File size: 12,770 Bytes
3b6a091 |
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 |
from torch.utils.data import Dataset
import pandas as pd
import os
import numpy as np
import torchaudio
import random
import torch
import glob
import h5py
from pathlib import Path
def to_mono(mixture, random_ch=False):
if mixture.ndim > 1: # multi channel
if not random_ch:
mixture = torch.mean(mixture, 0)
else: # randomly select one channel
indx = np.random.randint(0, mixture.shape[0] - 1)
mixture = mixture[indx]
return mixture
def pad_audio(audio, target_len, fs):
if audio.shape[-1] < target_len:
audio = torch.nn.functional.pad(
audio, (0, target_len - audio.shape[-1]), mode="constant"
)
padded_indx = [target_len / len(audio)]
onset_s = 0.000
elif len(audio) > target_len:
rand_onset = random.randint(0, len(audio) - target_len)
audio = audio[rand_onset:rand_onset + target_len]
onset_s = round(rand_onset / fs, 3)
padded_indx = [target_len / len(audio)]
else:
onset_s = 0.000
padded_indx = [1.0]
offset_s = round(onset_s + (target_len / fs), 3)
return audio, onset_s, offset_s, padded_indx
def process_labels(df, onset, offset):
df["onset"] = df["onset"] - onset
df["offset"] = df["offset"] - onset
df["onset"] = df.apply(lambda x: max(0, x["onset"]), axis=1)
df["offset"] = df.apply(lambda x: min(10, x["offset"]), axis=1)
df_new = df[(df.onset < df.offset)]
return df_new.drop_duplicates()
def read_audio(file, multisrc, random_channel, pad_to):
mixture, fs = torchaudio.load(file)
if not multisrc:
mixture = to_mono(mixture, random_channel)
if pad_to is not None:
mixture, onset_s, offset_s, padded_indx = pad_audio(mixture, pad_to, fs)
else:
padded_indx = [1.0]
onset_s = None
offset_s = None
mixture = mixture.float()
return mixture, onset_s, offset_s, padded_indx
class StronglyAnnotatedSet(Dataset):
def __init__(
self,
audio_folder,
tsv_entries,
encoder,
pad_to=10,
fs=16000,
return_filename=False,
random_channel=False,
multisrc=False,
feats_pipeline=None,
embeddings_hdf5_file=None,
embedding_type=None
):
self.encoder = encoder
self.fs = fs
self.pad_to = pad_to * fs
self.return_filename = return_filename
self.random_channel = random_channel
self.multisrc = multisrc
self.feats_pipeline = feats_pipeline
self.embeddings_hdf5_file = embeddings_hdf5_file
self.embedding_type = embedding_type
assert embedding_type in ["global", "frame", None], "embedding type are either frame or global or None, got {}".format(embedding_type)
tsv_entries = tsv_entries.dropna()
examples = {}
for i, r in tsv_entries.iterrows():
if r["filename"] not in examples.keys():
examples[r["filename"]] = {
"mixture": os.path.join(audio_folder, r["filename"]),
"events": [],
}
if not np.isnan(r["onset"]):
examples[r["filename"]]["events"].append(
{
"event_label": r["event_label"],
"onset": r["onset"],
"offset": r["offset"],
}
)
else:
if not np.isnan(r["onset"]):
examples[r["filename"]]["events"].append(
{
"event_label": r["event_label"],
"onset": r["onset"],
"offset": r["offset"],
}
)
# we construct a dictionary for each example
self.examples = examples
self.examples_list = list(examples.keys())
if self.embeddings_hdf5_file is not None:
assert self.embedding_type is not None, "If you use embeddings you need to specify also the type (global or frame)"
# fetch dict of positions for each example
self.ex2emb_idx = {}
f = h5py.File(self.embeddings_hdf5_file, "r")
for i, fname in enumerate(f["filenames"]):
self.ex2emb_idx[fname.decode('UTF-8')] = i
self._opened_hdf5 = None
def __len__(self):
return len(self.examples_list)
@property
def hdf5_file(self):
if self._opened_hdf5 is None:
self._opened_hdf5 = h5py.File(self.embeddings_hdf5_file, "r")
return self._opened_hdf5
def __getitem__(self, item):
c_ex = self.examples[self.examples_list[item]]
mixture, onset_s, offset_s, padded_indx = read_audio(
c_ex["mixture"], self.multisrc, self.random_channel, self.pad_to
)
# labels
labels = c_ex["events"]
# to steps
labels_df = pd.DataFrame(labels)
labels_df = process_labels(labels_df, onset_s, offset_s)
# check if labels exists:
if not len(labels_df):
max_len_targets = self.encoder.n_frames
strong = torch.zeros(max_len_targets, len(self.encoder.labels)).float()
else:
strong = self.encoder.encode_strong_df(labels_df)
strong = torch.from_numpy(strong).float()
out_args = [mixture, strong.transpose(0, 1), padded_indx]
if self.feats_pipeline is not None:
# use this function to extract features in the dataloader and apply possibly some data augm
feats = self.feats_pipeline(mixture)
out_args.append(feats)
if self.return_filename:
out_args.append(c_ex["mixture"])
if self.embeddings_hdf5_file is not None:
name = Path(c_ex["mixture"]).stem
index = self.ex2emb_idx[name]
if self.embedding_type == "global":
embeddings = torch.from_numpy(self.hdf5_file["global_embeddings"][index]).float()
elif self.embedding_type == "frame":
embeddings = torch.from_numpy(np.stack(self.hdf5_file["frame_embeddings"][index])).float()
else:
raise NotImplementedError
out_args.append(embeddings)
return out_args
class WeakSet(Dataset):
def __init__(
self,
audio_folder,
tsv_entries,
encoder,
pad_to=10,
fs=16000,
return_filename=False,
random_channel=False,
multisrc=False,
feats_pipeline=None,
embeddings_hdf5_file=None,
embedding_type=None,
):
self.encoder = encoder
self.fs = fs
self.pad_to = pad_to * fs
self.return_filename = return_filename
self.random_channel = random_channel
self.multisrc = multisrc
self.feats_pipeline = feats_pipeline
self.embeddings_hdf5_file = embeddings_hdf5_file
self.embedding_type = embedding_type
assert embedding_type in ["global", "frame",
None], "embedding type are either frame or global or None, got {}".format(
embedding_type)
examples = {}
for i, r in tsv_entries.iterrows():
if r["filename"] not in examples.keys():
examples[r["filename"]] = {
"mixture": os.path.join(audio_folder, r["filename"]),
"events": r["event_labels"].split(","),
}
self.examples = examples
self.examples_list = list(examples.keys())
if self.embeddings_hdf5_file is not None:
assert self.embedding_type is not None, "If you use embeddings you need to specify also the type (global or frame)"
# fetch dict of positions for each example
self.ex2emb_idx = {}
f = h5py.File(self.embeddings_hdf5_file, "r")
for i, fname in enumerate(f["filenames"]):
self.ex2emb_idx[fname.decode('UTF-8')] = i
self._opened_hdf5 = None
def __len__(self):
return len(self.examples_list)
@property
def hdf5_file(self):
if self._opened_hdf5 is None:
self._opened_hdf5 = h5py.File(self.embeddings_hdf5_file, "r")
return self._opened_hdf5
def __getitem__(self, item):
file = self.examples_list[item]
c_ex = self.examples[file]
mixture, _, _, padded_indx = read_audio(
c_ex["mixture"], self.multisrc, self.random_channel, self.pad_to
)
# labels
labels = c_ex["events"]
# check if labels exists:
max_len_targets = self.encoder.n_frames
weak = torch.zeros(max_len_targets, len(self.encoder.labels))
if len(labels):
weak_labels = self.encoder.encode_weak(labels)
weak[0, :] = torch.from_numpy(weak_labels).float()
out_args = [mixture, weak.transpose(0, 1), padded_indx]
if self.feats_pipeline is not None:
feats = self.feats_pipeline(mixture)
out_args.append(feats)
if self.return_filename:
out_args.append(c_ex["mixture"])
if self.embeddings_hdf5_file is not None:
name = Path(c_ex["mixture"]).stem
index = self.ex2emb_idx[name]
if self.embedding_type == "global":
embeddings = torch.from_numpy(self.hdf5_file["global_embeddings"][index]).float()
elif self.embedding_type == "frame":
embeddings = torch.from_numpy(np.stack(self.hdf5_file["frame_embeddings"][index])).float()
else:
raise NotImplementedError
out_args.append(embeddings)
return out_args
class UnlabeledSet(Dataset):
def __init__(
self,
unlabeled_folder,
encoder,
pad_to=10,
fs=16000,
return_filename=False,
random_channel=False,
multisrc=False,
feats_pipeline=None,
embeddings_hdf5_file=None,
embedding_type=None,
):
self.encoder = encoder
self.fs = fs
self.pad_to = pad_to * fs if pad_to is not None else None
self.examples = glob.glob(os.path.join(unlabeled_folder, "*.wav"))
self.return_filename = return_filename
self.random_channel = random_channel
self.multisrc = multisrc
self.feats_pipeline = feats_pipeline
self.embeddings_hdf5_file = embeddings_hdf5_file
self.embedding_type = embedding_type
assert embedding_type in ["global", "frame",
None], "embedding type are either frame or global or None, got {}".format(
embedding_type)
if self.embeddings_hdf5_file is not None:
assert self.embedding_type is not None, "If you use embeddings you need to specify also the type (global or frame)"
# fetch dict of positions for each example
self.ex2emb_idx = {}
f = h5py.File(self.embeddings_hdf5_file, "r")
for i, fname in enumerate(f["filenames"]):
self.ex2emb_idx[fname.decode('UTF-8')] = i
self._opened_hdf5 = None
def __len__(self):
return len(self.examples)
@property
def hdf5_file(self):
if self._opened_hdf5 is None:
self._opened_hdf5 = h5py.File(self.embeddings_hdf5_file, "r")
return self._opened_hdf5
def __getitem__(self, item):
c_ex = self.examples[item]
mixture, _, _, padded_indx = read_audio(
c_ex, self.multisrc, self.random_channel, self.pad_to
)
max_len_targets = self.encoder.n_frames
strong = torch.zeros(max_len_targets, len(self.encoder.labels)).float()
out_args = [mixture, strong.transpose(0, 1), padded_indx]
if self.feats_pipeline is not None:
feats = self.feats_pipeline(mixture)
out_args.append(feats)
if self.return_filename:
out_args.append(c_ex)
if self.embeddings_hdf5_file is not None:
name = Path(c_ex).stem
index = self.ex2emb_idx[name]
if self.embedding_type == "global":
embeddings = torch.from_numpy(self.hdf5_file["global_embeddings"][index]).float()
elif self.embedding_type == "frame":
embeddings = torch.from_numpy(np.stack(self.hdf5_file["frame_embeddings"][index])).float()
else:
raise NotImplementedError
out_args.append(embeddings)
return out_args
|