONFIG = spectrograms.SpectrogramConfig()

# Create two vocabulary configs, one default and one with only on-off velocity.
VOCAB_CONFIG_FULL = vocabularies.VocabularyConfig()
VOCAB_CONFIG_NOVELOCITY = vocabularies.VocabularyConfig(num_velocity_bins=1)

# Transcribe MAESTRO v1.
add_transcription_task_to_registry(
    dataset_config=datasets.MAESTROV1_CONFIG,
    spectrogram_config=SPECTROGRAM_CONFIG,
    vocab_config=VOCAB_CONFIG_FULL,
    tokenize_fn=functools.partial(
        preprocessors.tokenize_transcription_example,
        audio_is_samples=False,
        id_feature_key='id'),
    onsets_only=False,
    include_ties=False)

# Transcribe MAESTRO v3.
add_transcription_task_to_registry(
    dataset_config=datasets.MAESTROV3_CONFIG,
    spectrogram_config=SPECTROGRAM_CONFIG,
    vocab_config=VOCAB_CONFIG_FULL,
    tokenize_fn=functools.partial(
        preprocessors.tokenize_transcription_example,
        audio_is_samples=False,
        id_feature_key='id'),
    onsets_only=False,
    include_ties=False)

# Transcribe MAESTRO v3 without velocities, with ties.
add_transcription_task_to_registry(
    dataset_config=datasets.MAESTROV3_CONFIG,
    spectrogram_config=SPECTROGRAM_CONFIG,
    vocab_config=VOCAB_CONFIG_NOVELOCITY,
    tokenize_fn=functools.partial(
        preprocessors.tokenize_transcription_example,
        audio_is_samples=False,
        id_feature_key='id'),
    onsets_only=False,
    include_ties=True)

# Transcribe GuitarSet, with ties.
add_transcription_task_to_registry(
    dataset_config=datasets.GUITARSET_CONFIG,
    spectrogram_config=SPECTROGRAM_CONFIG,
    vocab_config=VOCAB_CONFIG_NOVELOCITY,
    tokenize_fn=preprocessors.tokenize_guitarset_example,
    onsets_only=False,
    include_ties=True)

# Transcribe URMP mixes, with ties.
add_transcription_task_to_registry(
    dataset_config=datasets.URMP_CONFIG,
    spectrogram_config=SPECTROGRAM_CONFIG,
    vocab_config=VOCAB_CONFIG_NOVELOCITY,
    tokenize_fn=functools.partial(
        preprocessors.tokenize_example_with_program_lookup,
        inst_name_to_program_fn=preprocessors.urmp_instrument_to_program,
        id_feature_key='id'),
    onsets_only=False,
    include_ties=True)

# Transcribe MusicNet, with ties.
add_transcription_task_to_registry(
    dataset_config=datasets.MUSICNET_CONFIG,
    spectrogram_config=SPECTROGRAM_CONFIG,
    vocab_config=VOCAB_CONFIG_NOVELOCITY,
    tokenize_fn=functools.partial(
        preprocessors.tokenize_transcription_example,
        audio_is_samples=True,
        id_feature_key='id'),
    onsets_only=False,
    include_ties=True)

# Transcribe MusicNetEM, with ties.
add_transcription_task_to_registry(
    dataset_config=datasets.MUSICNET_EM_CONFIG,
    spectrogram_config=SPECTROGRAM_CONFIG,
    vocab_config=VOCAB_CONFIG_NOVELOCITY,
    tokenize_fn=functools.partial(
        preprocessors.tokenize_transcription_example,
        audio_is_samples=True,
        id_feature_key='id'),
    onsets_only=False,
    include_ties=True)

# Transcribe Cerberus4 (piano-guitar-bass-drums quartets), with ties.
add_transcription_task_to_registry(
    dataset_config=datasets.CERBERUS4_CONFIG,
    spectrogram_config=SPECTROGRAM_CONFIG,
    vocab_config=VOCAB_CONFIG_NOVELOCITY,
    tokenize_fn=functools.partial(
        preprocessors.tokenize_slakh_example,
        track_specs=datasets.CERBERUS4_CONFIG.track_specs,
        ignore_pitch_bends=True),
    onsets_only=False,
    include_ties=True)

# Transcribe 10 random sub-mixes of each song from Slakh, with ties.
add_transcription_task_to_registry(
    dataset_config=datasets.SLAKH_CONFIG,
    spectrogram_config=SPECTROGRAM_CONFIG,
    vocab_config=VOCAB_CONFIG_NOVELOCITY,
    tokenize_fn=functools.partial(
        preprocessors.tokenize_slakh_example,
        track_specs=None,
        ignore_pitch_bends=True),
    onsets_only=False,
    include_ties=True)


# Construct task names to include in transcription mixture.
MIXTURE_DATASET_NAMES = [
    'maestrov3', 'guitarset', 'urmp', 'musicnet_em', 'cerberus4', 'slakh'
]
MIXTURE_TRAIN_TASK_NAMES = []
MIXTURE_EVAL_TASK_NAMES = []
MIXTURE_TEST_TASK_NAMES = []
for dataset_name in MIXTURE_DATASET_NAMES:
  MIXTURE_TRAIN_TASK_NAMES.append(
      construct_task_name(task_prefix=f'{dataset_name}_notes_ties',
                          spectrogram_config=SPECTROGRAM_CONFIG,
                          vocab_config=VOCAB_CONFIG_NOVELOCITY,
                          task_suffix='train'))
  MIXTURE_EVAL_TASK_NAMES.append(
      construct_task_name(task_prefix=f'{dataset_name}_notes_ties',
                          spectrogram_config=SPECTROGRAM_CONFIG,
                          vocab_config=VOCAB_CONFIG_NOVELOCITY,
                          task_suffix='validation'))
MIXING_TEMPERATURE = 10 / 3

# Add the mixture of all transcription tasks, with ties.
seqio.MixtureRegistry.add(
    construct_task_name(
        task_prefix='mega_notes_ties',
        spectrogram_config=SPECTROGRAM_CONFIG,
        vocab_config=VOCAB_CONFIG_NOVELOCITY,
        task_suffix='train'),
    MIXTURE_TRAIN_TASK_NAMES,
    default_rate=functools.partial(
        seqio.mixing_rate_num_examples,
        temperature=MIXING_TEMPERATURE))
seqio.MixtureRegistry.add(
    construct_task_name(
        task_prefix='mega_notes_ties',
        spectrogram_config=SPECTROGRAM_CONFIG,
        vocab_config=VOCAB_CONFIG_NOVELOCITY,
        task_suffix='eval'),
    MIXTURE_EVAL_TASK_NAMES,
    default_rate=functools.partial(
        seqio.mixing_rate_num_examples,
        temperature=MIXING_TEMPERATURE))