-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathsave_logits.py
197 lines (169 loc) · 7.79 KB
/
save_logits.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
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
from collections import defaultdict
import tempfile
from torch.cuda.amp import autocast
import os
from typing import List, Union
import torch
import numpy as np
from loguru import logger
from tqdm import tqdm
import pandas as pd
from pathlib import Path
from packaging import version
from h5py import File
from fire import Fire
import ignite.distributed as idist
import utils
import dataset
import models
SR = 16000
infer_mode = torch.inference_mode if version.parse(
torch.__version__) > version.parse('1.8.2') else torch.no_grad
torch.backends.cudnn.deterministic = True
def create_model_proxy(model_name: Union[str, List], model_args):
if isinstance(model_name, list):
mdls = [
getattr(models, mdl_name)(pretrained=True, **model_args) for mdl_name in model_name
]
return models.EnsembleModel(mdls)
else:
return getattr(models, model_name)(pretrained=True, **model_args)
def forward_spec_amp(model, data):
with autocast(enabled=True):
return model.forward_spectrogram(data).to(torch.float32)
def forward_spec(model, data):
return model.forward_spectrogram(data)
def _save_logits_parallel(rank, parquet_files, output_dir, wavtransforms,
spectransforms, config_parameters):
model = create_model_proxy(config_parameters.model,
config_parameters.model_args)
device = idist.device()
logger.debug(f"Starting inference on {device=}")
model = model.to(device).eval()
if idist.get_world_size() > 1 and rank == 0:
logger.info("\nDistributed setting:")
logger.info(f"\tbackend: {idist.backend()}")
logger.info(f"\tworld size: {idist.get_world_size()}")
logger.info("\n")
dataloader = idist.auto_dataloader(
dataset.SequentialHDF5DataLoader(
pd.read_parquet(parquet_files[rank]),
chunk_length=config_parameters.chunk_length,
wavtransforms=wavtransforms,
),
num_workers=config_parameters.num_workers,
collate_fn=dataset.sequential_pad_logits,
worker_init_fn=dataset.
worker_init_fn, # Required, otherwise workers have same data
persistent_workers=
True, # super important otherwise worker_init_fn is called again, will take forever
batch_size=config_parameters.batch_size)
forward_function = forward_spec_amp if config_parameters.mode == 'amp' else forward_spec
for epoch in tqdm(range(1, config_parameters.epochs + 1),
desc='Epoch',
disable=(rank != 0)):
output_hdf5 = output_dir / f"logits_ep{epoch}_{rank}.h5"
output_parquet = output_dir / f"logits_ep{epoch}_{rank}.parquet"
indexed_data = defaultdict(list)
c = 0
with File(output_hdf5, 'w') as store:
vals = store.create_dataset('topk_values',
dtype='float16',
maxshape=(None,
config_parameters.topk),
shape=(config_parameters.batch_size,
config_parameters.topk))
idxs = store.create_dataset('topk_indexes',
dtype='int16',
maxshape=(None,
config_parameters.topk),
shape=(config_parameters.batch_size,
config_parameters.topk))
seeds = store.create_dataset(
'seeds',
dtype='int32',
maxshape=(None, ),
shape=(config_parameters.batch_size, ))
for data, starts, ends, fnames, seed, specaug_vals, specaug_minvals in tqdm(
dataloader, leave=False, desc='sample', disable=(rank
!= 0)):
data = data.to(device)
with torch.inference_mode():
data = model.front_end(data)
data = spectransforms(
data,
values=torch.tensor(specaug_vals),
min_values=torch.tensor(specaug_minvals))
y = forward_function(model, data).to('cpu')
# y = model.forward_spectrogram(data).to('cpu')
y_vals, y_idxs = y.topk(config_parameters.topk)
if c >= vals.shape[0]:
vals.resize(
(c + data.shape[0], config_parameters.topk))
idxs.resize(
(c + data.shape[0], config_parameters.topk))
seeds.resize((c + data.shape[0], ))
vals[c:c + y.shape[0], :] = y_vals.numpy()
idxs[c:c + y.shape[0], :] = y_idxs.numpy()
seeds[c:c + y.shape[0]] = seed
indexed_data['filename'].extend(fnames)
indexed_data['start'].extend(starts)
indexed_data['end'].extend(ends)
c += y.shape[0]
# Reshape the data to not waste space
vals.resize((c, config_parameters.topk))
idxs.resize((c, config_parameters.topk))
seeds.resize((c, ))
ret_df = pd.DataFrame(indexed_data)
ret_df['start'] = ret_df['start'].astype('float32')
ret_df['end'] = ret_df['end'].astype('float32')
ret_df.to_parquet(output_parquet, index=False, compression='brotli')
def save_parallel(config, **override_kwargs):
# Best use with torchrun:
# torchrun --nproc_per_node 8 save_logits.py logitconfig/balanced.yaml --num_workers 8
config_parameters = utils.parse_config_or_kwargs(config,
config_type=utils.Config,
**override_kwargs)
world_size = int(os.environ.get('WORLD_SIZE', 1))
model_name = config_parameters.model
if isinstance(config_parameters.model, List):
model_name = [
mdl_name.replace('audiotransformer_', '')
for mdl_name in config_parameters.model
]
model_name = '_'.join(model_name)
output_dir = Path(config_parameters.outputpath) / model_name / Path(
config_parameters.train_data
).stem / f"chunk_{config_parameters.chunk_length:.0f}"
output_dir.mkdir(exist_ok=True, parents=True)
wavtransforms = utils.parse_wavtransforms(config_parameters.wavtransforms)
spectransforms = utils.parse_spectransforms(
config_parameters.spectransforms)
torch.save(
{
'wavtransforms': wavtransforms,
'spectransforms': spectransforms,
'config': config_parameters
}, output_dir / 'meta.pt')
df = pd.read_csv(config_parameters.train_data,
converters={'filename': lambda x: Path(x).name},
sep='\t')
#Split dataframe into parts
dfs: List[pd.DataFrame] = np.array_split(df, world_size)
tempfiles = [
tempfile.NamedTemporaryFile(suffix='.parquet') for _ in range(len(dfs))
]
# Dump splits for each worker process
for i in range(len(dfs)):
dfs[i].to_parquet(tempfiles[i].name, compression='brotli')
local_rank = int(os.environ.get('LOCAL_RANK', 0))
if local_rank == 0:
logger.info(f"Output can be found at {output_dir}")
with idist.Parallel(backend='gloo') as parallel:
parallel.run(_save_logits_parallel, [tmp.name for tmp in tempfiles],
output_dir, wavtransforms, spectransforms,
config_parameters)
for tmp in tempfiles:
tmp.close()
if __name__ == "__main__":
Fire(save_parallel)