/
train_model.py
223 lines (167 loc) · 8.59 KB
/
train_model.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
import datasets.utils.logging
from segmenter import arguments, vocab, utils
from segmenter.huggingface_dataset import get_datasets
from segmenter.utils import load_text_model
from segmenter.models.segmenter_model import FeatureExtractorSegmenterModel, HuggingFaceSegmenterModel
import argparse
import torch
import torch.utils.data as data
import torch.optim as optim
import time
import os
import math
import numpy as np
import random
import logging
from sklearn.metrics import accuracy_score, precision_recall_fscore_support, classification_report
from transformers import default_data_collator, DataCollatorWithPadding
logger = logging.getLogger(__name__)
def save_model(model, args, vocabulary, model_str):
if not os.path.exists(args.output_folder):
os.makedirs(args.output_folder)
torch.save({
'model_state_dict': model.state_dict(),
'vocabulary': vocabulary,
'args': args
}, args.output_folder + "/model." + model_str + ".pt", pickle_protocol=4)
def eval_model(dev_dataloader, model, epoch, update):
with torch.no_grad():
predicted_l = []
true_l = []
epoch_cost = 0
model.eval()
for batch in dev_dataloader:
y = batch["labels"].to(device)
model_output = model.forward(batch, device)
results = model.get_sentence_prediction(model_output)
yhat = torch.argmax(results, dim=1)
predicted_l.extend(yhat.detach().cpu().numpy().tolist())
true_l.extend(y.detach().cpu().numpy().tolist())
cost = loss(results, y)
epoch_cost += cost.detach().cpu().numpy()
logger.info(f"Epoch {epoch}, update {update}, dev cost/batch: {epoch_cost / len(dev_dataloader)}")
logger.info(f"Dev Accuracy:{accuracy_score(true_l, predicted_l)}")
precision, recall, f1, _ = precision_recall_fscore_support(true_l, predicted_l, average='macro')
logger.info(f"Dev precision, Recall, F1 (Macro): {precision} {recall} {f1}")
logger.info(classification_report(true_l, predicted_l))
return f1
if __name__ == "__main__":
start_prep = time.time()
logging.basicConfig(level=logging.INFO)
datasets.utils.disable_progress_bar()
parser = argparse.ArgumentParser()
arguments.add_train_arguments(parser)
arguments.add_general_arguments(parser)
known_args, unknown_args = parser.parse_known_args()
model_arguments_parser = argparse.ArgumentParser()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
os.environ['PYTHONHASHSEED'] = str(known_args.seed)
random.seed(known_args.seed)
np.random.seed(known_args.seed)
torch.manual_seed(known_args.seed)
dataset_workers = 2
vocabulary = vocab.VocabDictionary()
vocabulary.create_from_count_file(path=known_args.vocabulary,
vocab_max_size=known_args.vocabulary_max_size,
word_min_frequency=known_args.vocabulary_min_frequency)
model_class, requires_vocab = utils.model_picker(known_args)
model_class.add_model_args(model_arguments_parser)
model_specific_args = model_arguments_parser.parse_args(unknown_args)
args = argparse.Namespace(**vars(known_args), **vars(model_specific_args))
if hasattr(args, "frozen_text_model_path"):
text_model, text_vocab, saved_model_args = load_text_model(args.frozen_text_model_path)
text_model = text_model.to(device)
assert isinstance(text_model, FeatureExtractorSegmenterModel)
model = model_class(args, text_model, saved_model_args.rnn_layer_size).to(device)
# overwrite the vocabulary so that it uses the one from the stored model
vocabulary = text_vocab
else:
if requires_vocab:
model = model_class(args, vocabulary).to(device)
else:
model = model_class(args).to(device)
if args.optimizer == "adam":
optimizer = optim.Adam(model.parameters(), lr=args.lr, betas=(args.adam_b1, args.adam_b2), eps=args.adam_eps)
else:
optimizer = optim.SGD(model.parameters(), lr=args.lr)
if args.lr_schedule == "reduce_on_plateau":
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', factor=args.lr_reduce_factor,
patience=args.lr_reduce_patience, verbose=True)
loss = torch.nn.CrossEntropyLoss()
best_result = -math.inf
best_epoch = -math.inf
if hasattr(args, "train_audio_features_corpus") and hasattr(args, "dev_audio_features_corpus"):
hf_datasets = get_datasets(train_text_file=args.train_corpus, dev_text_file=args.dev_corpus,
temperature=args.sampling_temperature,
train_audio_features_file=args.train_audio_features_corpus,
dev_audio_features_file=args.dev_audio_features_corpus, num_classes=args.n_classes)
else:
hf_datasets = get_datasets(train_text_file=args.train_corpus, dev_text_file=args.dev_corpus,
temperature=args.sampling_temperature, num_classes=args.n_classes)
if isinstance(model, HuggingFaceSegmenterModel):
hf_datasets = hf_datasets.map(model.apply_tokenizer).remove_columns(column_names="words")
collater = DataCollatorWithPadding(tokenizer=model.tokenizer)
else:
def text_to_idx(sample):
return {"idx": [vocabulary.get_index(token) for token in sample["words"].split()]}
hf_datasets = hf_datasets.map(text_to_idx)
collater = default_data_collator
train_dataset = hf_datasets["train"]
dev_dataset = hf_datasets["dev"]
train_dataloader = data.DataLoader(train_dataset, num_workers=dataset_workers, batch_size=args.batch_size,
shuffle=True, collate_fn=collater)
dev_dataloader = data.DataLoader(dev_dataset, num_workers=dataset_workers, batch_size=args.batch_size,
shuffle=False, drop_last=False, collate_fn=collater)
end_prep = time.time()
logger.info(f"Training preparation took {str(end_prep - start_prep)} seconds.")
if args.amp:
if not torch.cuda.is_available():
raise Exception
scaler = torch.cuda.amp.GradScaler()
update = 0
for epoch in range(1, args.epochs + 1):
last_time = time.time()
epoch_cost = 0
optimizer.zero_grad()
model.train()
for batch in train_dataloader:
with torch.cuda.amp.autocast(enabled=args.amp):
model_output = model.forward(batch, device)
results = model.get_sentence_prediction(model_output)
# Scale loss by gradient accumulation steps, so that the mean is properly computed over the virtual
# batch.
cost = loss(results, batch["labels"].to(device)) / args.gradient_accumulation
if args.amp:
scaler.scale(cost).backward()
else:
cost.backward()
epoch_cost += cost.detach().cpu().numpy()
update += 1
if update % args.gradient_accumulation == 0:
if args.amp:
scaler.step(optimizer)
scaler.update()
else:
optimizer.step()
optimizer.zero_grad()
if update % args.log_every == 0:
curr_time = time.time()
diff_t = curr_time - last_time
last_time = curr_time
logger.debug(f"Epoch {epoch} update {update} cost: {cost.detach().cpu().numpy()}, batches per second: "
f"{str(float(args.log_every) / float(diff_t))}")
if args.checkpoint_every_n_updates is not None and (update + 1) % args.checkpoint_every_n_updates == 0:
_ = eval_model(dev_dataloader, model, epoch, update)
save_model(model, args, vocabulary, str(epoch) + "." + str(update))
logger.info(f"Epoch {epoch}, train cost/batch: {epoch_cost / len(train_dataloader)}")
f1 = eval_model(dev_dataloader, model, epoch, update)
if args.lr_schedule == "reduce_on_plateau":
scheduler.step(f1)
if f1 > best_result:
best_result = f1
best_epoch = epoch
save_model(model, args, vocabulary, "best")
if args.checkpoint_interval > 0 and epoch % args.checkpoint_interval == 0:
save_model(model, args, vocabulary, str(epoch))
logger.info("Training finished.")
logger.info(f"Best checkpoint F1: {best_result}, achieved at epoch: {best_epoch}")