/
main.py
295 lines (221 loc) · 11.5 KB
/
main.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
import numpy as np
import torch
import torch.optim as optim
import torch.nn as nn
from torch.utils.data import DataLoader
import Levenshtein as LS
from tqdm import tqdm
import logging
import os
import sys
import pandas as pd
import argparse
from dataloader import transform_letter_to_index, create_dictionaries, transform_index_to_letter, Speech2TextDataset, LanguageModelDataLoader, collate_train, collate
from models import Seq2Seq
LETTER_LIST = ['<pad>', 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L', 'M', 'N', 'O', 'P', 'Q', \
'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z', '-', "'", '.', '_', '+', ' ','<sos>','<eos>']
DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
def pre_train(model, train_loader, vali_loader, criterion, args, logger, letter2index, index2letter):
optimizer = optim.Adam(model.decoder.parameters(), lr=args.lr, weight_decay=5e-4)
model.train()
model.to(DEVICE)
train_loss_ = []
len_train_loader = train_loader.num_batch
for epoch in range(args.pre_trained_epochs):
# 1) Iterate through your loader
train_loss = 0.0
num_data = 0.0
for batch_id, (x, y) in tqdm(enumerate(train_loader), total=len_train_loader):
x = x.to(DEVICE)
y = y.to(DEVICE)
key = torch.zeros([x.size(0), 128]).to(DEVICE)
values = torch.zeros([x.size(0), x.size(1), 128]).to(DEVICE)
lengths = torch.zeros([x.size(0)]).to(DEVICE)
preds = model.decoder(key, values, lengths, text=y, isTrain=True, isAttended=False)[0]
loss = criterion(preds.view(-1, preds.size(2)), y[:,1:].contiguous().view(-1)).mean() # no <sos>
loss.backward()
optimizer.step()
train_loss += loss.item() * x.size(0)
num_data += x.size(0)
train_loss = train_loss / num_data
train_loss_.append(train_loss)
print('Epoch %i, train loss: %f'%(epoch+1, train_loss))
def vali_inference(model, vali_loader, letter2index, index2letter):
model.eval()
LS_dist = 0.0
num_data = 0.0
for batch_id, (x, x_length, y, y_length) in enumerate(vali_loader):
x = x.to(DEVICE)
x_length = x_length.to(DEVICE)
preds = model(x, x_length, isTrain=False)
preds_transcript = transform_index_to_letter(preds.argmax(-1).detach().cpu().numpy(), letter2index, index2letter)
y_transcript = transform_index_to_letter(y[:,1:].numpy(), letter2index, index2letter) # note y is: <sos> ... <eos>
for i in range(y.size(0)):
LS_dist += LS.distance(preds_transcript[i], y_transcript[i])
num_data += x.size(0)
LS_dist /= num_data
return LS_dist
def train(model, train_loader, vali_loader, criterion, args, logger, letter2index, index2letter):
optimizer = optim.Adam(model.parameters(), lr=args.lr, weight_decay=5e-4)
scheduler_p = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=args.pt_factor, patience=1, verbose=True)
model.train()
model.to(DEVICE)
train_loss_ = []
vali_loss_ = []
len_train_loader = len(train_loader)
for epoch in range(args.epochs):
# 1) Iterate through your loader
train_loss = 0.0
num_data = 0.0
for batch_id, (x, x_length, y, y_length) in tqdm(enumerate(train_loader), total=len_train_loader):
model.train()
# 2) Use torch.autograd.set_detect_anomaly(True) to get notices about gradient explosion
# ignore
# 3) Set the inputs to the device.
x = x.to(DEVICE)
x_length = x_length.to(DEVICE)
y = y.to(DEVICE)
y_length = y_length.to(DEVICE)
# 4) Pass your inputs, and length of speech into the model.
preds = model(x, x_length, y)[0]
# 5) Generate a mask based on the lengths of the text to create a masked loss.
# 5.1) Ensure the mask is on the device and is the correct shape.
mask = y[:, 1:] > 0 # note y is <sos> ... <eos>
# 6) If necessary, reshape your predictions and origianl text input
# 6.1) Use .contiguous() if you need to.
preds = preds.view(-1, preds.size(-1))
y = y[:, 1:].contiguous().view(-1) # note y is <sos> ... <eos>
# 7) Use the criterion to get the loss.
loss = criterion(preds, y) # no <sos>
# 8) Use the mask to calculate a masked loss.
masked_loss = (loss * mask.view(-1)).sum()
# 9) Run the backward pass on the masked loss.
masked_loss.backward()
# 10) Use torch.nn.utils.clip_grad_norm(model.parameters(), 2)
torch.nn.utils.clip_grad_norm(model.parameters(), 2)
# 11) Take a step with your optimizer
optimizer.step()
# 12) Normalize the masked loss
masked_loss = masked_loss / mask.sum()
train_loss += masked_loss.item() * x.size(0)
num_data += x.size(0)
# 13) Optionally print the training loss after every N batches
if batch_id > 1 and batch_id % (int(len_train_loader/2)+1) == 0:
vali_loss = vali_inference(model, vali_loader, letter2index, index2letter)
print('train loss: %f, vali loss: %f'%(train_loss/num_data, vali_loss))
train_loss = train_loss / num_data
train_loss_.append(train_loss)
vali_loss = vali_inference(model, vali_loader, letter2index, index2letter)
vali_loss_.append(vali_loss)
scheduler_p.step(vali_loss)
torch.save(model.state_dict(), os.path.join(args.save_path,args.model_name,"model_%i.t7"%epoch))
np.save(os.path.join(args.save_path, args.model_name, "train_loss_.npy"), train_loss_)
np.save(os.path.join(args.save_path, args.model_name, "vali_loss_.npy"), vali_loss_)
logger.info('Epoch %i, train loss: %f, vali loss: %f'%(epoch+1, train_loss, vali_loss))
###########################
# argument #
###########################
parser = argparse.ArgumentParser(description='Hw4p2')
parser.add_argument("--lr", type=float, default=1e-3, help='learning rate')
parser.add_argument("--epochs", type=int, default=50, help='training epochs')
parser.add_argument("--batch_size", type=int, default=100, help='batch size')
parser.add_argument("--pre_trained_epochs", type=int, default=20, help='pre-training epochs')
parser.add_argument("--pre_trained_batch_size", type=int, default=3000, help='pre-training batch size')
parser.add_argument("--batch_size_test", type=int, default=300, help='batch size for test')
parser.add_argument("--pt_factor", type=float, default=0.5, help='plateau scheduler reduction factor')
parser.add_argument("--name", type=str, default='hw4p2', help='model name')
parser.add_argument("--data_path", type=str, help='data path')
parser.add_argument("--save_path", type=str, help='where to save')
args = parser.parse_args()
###########################
# logger #
###########################
logger = logging.getLogger()
logger.setLevel(logging.INFO)
if not os.path.isdir(os.path.join(args.save_path, args.model_name)):
os.makedirs(os.path.join(args.save_path, args.model_name))
if os.path.isfile(os.path.join(args.save_path, args.model_name, 'logfile.log')):
os.remove(os.path.join(args.save_path, args.model_name, 'logfile.log'))
file_log_handler = logging.FileHandler(os.path.join(args.save_path, args.model_name, 'logfile.log'))
logger.addHandler(file_log_handler)
stdout_log_handler = logging.StreamHandler(sys.stdout)
logger.addHandler(stdout_log_handler)
attrs = vars(args)
for item in attrs.items():
logger.info("%s: %s"%item)
###########################
# Main #
###########################
# Data loading
letter2index, index2letter = create_dictionaries(LETTER_LIST)
speech_train = np.load(os.path.join(args.data_path, 'train_new.npy'), allow_pickle=True, encoding='bytes')
transcript_train = np.load(os.path.join(args.data_path, 'train_transcripts.npy'), allow_pickle=True,encoding='bytes')
speech_valid = np.load(os.path.join(args.data_path, 'dev_new.npy'), allow_pickle=True, encoding='bytes')
transcript_valid = np.load(os.path.join(args.data_path, 'dev_transcripts.npy'), allow_pickle=True,encoding='bytes')
speech_test = np.load(os.path.join(args.data_path, 'test_new.npy'), allow_pickle=True, encoding='bytes')
train_dataset = Speech2TextDataset(speech_train, text=transcript_train)
train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, collate_fn=collate_train, num_workers=os.cpu_count())
vali_dataset = Speech2TextDataset(speech_valid, text=transcript_valid)
vali_loader = DataLoader(vali_dataset, batch_size=args.batch_size, shuffle=False, collate_fn=collate_train, num_workers=os.cpu_count())
test_dataset = Speech2TextDataset(speech_test)
test_loader = DataLoader(test_dataset, batch_size=args.batch_size_test, shuffle=False, collate_fn=collate, num_workers=os.cpu_count())
pre_train_loader = LanguageModelDataLoader(transcript_train, length=50, batch_size=args.pre_trained_batch_size)
# Model Initialization
DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
model = Seq2Seq(40, len(LETTER_LIST), 256, value_size=256, key_size=256).to(DEVICE)
criterion = nn.CrossEntropyLoss(reduction='none').to(DEVICE)
# Pre-train
pre_train(model, pre_train_loader, pre_train_loader, criterion, args, logger, letter2index, index2letter)
# Train
train(model, train_loader, vali_loader, criterion, args, logger, letter2index, index2letter)
# Inference
# greedy search
result = []
result_probs = []
model.eval()
with torch.no_grad():
for batch_id, (x, x_length) in enumerate(test_loader):
x = x.to(DEVICE)
x_length = x_length.to(DEVICE)
preds, probs = model(x, x_length, isTrain=False, random=False)
preds_transcript = transform_index_to_letter(preds.argmax(-1).detach().cpu().numpy(), letter2index, index2letter)
result.extend(preds_transcript)
result_probs.append(probs)
result_probs = torch.cat(result_probs)
result_greedy = result
result_probs_greedy = result_probs
# random search
result = []
result_probs = []
model.eval()
with torch.no_grad():
for batch_id, (x, x_length) in enumerate(test_loader):
x = x.to(DEVICE)
x_length = x_length.to(DEVICE)
random_result = []
random_probs = []
for i in tqdm(range(1000)):
preds, probs = model(x, x_length, isTrain=False, random=True)
preds_transcript = transform_index_to_letter(preds.argmax(-1).detach().cpu().numpy(), letter2index, index2letter)
random_result.append(preds_transcript)
random_probs.append(probs)
random_probs = torch.stack(random_probs)
max_probs = random_probs.max(0)[0]
max_index = random_probs.argmax(0)
for i in range(len(max_index)):
result.append(random_result[max_index[i]][i])
result_probs.append(max_probs)
result_probs = torch.cat(result_probs)
result_random = result
result_probs_random = result_probs
# combine greedy search and random search
result = []
for i in range(len(result_random)):
if result_probs_random[i] >= result_probs_greedy[i]:
result.append(result_random[i])
else:
result.append(result_greedy[i])
# save result
f = pd.read_csv(os.path.join(args.data_path, 'test_sample_submission.csv'))
f['Predicted'] = result
f.to_csv(os.path.join(args.data_path, 'test_sample_submission.csv'), index=False)