-
Notifications
You must be signed in to change notification settings - Fork 227
/
extract_durations.py
131 lines (113 loc) · 6.91 KB
/
extract_durations.py
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
import argparse
import pickle
import tensorflow as tf
import numpy as np
from tqdm import tqdm
from p_tqdm import p_umap
from utils.training_config_manager import TrainingConfigManager
from utils.logging_utils import SummaryManager
from data.datasets import AlignerPreprocessor
from utils.alignments import get_durations_from_alignment
from utils.scripts_utils import dynamic_memory_allocation
from data.datasets import AlignerDataset
from data.datasets import DataReader
np.random.seed(42)
tf.random.set_seed(42)
dynamic_memory_allocation()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--config', dest='config', type=str)
parser.add_argument('--best', dest='best', action='store_true',
help='Use best head instead of weighted average of heads.')
parser.add_argument('--autoregressive_weights', type=str, default=None,
help='Explicit path to autoregressive model weights.')
parser.add_argument('--skip_char_pitch', dest='skip_char_pitch', action='store_true')
parser.add_argument('--skip_durations', dest='skip_durations', action='store_true')
args = parser.parse_args()
weighted = not args.best
tag_description = ''.join([
f'{"_weighted" * weighted}{"_best" * (not weighted)}',
])
writer_tag = f'DurationExtraction{tag_description}'
print(writer_tag)
config_manager = TrainingConfigManager(config_path=args.config, aligner=True)
config = config_manager.config
config_manager.print_config()
if not args.skip_durations:
model = config_manager.load_model(args.autoregressive_weights)
if model.r != 1:
print(f"ERROR: model's reduction factor is greater than 1, check config. (r={model.r}")
data_prep = AlignerPreprocessor.from_config(config=config_manager,
tokenizer=model.text_pipeline.tokenizer)
data_handler = AlignerDataset.from_config(config_manager,
preprocessor=data_prep,
kind='phonemized')
target_dir = config_manager.duration_dir
config_manager.dump_config()
dataset = data_handler.get_dataset(bucket_batch_sizes=config['bucket_batch_sizes'],
bucket_boundaries=config['bucket_boundaries'],
shuffle=False,
drop_remainder=False)
last_layer_key = 'Decoder_LastBlock_CrossAttention'
print(f'Extracting attention from layer {last_layer_key}')
summary_manager = SummaryManager(model=model, log_dir=config_manager.log_dir / 'Duration Extraction',
config=config,
default_writer='Duration Extraction')
all_durations = np.array([])
new_alignments = []
iterator = tqdm(enumerate(dataset.all_batches()))
step = 0
for c, (mel_batch, text_batch, stop_batch, file_name_batch) in iterator:
iterator.set_description(f'Processing dataset')
outputs = model.val_step(inp=text_batch,
tar=mel_batch,
stop_prob=stop_batch)
attention_values = outputs['decoder_attention'][last_layer_key].numpy()
text = text_batch.numpy()
mel = mel_batch.numpy()
durations, final_align, jumpiness, peakiness, diag_measure = get_durations_from_alignment(
batch_alignments=attention_values,
mels=mel,
phonemes=text,
weighted=weighted)
batch_avg_jumpiness = tf.reduce_mean(jumpiness, axis=0)
batch_avg_peakiness = tf.reduce_mean(peakiness, axis=0)
batch_avg_diag_measure = tf.reduce_mean(diag_measure, axis=0)
for i in range(tf.shape(jumpiness)[1]):
summary_manager.display_scalar(tag=f'DurationAttentionJumpiness/head{i}',
scalar_value=tf.reduce_mean(batch_avg_jumpiness[i]), step=c)
summary_manager.display_scalar(tag=f'DurationAttentionPeakiness/head{i}',
scalar_value=tf.reduce_mean(batch_avg_peakiness[i]), step=c)
summary_manager.display_scalar(tag=f'DurationAttentionDiagonality/head{i}',
scalar_value=tf.reduce_mean(batch_avg_diag_measure[i]), step=c)
for i, name in enumerate(file_name_batch):
all_durations = np.append(all_durations, durations[i]) # for plotting only
summary_manager.add_image(tag='ExtractedAlignments',
image=tf.expand_dims(tf.expand_dims(final_align[i], 0), -1),
step=step)
step += 1
np.save(str(target_dir / f"{name.numpy().decode('utf-8')}.npy"), durations[i])
all_durations[all_durations >= 20] = 20 # for plotting only
buckets = len(set(all_durations)) # for plotting only
summary_manager.add_histogram(values=all_durations, tag='ExtractedDurations', buckets=buckets)
if not args.skip_char_pitch:
def _pitch_per_char(pitch, durations, mel_len):
durs_cum = np.cumsum(np.pad(durations, (1, 0)))
pitch_char = np.zeros((durations.shape[0],), dtype=np.float)
for idx, a, b in zip(range(mel_len), durs_cum[:-1], durs_cum[1:]):
values = pitch[a:b][np.where(pitch[a:b] != 0.0)[0]]
values = values[np.where((values * pitch_stats['pitch_std'] + pitch_stats['pitch_mean']) < 400)[0]]
pitch_char[idx] = np.mean(values) if len(values) > 0 else 0.0
return pitch_char
def process_per_char_pitch(sample_name: str):
pitch = np.load((config_manager.pitch_dir / sample_name).with_suffix('.npy').as_posix())
durations = np.load((config_manager.duration_dir / sample_name).with_suffix('.npy').as_posix())
mel = np.load((config_manager.mel_dir / sample_name).with_suffix('.npy').as_posix())
char_wise_pitch = _pitch_per_char(pitch, durations, mel.shape[0])
np.save((config_manager.pitch_per_char / sample_name).with_suffix('.npy').as_posix(), char_wise_pitch)
metadatareader = DataReader.from_config(config_manager, kind='phonemized', scan_wavs=False)
pitch_stats = pickle.load(open(config_manager.data_dir / 'pitch_stats.pkl', 'rb'))
print(f'\nComputing phoneme-wise pitch')
print(f'{len(metadatareader.filenames)} items found in {metadatareader.metadata_path}.')
wav_iter = p_umap(process_per_char_pitch, metadatareader.filenames)
print('Done.')