File size: 10,974 Bytes
affcd23 |
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 |
import random
from pathlib import Path
from random import shuffle
import PIL
import pandas as pd
import numpy as np
import torch
import torch.nn.functional as F
from torch.nn.utils import clip_grad_norm_
from torchvision.transforms import ToTensor
from tqdm import tqdm
from hw_asr.base import BaseTrainer
from hw_asr.base.base_text_encoder import BaseTextEncoder
from hw_asr.logger.utils import plot_spectrogram_to_buf
from hw_asr.metric.utils import calc_cer, calc_wer
from hw_asr.utils import inf_loop, MetricTracker
class Trainer(BaseTrainer):
"""
Trainer class
"""
def __init__(
self,
model,
criterion,
metrics,
optimizer,
config,
device,
dataloaders,
text_encoder,
lr_scheduler=None,
len_epoch=None,
skip_oom=True,
):
super().__init__(model, criterion, metrics, optimizer, config, device)
self.skip_oom = skip_oom
self.text_encoder = text_encoder
self.config = config
self.train_dataloader = dataloaders["train"]
if len_epoch is None:
# epoch-based training
self.len_epoch = len(self.train_dataloader)
else:
# iteration-based training
self.train_dataloader = inf_loop(self.train_dataloader)
self.len_epoch = len_epoch
self.evaluation_dataloaders = {k: v for k, v in dataloaders.items() if k != "train"}
self.lr_scheduler = lr_scheduler
self.log_step = 50
self.train_metrics = MetricTracker("loss", "grad norm", *[m.name for m in self.metrics], writer=self.writer)
self.evaluation_metrics = MetricTracker("loss", *[m.name for m in self.metrics], writer=self.writer)
@staticmethod
def move_batch_to_device(batch, device: torch.device):
"""
Move all necessary tensors to the HPU
"""
for tensor_for_gpu in ["spectrogram", "text_encoded"]:
batch[tensor_for_gpu] = batch[tensor_for_gpu].to(device)
return batch
def _clip_grad_norm(self):
if self.config["trainer"].get("grad_norm_clip", None) is not None:
clip_grad_norm_(self.model.parameters(), self.config["trainer"]["grad_norm_clip"])
def _train_epoch(self, epoch):
"""
Training logic for an epoch
:param epoch: Integer, current training epoch.
:return: A log that contains average loss and metric in this epoch.
"""
self.model.train()
self.train_metrics.reset()
self.writer.add_scalar("epoch", epoch)
for batch_idx, batch in enumerate(tqdm(self.train_dataloader, desc="train", total=self.len_epoch - 1)):
try:
batch = self.process_batch(
batch,
is_train=True,
metrics=self.train_metrics,
)
except RuntimeError as e:
if "out of memory" in str(e) and self.skip_oom:
self.logger.warning("OOM on batch. Skipping batch.")
for p in self.model.parameters():
if p.grad is not None:
del p.grad # free some memory
torch.cuda.empty_cache()
continue
else:
raise e
self.train_metrics.update("grad norm", self.get_grad_norm())
if batch_idx % self.log_step == 0:
self.writer.set_step((epoch - 1) * self.len_epoch + batch_idx)
self.logger.debug("Train Epoch: {} {} Loss: {:.6f}".format(epoch, self._progress(batch_idx), batch["loss"].item()))
self.writer.add_scalar("learning rate", self.lr_scheduler.get_last_lr()[0])
self._log_predictions(**batch)
self._log_spectrogram(batch["spectrogram"])
self._log_scalars(self.train_metrics)
# we don't want to reset train metrics at the start of every epoch
# because we are interested in recent train metrics
last_train_metrics = self.train_metrics.result()
self.train_metrics.reset()
if batch_idx + 1 >= self.len_epoch:
break
log = last_train_metrics
for part, dataloader in self.evaluation_dataloaders.items():
val_log = self._evaluation_epoch(epoch, part, dataloader)
log.update(**{f"{part}_{name}": value for name, value in val_log.items()})
return log
def process_batch(self, batch, is_train: bool, metrics: MetricTracker, part: str = None, epoch: int = None):
batch = self.move_batch_to_device(batch, self.device)
if is_train:
self.optimizer.zero_grad()
outputs = self.model(**batch)
if type(outputs) is dict:
batch.update(outputs)
else:
batch["logits"] = outputs
batch["log_probs"] = F.log_softmax(batch["logits"], dim=-1)
batch["log_probs_length"] = self.model.transform_input_lengths(batch["spectrogram_length"])
batch["loss"] = self.criterion(**batch)
if is_train:
batch["loss"].backward()
self._clip_grad_norm()
self.optimizer.step()
if self.lr_scheduler is not None:
self.lr_scheduler.step()
metrics.update("loss", batch["loss"].item())
for met in self.metrics:
is_not_test = is_train or ("val" in part)
is_test = not is_not_test
hard_to_calc_metric = "beam search" in met.name or "LM" in met.name
if hard_to_calc_metric and (is_not_test or (is_test and (epoch % 25) != 0)):
continue
metrics.update(met.name, met(**batch))
return batch
def _evaluation_epoch(self, epoch, part, dataloader):
"""
Validate after training an epoch
:param epoch: Integer, current training epoch.
:return: A log that contains information about validation
"""
self.model.eval()
self.evaluation_metrics.reset()
with torch.no_grad():
for batch_idx, batch in tqdm(
enumerate(dataloader),
desc=part,
total=len(dataloader),
):
batch = self.process_batch(batch, is_train=False, metrics=self.evaluation_metrics, part=part, epoch=epoch)
self.writer.set_step(epoch * self.len_epoch, part)
self._log_predictions(**batch)
self._log_spectrogram(batch["spectrogram"])
self._log_scalars(self.evaluation_metrics)
# add histogram of model parameters to the tensorboard
# for name, p in self.model.named_parameters():
# self.writer.add_histogram(name, p, bins="auto")
return self.evaluation_metrics.result()
def _progress(self, batch_idx):
base = "[{}/{} ({:.0f}%)]"
if hasattr(self.train_dataloader, "n_samples"):
current = batch_idx * self.train_dataloader.batch_size
total = self.train_dataloader.n_samples
else:
current = batch_idx
total = self.len_epoch
return base.format(current, total, 100.0 * current / total)
def _log_predictions(
self,
text,
logits,
log_probs,
log_probs_length,
audio_path,
audio,
examples_to_log=10,
*args,
**kwargs,
):
# TODO: implement logging of beam search results
if self.writer is None:
return
ids = np.random.choice(len(text), examples_to_log, replace=False)
text = [text[i] for i in ids]
logits = logits[ids]
log_probs = log_probs[ids]
log_probs_length = log_probs_length[ids]
audio_path = [audio_path[i] for i in ids]
audio = [audio[i] for i in ids]
argmax_inds = log_probs.cpu().argmax(-1).numpy()
argmax_inds = [inds[: int(ind_len)] for inds, ind_len in zip(argmax_inds, log_probs_length.numpy())]
argmax_texts_raw = [self.text_encoder.decode(inds) for inds in argmax_inds]
argmax_texts = [self.text_encoder.ctc_decode(inds) for inds in argmax_inds]
probs = np.exp(log_probs.detach().cpu().numpy())
probs_length = log_probs_length.detach().cpu().numpy()
bs_preds = [self.text_encoder.ctc_beam_search(prob[:prob_length], 4) for prob, prob_length in zip(probs, probs_length)]
logits = logits.detach().cpu().numpy()
lm_preds = [self.text_encoder.ctc_lm_beam_search(logit[:length]) for logit, length in zip(logits, probs_length)]
tuples = list(zip(argmax_texts, bs_preds, lm_preds, text, argmax_texts_raw, audio_path, audio))
rows = {}
for pred, bs_pred, lm_pred, target, raw_pred, audio_path, audio in tuples:
target = BaseTextEncoder.normalize_text(target)
wer = calc_wer(target, pred) * 100
cer = calc_cer(target, pred) * 100
bs_wer = calc_wer(target, bs_pred) * 100
bs_cer = calc_cer(target, bs_pred) * 100
lm_wer = calc_wer(target, lm_pred) * 100
lm_cer = calc_cer(target, lm_pred) * 100
rows[Path(audio_path).name] = {
"orig_audio": self.writer.wandb.Audio(audio_path), # inaccurate, but no changes in the template
"augm_audio": self.writer.wandb.Audio(audio.squeeze().numpy(), sample_rate=16000), # inaccurate, but no changes in the template
"target": target,
"raw pred": raw_pred,
"pred": pred,
"bs pred": bs_pred,
"lm pred": lm_pred,
"wer": wer,
"cer": cer,
"bs wer": bs_wer,
"bs cer": bs_cer,
"lm wer": lm_wer,
"lm cer": lm_cer,
}
self.writer.add_table("predictions", pd.DataFrame.from_dict(rows, orient="index"))
def _log_spectrogram(self, spectrogram_batch):
spectrogram = random.choice(spectrogram_batch.cpu())
image = PIL.Image.open(plot_spectrogram_to_buf(spectrogram))
self.writer.add_image("spectrogram", ToTensor()(image))
@torch.no_grad()
def get_grad_norm(self, norm_type=2):
parameters = self.model.parameters()
if isinstance(parameters, torch.Tensor):
parameters = [parameters]
parameters = [p for p in parameters if p.grad is not None]
total_norm = torch.norm(
torch.stack([torch.norm(p.grad.detach(), norm_type).cpu() for p in parameters]),
norm_type,
)
return total_norm.item()
def _log_scalars(self, metric_tracker: MetricTracker):
if self.writer is None:
return
for metric_name in metric_tracker.keys():
self.writer.add_scalar(f"{metric_name}", metric_tracker.avg(metric_name))
|