-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
417 lines (331 loc) · 15.9 KB
/
trainer.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
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
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
import sys
import os
import io
import csv
import time
import uuid
import argparse
import torch
import apex
from torch import nn
from torch.nn import functional as F
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from torch.utils.tensorboard import SummaryWriter
from transformers import BertTokenizer, BertForSequenceClassification, AdamW
from logzero import setup_logger
from sklearn import metrics
import pycm
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
import cv2
from tabulate import tabulate
class EmotionDataset(Dataset):
label_index_map = {
'anger': 0,
'disgust': 1,
'joy': 2,
'sadness': 3,
'surprise': 4,
}
def __init__(self, config, phase):
self.config = config
n_labels = len(self.label_index_map.keys())
filepath = os.path.join(config.dataroot, f'{phase}.txt')
self.texts = []
self.labels = torch.empty(0)
with io.open(filepath, encoding='utf-8') as f:
reader = csv.reader(f, delimiter='\t')
for row in reader:
text = row[0]
label_name = 'none' if len(row) == 1 else row[1]
if label_name not in self.label_index_map and phase != 'predict':
self.config.logger.warn(f'{label_name} is invalid label name, skipped')
continue
self.texts.append(text)
labels = torch.zeros(1, n_labels)
if label_name == 'none':
assert self.config.predict, 'label is necessary not when args.predict == true'
else:
index = self.label_index_map[label_name]
labels[0][index] = 1
self.labels = torch.cat([self.labels, labels])
def __getitem__(self, index):
return self.texts[index], self.labels[index]
def __len__(self):
return len(self.texts)
class SemEval2018EmotionDataset(Dataset):
label_index_map = {
'anger' : 0,
'anticipation' : 1,
'disgust' : 2,
'fear' : 3,
'joy' : 4,
'love' : 5,
'optimism' : 6,
'pessimism' : 7,
'sadness' : 8,
'surprise' : 9,
'trust' : 10,
}
def __init__(self, config, phase):
self.config = config
n_labels = len(self.label_index_map.keys())
filepath = os.path.join(config.dataroot, f'{phase}.txt')
self.texts = []
self.labels = torch.empty(0)
with io.open(filepath, encoding='utf-8') as f:
reader = csv.reader(f, delimiter='\t')
for i, row in enumerate(reader):
if i == 0:
continue
self.texts.append(row[1])
labels = torch.zeros(1, n_labels)
for i in self.label_index_map.values():
column_index = i + 2
labels[0][i] = int(row[column_index])
self.labels = torch.cat([self.labels, labels])
def __getitem__(self, index):
return self.texts[index], self.labels[index]
def __len__(self):
return len(self.texts)
class CustomClassificationHead(nn.Module):
def __init__(self, config, input_dim):
super().__init__()
self.config = config
self.fc1 = nn.Linear(input_dim, 4096)
self.fc2 = nn.Linear(4096, 2048)
self.fc3 = nn.Linear(2048, 1024)
self.fc4 = nn.Linear(1024, config.n_labels)
self.dropout = nn.Dropout(p=0.3)
self.prelu1 = nn.PReLU()
self.prelu2 = nn.PReLU()
self.prelu3 = nn.PReLU()
nn.init.kaiming_normal_(self.fc1.weight)
nn.init.kaiming_normal_(self.fc2.weight)
nn.init.kaiming_normal_(self.fc3.weight)
nn.init.kaiming_normal_(self.fc4.weight)
def forward(self, x):
# dropout is applied before this method is called
# https://github.com/huggingface/transformers/blob/v4.1.1/src/transformers/models/bert/modeling_bert.py#L1380
x = self.prelu1(self.fc1(x))
x = self.prelu2(self.fc2(self.dropout(x)))
x = self.prelu3(self.fc3(self.dropout(x)))
return self.fc4(self.dropout(x))
class Trainer:
def __init__(self, config):
self.config = config
model_name = 'cl-tohoku/bert-base-japanese-whole-word-masking' if config.lang == 'ja' else 'bert-base-uncased'
self.tokenizer = BertTokenizer.from_pretrained(model_name, padding=True)
self.model = self.__create_model(model_name)
self.optimizer = AdamW(self.model.parameters(), lr=config.lr)
self.warmup_scheduler = optim.lr_scheduler.LambdaLR(self.optimizer, lr_lambda=lambda step: min(1.0, (step + 1) / config.warmup_steps))
if not self.config.predict:
data_train= self.config.dataset_class(self.config, 'train')
self.dataloader_train = DataLoader(data_train, batch_size=self.config.batch_size, shuffle=True)
data_eval= self.config.dataset_class(self.config, 'eval')
self.dataloader_eval = DataLoader(data_eval, batch_size=self.config.batch_size, shuffle=False)
self.writer = SummaryWriter(log_dir=config.tensorboard_log_dir)
self.best_f1_score = 0.0
if config.fp16:
self.model, self.optimizer = apex.amp.initialize(self.model, self.optimizer, 'O1')
self.load(self.config.model_path)
if self.config.freeze_base:
for param in self.model.base_model.parameters():
param.requires_grad = False
def __create_model(self, model_name):
model = BertForSequenceClassification.from_pretrained(model_name, num_labels=self.config.n_labels, return_dict=True)
if self.config.freeze_base_model:
for param in model.base_model.parameters():
param.requires_grad = False
if self.config.custom_head:
model.classifier = CustomClassificationHead(self.config, model.config.hidden_size)
return model.to(self.config.device)
def forward(self, inputs, labels):
if self.config.multi_labels:
outputs = self.model(**inputs)
loss = F.binary_cross_entropy_with_logits(outputs.logits, labels)
return loss, 0 < outputs.logits
outputs = self.model(**inputs, labels=torch.argmax(labels, dim=1))
return outputs.loss, torch.argmax(outputs.logits, dim=1)
def train(self, epoch):
self.model.train()
for i, (texts, labels) in enumerate(self.dataloader_train):
start_time = time.time()
inputs = self.tokenizer(texts, return_tensors='pt', padding=True).to(self.config.device)
labels = labels.to(self.config.device)
loss, _ = self.forward(inputs, labels)
if self.config.fp16:
with apex.amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
self.warmup_scheduler.step()
if i % self.config.log_interval == 0:
elapsed_time = time.time() - start_time
self.config.logger.info('train epoch: {}, step: {}, loss: {:.2f}, time: {:.2f}, lr: {:.2e}'.format(epoch, i, loss, elapsed_time, self.warmup_scheduler.get_lr()[0]))
self.writer.add_scalar('loss/train', loss, epoch, start_time)
self.save(self.config.model_path)
@torch.no_grad()
def eval(self, epoch):
self.model.eval()
all_labels = torch.empty(0)
all_preds = torch.empty(0)
losses = []
start_time = time.time()
for i, (texts, labels) in enumerate(self.dataloader_eval):
inputs = self.tokenizer(texts, return_tensors='pt', padding=True).to(self.config.device)
labels = labels.to(self.config.device)
loss, preds = self.forward(inputs, labels)
losses.append(loss)
if not self.config.multi_labels:
labels = torch.argmax(labels, dim=1)
all_labels = torch.cat([all_labels, labels.cpu()])
all_preds = torch.cat([all_preds, preds.cpu()])
elapsed_time = time.time() - start_time
average_loss = sum(losses)/len(losses)
self.config.logger.info('eval epoch: {}, loss: {:.2f}, time: {:.2f}'.format(epoch, average_loss, elapsed_time))
self.__log_confusion_matrix(all_preds, all_labels, epoch)
columns = self.config.dataset_class.label_index_map.keys()
df = pd.DataFrame(metrics.classification_report(all_labels, all_preds, output_dict=True))
print(tabulate(df, headers='keys', tablefmt="github", floatfmt='.3f'))
if not self.config.eval_only:
f1_score = df.loc['f1-score']
micro = f1_score['micro avg'] if 'micro avg' in f1_score else f1_score['accuracy']
macro = f1_score['macro avg']
self.writer.add_scalar('loss/eval', average_loss, epoch, start_time)
self.writer.add_scalar('metrics/f1_score_micro(accuracy)', micro, epoch, start_time)
self.writer.add_scalar('metrics/f1_score_macro', macro, epoch, start_time)
if self.best_f1_score < micro:
self.best_f1_score = micro
self.save(self.config.best_model_path)
@torch.no_grad()
def predict(self):
label_map = {value: key for key, value in self.config.dataset_class.label_index_map.items()}
label_map[-1] = 'none'
np.set_printoptions(precision=0)
dataset = self.config.dataset_class(self.config, 'predict')
loader = DataLoader(dataset, batch_size=self.config.batch_size)
output_path = os.path.join(self.config.dataroot, 'predict_result')
with open(output_path, 'w') as f:
for i, (texts, labels) in enumerate(loader):
inputs = self.tokenizer(texts, return_tensors='pt', padding=True).to(self.config.device)
outputs = self.model(**inputs)
preds = torch.argmax(outputs.logits, dim=1)
probs = F.softmax(outputs.logits, dim=1) * 100
for j in range(len(texts)):
pred_label_name = label_map[preds[j].item()]
true_label_name = label_map[labels[j].item()]
prob = probs[j].cpu().numpy()
f.write(f'{pred_label_name}\t{prob}\t{true_label_name}\t{texts[j]}\n')
def __log_confusion_matrix(self, all_preds, all_labels, epoch):
buf = io.BytesIO()
label_map = {value: key for key, value in self.config.dataset_class.label_index_map.items()}
np.set_printoptions(precision=3)
if self.config.multi_labels:
fig, axes = plt.subplots(1, len(label_map.keys()), figsize=(25, 5))
cm = metrics.multilabel_confusion_matrix(y_pred=all_preds.numpy(), y_true=all_labels.numpy())
for i in range(len(label_map.keys())):
mat = np.array([[cm[i][1][1], cm[i][1][0]], [cm[i][0][1], cm[i][0][0]]])
result = mat / mat.sum(axis=1, keepdims=True)
print(f'{label_map[i]}\n{result}\n')
display = metrics.ConfusionMatrixDisplay(result, display_labels=['P', 'N'])
display.plot(ax=axes[i], cmap=plt.cm.Blues, values_format='.2f')
display.ax_.set_title(label_map[i])
display.ax_.set_ylabel('True label' if i == 0 else '')
display.ax_.set_yticklabels(['P', 'N'] if i == 0 else [])
display.im_.colorbar.remove()
plt.subplots_adjust(wspace=0.1, hspace=0.1)
fig.colorbar(display.im_, ax=axes)
plt.savefig(buf, format="png", dpi=180)
else:
cm = metrics.confusion_matrix(y_pred=all_preds.numpy(), y_true=all_labels.numpy(), normalize='true')
display = metrics.ConfusionMatrixDisplay(cm, display_labels=label_map.values())
display.plot(cmap=plt.cm.Blues)
display.figure_.savefig(buf, format="png", dpi=180)
cm = pycm.ConfusionMatrix(actual_vector=all_labels.numpy(), predict_vector=all_preds.numpy())
cm.relabel(mapping=label_map)
cm.print_normalized_matrix()
buf.seek(0)
img_arr = np.frombuffer(buf.getvalue(), dtype=np.uint8)
buf.close()
img = cv2.imdecode(img_arr, 1)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
self.writer.add_image('confusion_maatrix', img, epoch, dataformats='HWC')
def save(self, model_path):
if self.config.no_save:
return
data = {
'model': self.model.state_dict(),
'optimizer': self.optimizer.state_dict(),
'amp': apex.amp.state_dict() if self.config.fp16 else None,
'batch_size': self.config.batch_size,
'fp16': self.config.fp16,
}
torch.save(data, model_path)
self.config.logger.info(f'save model to {model_path}')
def load(self, model_path):
if not os.path.isfile(model_path):
return
data = torch.load(model_path, map_location=self.config.device_name)
self.model.load_state_dict(data['model'])
self.optimizer.load_state_dict(data['optimizer'])
if self.config.fp16:
apex.amp.load_state_dict(data['amp'])
self.config.logger.info(f'load model from {model_path}')
if __name__ == '__main__':
parser = argparse.ArgumentParser(add_help=True)
parser.add_argument('--cpu', action='store_true', help='use cpu')
parser.add_argument('--loglevel', default='DEBUG')
parser.add_argument('--log_interval', type=int, default=1)
parser.add_argument('--eval_interval', type=int, default=1)
parser.add_argument('--n_labels', type=int, default=6, help='number of classes to train')
parser.add_argument('--dataroot', default='data', help='path to data directory')
parser.add_argument('--batch_size', type=int, default=64, help='size of batch')
parser.add_argument('--epochs', type=int, default=10, help='epoch count')
parser.add_argument('--fp16', action='store_true', help='run model with float16')
parser.add_argument('--lang', default='ja', choices=['en', 'ja'])
parser.add_argument('--eval_only', action='store_true')
parser.add_argument('--predict', action='store_true')
parser.add_argument('--no_save', action='store_true')
parser.add_argument('--name', default=None)
parser.add_argument('--freeze_base', action='store_true')
parser.add_argument('--lr', type=float, default=1e-5)
parser.add_argument('--warmup_steps', type=int, default=1000)
parser.add_argument('--multi_labels', action='store_true')
parser.add_argument('--dataset_class_name', default='EmotionDataset', choices=['EmotionDataset', 'SemEval2018EmotionDataset'])
parser.add_argument('--custom_head', action='store_true')
parser.add_argument('--freeze_base_model', action='store_true')
args = parser.parse_args()
pd.options.display.precision = 3
pd.options.display.max_columns = 30
is_cpu = args.cpu or not torch.cuda.is_available()
args.device_name = "cpu" if is_cpu else "cuda:0"
args.device = torch.device(args.device_name)
logger = setup_logger(name=__name__, level=args.loglevel)
logger.info(args)
args.logger = logger
if args.name is None:
args.name = str(uuid.uuid4())[:8]
args.tensorboard_log_dir = f'{args.dataroot}/runs/{args.name}'
args.model_path = f'{args.dataroot}/{args.name}.pth'
args.best_model_path = f'{args.dataroot}/{args.name}.best.pth'
args.dataset_class = globals()[args.dataset_class_name]
if args.dataset_class_name == 'SemEval2018EmotionDataset':
args.n_labels = 11
args.multi_labels = True
trainer = Trainer(args)
if args.eval_only:
trainer.eval(0)
sys.exit()
if args.predict:
trainer.predict()
sys.exit()
for epoch in range(args.epochs):
trainer.train(epoch)
if epoch % args.eval_interval == 0:
trainer.eval(epoch)
trainer.eval(epoch)