Spaces:
Build error
Build error
File size: 16,018 Bytes
daf1ccd |
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 |
# Ke Chen
# knutchen@ucsd.edu
# HTS-AT: A HIERARCHICAL TOKEN-SEMANTIC AUDIO TRANSFORMER FOR SOUND CLASSIFICATION AND DETECTION
# The Model Training Wrapper
import numpy as np
import librosa
import os
import bisect
from numpy.lib.function_base import average
from sklearn.metrics import average_precision_score, roc_auc_score, accuracy_score
from utils import get_loss_func, get_mix_lambda, d_prime
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as cp
import torch.optim as optim
from torch.nn.parameter import Parameter
import torch.distributed as dist
import pytorch_lightning as pl
from utils import do_mixup, get_mix_lambda, do_mixup_label
class SEDWrapper(pl.LightningModule):
def __init__(self, sed_model, config, dataset):
super().__init__()
self.sed_model = sed_model
self.config = config
self.dataset = dataset
self.loss_func = get_loss_func(config.loss_type)
def evaluate_metric(self, pred, ans):
ap = []
if self.config.dataset_type == "audioset":
mAP = np.mean(average_precision_score(ans, pred, average = None))
mAUC = np.mean(roc_auc_score(ans, pred, average = None))
dprime = d_prime(mAUC)
return {"mAP": mAP, "mAUC": mAUC, "dprime": dprime}
else:
acc = accuracy_score(ans, np.argmax(pred, 1))
return {"acc": acc}
def forward(self, x, mix_lambda = None):
output_dict = self.sed_model(x, mix_lambda)
return output_dict["clipwise_output"], output_dict["framewise_output"]
def inference(self, x):
self.device_type = next(self.parameters()).device
self.eval()
x = torch.from_numpy(x).float().to(self.device_type)
print(x.shape)
output_dict = self.sed_model(x, None, True)
for key in output_dict.keys():
output_dict[key] = output_dict[key].detach().cpu().numpy()
return output_dict
def training_step(self, batch, batch_idx):
self.device_type = next(self.parameters()).device
mix_lambda = torch.from_numpy(get_mix_lambda(0.5, len(batch["waveform"]))).to(self.device_type)
# Another Choice: also mixup the target, but AudioSet is not a perfect data
# so "adding noise" might be better than purly "mix"
# batch["target"] = do_mixup_label(batch["target"])
# batch["target"] = do_mixup(batch["target"], mix_lambda)
pred, _ = self(batch["waveform"], mix_lambda)
loss = self.loss_func(pred, batch["target"])
self.log("loss", loss, on_epoch= True, prog_bar=True)
return loss
def training_epoch_end(self, outputs):
# Change: SWA, deprecated
# for opt in self.trainer.optimizers:
# if not type(opt) is SWA:
# continue
# opt.swap_swa_sgd()
self.dataset.generate_queue()
def validation_step(self, batch, batch_idx):
pred, _ = self(batch["waveform"])
return [pred.detach(), batch["target"].detach()]
def validation_epoch_end(self, validation_step_outputs):
self.device_type = next(self.parameters()).device
pred = torch.cat([d[0] for d in validation_step_outputs], dim = 0)
target = torch.cat([d[1] for d in validation_step_outputs], dim = 0)
gather_pred = [torch.zeros_like(pred) for _ in range(dist.get_world_size())]
gather_target = [torch.zeros_like(target) for _ in range(dist.get_world_size())]
dist.barrier()
if self.config.dataset_type == "audioset":
metric_dict = {
"mAP": 0.,
"mAUC": 0.,
"dprime": 0.
}
else:
metric_dict = {
"acc":0.
}
dist.all_gather(gather_pred, pred)
dist.all_gather(gather_target, target)
if dist.get_rank() == 0:
gather_pred = torch.cat(gather_pred, dim = 0).cpu().numpy()
gather_target = torch.cat(gather_target, dim = 0).cpu().numpy()
if self.config.dataset_type == "scv2":
gather_target = np.argmax(gather_target, 1)
metric_dict = self.evaluate_metric(gather_pred, gather_target)
print(self.device_type, dist.get_world_size(), metric_dict, flush = True)
if self.config.dataset_type == "audioset":
self.log("mAP", metric_dict["mAP"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
self.log("mAUC", metric_dict["mAUC"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
self.log("dprime", metric_dict["dprime"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
else:
self.log("acc", metric_dict["acc"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
dist.barrier()
def time_shifting(self, x, shift_len):
shift_len = int(shift_len)
new_sample = torch.cat([x[:, shift_len:], x[:, :shift_len]], axis = 1)
return new_sample
def test_step(self, batch, batch_idx):
print(batch['waveform'].shape)
exit()
self.device_type = next(self.parameters()).device
preds = []
# time shifting optimization
if self.config.fl_local or self.config.dataset_type != "audioset":
shift_num = 1 # framewise localization cannot allow the time shifting
else:
shift_num = 10
for i in range(shift_num):
pred, pred_map = self(batch["waveform"])
preds.append(pred.unsqueeze(0))
batch["waveform"] = self.time_shifting(batch["waveform"], shift_len = 100 * (i + 1))
preds = torch.cat(preds, dim=0)
pred = preds.mean(dim = 0)
if self.config.fl_local:
return [
pred.detach().cpu().numpy(),
pred_map.detach().cpu().numpy(),
batch["audio_name"],
batch["real_len"].cpu().numpy()
]
else:
return [pred.detach(), batch["target"].detach()]
def test_epoch_end(self, test_step_outputs):
self.device_type = next(self.parameters()).device
if self.config.fl_local:
pred = np.concatenate([d[0] for d in test_step_outputs], axis = 0)
pred_map = np.concatenate([d[1] for d in test_step_outputs], axis = 0)
audio_name = np.concatenate([d[2] for d in test_step_outputs], axis = 0)
real_len = np.concatenate([d[3] for d in test_step_outputs], axis = 0)
heatmap_file = os.path.join(self.config.heatmap_dir, self.config.test_file + "_" + str(self.device_type) + ".npy")
save_npy = [
{
"audio_name": audio_name[i],
"heatmap": pred_map[i],
"pred": pred[i],
"real_len":real_len[i]
}
for i in range(len(pred))
]
np.save(heatmap_file, save_npy)
else:
self.device_type = next(self.parameters()).device
pred = torch.cat([d[0] for d in test_step_outputs], dim = 0)
target = torch.cat([d[1] for d in test_step_outputs], dim = 0)
gather_pred = [torch.zeros_like(pred) for _ in range(dist.get_world_size())]
gather_target = [torch.zeros_like(target) for _ in range(dist.get_world_size())]
dist.barrier()
if self.config.dataset_type == "audioset":
metric_dict = {
"mAP": 0.,
"mAUC": 0.,
"dprime": 0.
}
else:
metric_dict = {
"acc":0.
}
dist.all_gather(gather_pred, pred)
dist.all_gather(gather_target, target)
if dist.get_rank() == 0:
gather_pred = torch.cat(gather_pred, dim = 0).cpu().numpy()
gather_target = torch.cat(gather_target, dim = 0).cpu().numpy()
if self.config.dataset_type == "scv2":
gather_target = np.argmax(gather_target, 1)
metric_dict = self.evaluate_metric(gather_pred, gather_target)
print(self.device_type, dist.get_world_size(), metric_dict, flush = True)
if self.config.dataset_type == "audioset":
self.log("mAP", metric_dict["mAP"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
self.log("mAUC", metric_dict["mAUC"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
self.log("dprime", metric_dict["dprime"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
else:
self.log("acc", metric_dict["acc"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
dist.barrier()
def configure_optimizers(self):
optimizer = optim.AdamW(
filter(lambda p: p.requires_grad, self.parameters()),
lr = self.config.learning_rate,
betas = (0.9, 0.999), eps = 1e-08, weight_decay = 0.05,
)
# Change: SWA, deprecated
# optimizer = SWA(optimizer, swa_start=10, swa_freq=5)
def lr_foo(epoch):
if epoch < 3:
# warm up lr
lr_scale = self.config.lr_rate[epoch]
else:
# warmup schedule
lr_pos = int(-1 - bisect.bisect_left(self.config.lr_scheduler_epoch, epoch))
if lr_pos < -3:
lr_scale = max(self.config.lr_rate[0] * (0.98 ** epoch), 0.03 )
else:
lr_scale = self.config.lr_rate[lr_pos]
return lr_scale
scheduler = optim.lr_scheduler.LambdaLR(
optimizer,
lr_lambda=lr_foo
)
return [optimizer], [scheduler]
class Ensemble_SEDWrapper(pl.LightningModule):
def __init__(self, sed_models, config, dataset):
super().__init__()
self.sed_models = nn.ModuleList(sed_models)
self.config = config
self.dataset = dataset
def evaluate_metric(self, pred, ans):
if self.config.dataset_type == "audioset":
mAP = np.mean(average_precision_score(ans, pred, average = None))
mAUC = np.mean(roc_auc_score(ans, pred, average = None))
dprime = d_prime(mAUC)
return {"mAP": mAP, "mAUC": mAUC, "dprime": dprime}
else:
acc = accuracy_score(ans, np.argmax(pred, 1))
return {"acc": acc}
def forward(self, x, sed_index, mix_lambda = None):
self.sed_models[sed_index].eval()
preds = []
pred_maps = []
# time shifting optimization
if self.config.fl_local or self.config.dataset_type != "audioset":
shift_num = 1 # framewise localization cannot allow the time shifting
else:
shift_num = 10
for i in range(shift_num):
pred, pred_map = self.sed_models[sed_index](x)
pred_maps.append(pred_map.unsqueeze(0))
preds.append(pred.unsqueeze(0))
x = self.time_shifting(x, shift_len = 100 * (i + 1))
preds = torch.cat(preds, dim=0)
pred_maps = torch.cat(pred_maps, dim = 0)
pred = preds.mean(dim = 0)
pred_map = pred_maps.mean(dim = 0)
return pred, pred_map
def time_shifting(self, x, shift_len):
shift_len = int(shift_len)
new_sample = torch.cat([x[:, shift_len:], x[:, :shift_len]], axis = 1)
return new_sample
def test_step(self, batch, batch_idx):
self.device_type = next(self.parameters()).device
if self.config.fl_local:
pred = torch.zeros(len(batch["waveform"]), self.config.classes_num).float().to(self.device_type)
pred_map = torch.zeros(len(batch["waveform"]), 1024, self.config.classes_num).float().to(self.device_type)
for j in range(len(self.sed_models)):
temp_pred, temp_pred_map = self(batch["waveform"], j)
pred = pred + temp_pred
pred_map = pred_map + temp_pred_map
pred = pred / len(self.sed_models)
pred_map = pred_map / len(self.sed_models)
return [
pred.detach().cpu().numpy(),
pred_map.detach().cpu().numpy(),
batch["audio_name"],
batch["real_len"].cpu().numpy()
]
else:
pred = torch.zeros(len(batch["waveform"]), self.config.classes_num).float().to(self.device_type)
for j in range(len(self.sed_models)):
temp_pred, _ = self(batch["waveform"], j)
pred = pred + temp_pred
pred = pred / len(self.sed_models)
return [
pred.detach(),
batch["target"].detach(),
]
def test_epoch_end(self, test_step_outputs):
self.device_type = next(self.parameters()).device
if self.config.fl_local:
pred = np.concatenate([d[0] for d in test_step_outputs], axis = 0)
pred_map = np.concatenate([d[1] for d in test_step_outputs], axis = 0)
audio_name = np.concatenate([d[2] for d in test_step_outputs], axis = 0)
real_len = np.concatenate([d[3] for d in test_step_outputs], axis = 0)
heatmap_file = os.path.join(self.config.heatmap_dir, self.config.test_file + "_" + str(self.device_type) + ".npy")
print(pred.shape)
print(pred_map.shape)
print(real_len.shape)
save_npy = [
{
"audio_name": audio_name[i],
"heatmap": pred_map[i],
"pred": pred[i],
"real_len":real_len[i]
}
for i in range(len(pred))
]
np.save(heatmap_file, save_npy)
else:
pred = torch.cat([d[0] for d in test_step_outputs], dim = 0)
target = torch.cat([d[1] for d in test_step_outputs], dim = 0)
gather_pred = [torch.zeros_like(pred) for _ in range(dist.get_world_size())]
gather_target = [torch.zeros_like(target) for _ in range(dist.get_world_size())]
dist.barrier()
if self.config.dataset_type == "audioset":
metric_dict = {
"mAP": 0.,
"mAUC": 0.,
"dprime": 0.
}
else:
metric_dict = {
"acc":0.
}
dist.all_gather(gather_pred, pred)
dist.all_gather(gather_target, target)
if dist.get_rank() == 0:
gather_pred = torch.cat(gather_pred, dim = 0).cpu().numpy()
gather_target = torch.cat(gather_target, dim = 0).cpu().numpy()
if self.config.dataset_type == "scv2":
gather_target = np.argmax(gather_target, 1)
metric_dict = self.evaluate_metric(gather_pred, gather_target)
print(self.device_type, dist.get_world_size(), metric_dict, flush = True)
if self.config.dataset_type == "audioset":
self.log("mAP", metric_dict["mAP"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
self.log("mAUC", metric_dict["mAUC"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
self.log("dprime", metric_dict["dprime"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
else:
self.log("acc", metric_dict["acc"] * float(dist.get_world_size()), on_epoch = True, prog_bar=True, sync_dist=True)
dist.barrier()
|