-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
256 lines (205 loc) · 10.6 KB
/
train.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
from load import LoadData
from preprocess import PreprocessData
from model import BidaF
from evaluate import evaluate as eval
from load import ptbtokenizer
import torch.optim as optim
import torch.nn as nn
import torch
import pickle
import time
import sys
import argparse
import copy
import numpy as np
import datetime
import os
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def str2bool(v):
if isinstance(v, bool):
return v
if v.lower() in ('yes', 'true', 't', 'y', '1'):
return True
elif v.lower() in ('no', 'false', 'f', 'n', '0'):
return False
else:
raise argparse.ArgumentTypeError('Boolean value expected.')
def create_checkpoint_paths(folder_path):
if not os.path.isdir(folder_path):
os.mkdir(folder_path)
def save_vocab(vocab, path):
output = open(path, 'wb')
pickle.dump(vocab, output)
output.close()
def predict_data(context_tokenized, question_tokenized, word_vocab, char_vocab):
longest_context_word = max([len(w) for w in context_tokenized])
longest_question_word = max([len(w) for w in question_tokenized])
context_words = (torch.tensor([[word_vocab[word.lower()] for word in context_tokenized]]),
torch.tensor([len(context_tokenized)]))
question_words = (torch.tensor([[word_vocab[word.lower()] for word in question_tokenized]]),
torch.tensor([len(question_tokenized)]))
context_char = []
for word in context_tokenized:
_context_word = []
for c_index in range(longest_context_word):
if c_index < len(word):
_context_word.append(char_vocab[word[c_index]])
else:
_context_word.append(char_vocab['<pad>'])
context_char.append(_context_word)
context_char = torch.tensor([context_char])
question_char = []
for word in question_tokenized:
_question_word = []
for c_index in range(longest_question_word):
if c_index < len(word):
_question_word.append(char_vocab[word[c_index]])
else:
_question_word.append(char_vocab['<pad>'])
question_char.append(_question_word)
question_char = torch.tensor([question_char])
predict_data.context_words = context_words
predict_data.question_words = question_words
predict_data.context_char = context_char
predict_data.question_char = question_char
predict_data.batch_size = 1
def get_prediction(context, question, model, word_vocab, char_vocab):
model.eval()
context_tokenized = ptbtokenizer(context, context=True)
question_tokenized = ptbtokenizer(question)
predict_data(context_tokenized, question_tokenized, word_vocab, char_vocab)
p1, p2 = model(predict_data)
answer = " ".join(context_tokenized[p1.argmax(1): p2.argmax(1)])
return answer, p1.argmax(1), p2.argmax(1)
def train(model, optimizer, criterion, path, epochs, epochs_log, model_name):
start_time = time.time()
for epoch in range(epochs):
# SETTING MODEL IN TRAINING MODE
model.train()
epoch_loss = 0
batch_num = 0.0
exmaples_count = 0.0
best_dev_acc_exact = -1.0
best_dev_acc_f1 = -1.0
for train_data in iter(data_preprocessor.train_iter):
batch_num += 1.0
exmaples_count += train_data.batch_size
p1, p2 = model(train_data)
optimizer.zero_grad()
try:
batch_loss = criterion(p1, train_data.start_idx) + criterion(p2, train_data.end_idx)
except Exception as e:
print(e)
return (p1, p2, train_data)
epoch_loss += batch_loss.item()
batch_loss.backward()
optimizer.step()
time_delta = datetime.timedelta(seconds=np.round(time.time() - start_time, 0))
sys.stdout.write(f'\rEpoch:{epoch} | Batch:{batch_num} | Time Running: {time_delta}')
break
if epoch % epochs_log == 0:
train_loss = epoch_loss/(exmaples_count)
dev_accuracy, dev_loss = eval(data_preprocessor.dev_iter,
model,
criterion,
data_preprocessor.WORDS.vocab,
calculate_loss=True,
calculate_accuracy=True)
dev_accuracy_exact = dev_accuracy.groupby('id')['Exact'].max().mean()
dev_accuracy_f1 = dev_accuracy.groupby('id')['F1'].max().mean()
train_accuracy, _ = eval(data_preprocessor.train_iter,
model,
criterion,
data_preprocessor.WORDS.vocab,
calculate_loss=False,
calculate_accuracy=True)
train_accuracy_exact = train_accuracy.groupby('id')['Exact'].max().mean()
train_accuracy_f1 = train_accuracy.groupby('id')['F1'].max().mean()
print(
f'\nTrain Loss:{train_loss:.4f} Train Acc Exact:{train_accuracy_exact:.4f} Train Acc F1:{train_accuracy_f1:.4f}')
print(
f'Validation Loss :{dev_loss:.4f} Dev Acc Exact:{dev_accuracy_exact:.4f} Dev Acc F1:{dev_accuracy_f1:.4f}')
print('Test Prediction Results')
predict_context = "He was speaking after figures showed that the country's economy shrank by 20.4% in April - " \
"the largest monthly contraction on record - as the country spent its first full month in lockdown."
predict_ques = "By how much did the country's economy shrank"
print(get_prediction(predict_context,
predict_ques,
model,
data_preprocessor.WORDS.vocab,
data_preprocessor.CHAR.vocab))
torch.save({
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': epoch_loss,
}, path + '/' + model_name + '.torch')
if dev_accuracy_f1 > best_dev_acc_f1:
best_dev_acc_f1 = dev_accuracy_f1
best_dev_acc_exact = dev_accuracy_exact
torch.save(model, path + '/' + 'best_' + model_name + '.torch')
print (f'Best Validation Results '
f'Dev Acc Exact:{best_dev_acc_exact:.4f} '
f'Dev Acc F1:{best_dev_acc_f1:.4f}')
parser = argparse.ArgumentParser(description='BiDaF for Machine Comprehension & Cloze-Style Reading Comprehension - Training')
# Input data
parser.add_argument('-data', type=str, default='../../../../data/squad v1.1/', help='Path to input data')
# Checkpoints paths
parser.add_argument('-data_checkpoint', type=str, default='./data_checkpoints', help='Path to store preprocessed data checkpoints')
parser.add_argument('-model_checkpoint', type=str, default='./model_checkpoints', help='Path to store modelled data checkpoints')
parser.add_argument('-model_name', type=str, default=None, required=True, help='provide a name to the model for storing at chekpoints')
parser.add_argument('-dataset_name', type=str, default=None, required=True, help='Name of the Dataset')
parser.add_argument('-load_data', type=str2bool, nargs='?', default=False, help='To Load data of use preprocessed data')
# Modelling parameters
parser.add_argument('-epochs', type=int, default=20, help='No. of Epoch to run')
parser.add_argument('-batch_size', type=int, default=60, help='Number of examples in each batch')
parser.add_argument('-glove_size', type=int, default=100, help='Size of Glove vector to use')
parser.add_argument('-char_embedding_size', type=int, default=100, help='Size of Character embeddings to be used')
parser.add_argument('-kernel_size', type=int, default=5, help='Kernel Size')
parser.add_argument('-channels_count', type=int, default=100 ,help='Count of channels for character embeddings')
parser.add_argument('-learning_rate', type=float, default=0.5, help='Learning Rate')
parser.add_argument('-epoch_log', type=int, default=2, help='Print logs after xx epochs')
args = parser.parse_args()
if args.load_data:
_ = LoadData(data_path=args.data,
checkpoint_path=args.data_checkpoint,
train_file=args.dataset_name + '_' + 'train.csv',
dev_file=args.dataset_name + '_' + 'dev.csv')
# Create Checkpoint folders for in between sessions storage
create_checkpoint_paths(args.data_checkpoint)
create_checkpoint_paths(args.model_checkpoint)
data_preprocessor = PreprocessData(data_path=args.data_checkpoint,
glove_size=args.glove_size,
batch_size=args.batch_size,
train_file='train1.csv',
dev_file='dev1.csv')
# train_file=args.dataset_name + '_' + 'train.csv',
# dev_file=args.dataset_name + '_' + 'dev.csv')
save_vocab(data_preprocessor.WORDS.vocab, args.model_checkpoint + '/' + args.dataset_name + '_' + 'WORDS.vocab')
save_vocab(data_preprocessor.CHAR.vocab, args.model_checkpoint + '/' + args.dataset_name + '_' + 'CHAR.vocab')
# Initializing Bidaf Model
model = BidaF(data_preprocessor.WORDS,
data_preprocessor.CHAR,
char_embedding_size=args.char_embedding_size,
char_conv_kernel_size=args.kernel_size,
char_conv_channels_count=args.channels_count).to(device)
# Intialize Optimizer
# optimizer = optim.Adam(model.parameters(), lr=0.001)
optimizer = optim.Adadelta(model.parameters(), lr=args.learning_rate, rho=0.999)
criterion = nn.CrossEntropyLoss()
## This is a piece of code for retraining the model from where left
# if os.path.isfile(args.model_checkpoint + '/model.torch'):
# checkpoint = torch.load(args.model_checkpoint + '/model.torch')
# model.load_state_dict(checkpoint['model_state_dict'])
# optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
# epoch = checkpoint['epoch']
# loss = checkpoint['loss']
# print('Loaded Trained Model')
# print(f'Trained for {epoch} Epochs, Achieved {loss} Training Loss')
_error = train(model=model,
optimizer=optimizer,
criterion=criterion,
path=args.model_checkpoint,
epochs=args.epochs,
epochs_log=args.epoch_log,
model_name=args.dataset_name + '_' + args.model_name)