-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
224 lines (195 loc) · 5.98 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
import os
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from dataset import WMTCollator, WMTDataset
from language import Language
from network import Seq2SeqModel
from train_utils import train
from utils import set_global_seed
# path
INPUT_LANG_TRAIN_DATA_PATH = "data/IWSLT15_English_Vietnamese/train.en"
OUTPUT_LANG_TRAIN_DATA_PATH = "data/IWSLT15_English_Vietnamese/train.vi"
INPUT_LANG_VAL_DATA_PATH = "data/IWSLT15_English_Vietnamese/tst2012.en"
OUTPUT_LANG_VAL_DATA_PATH = "data/IWSLT15_English_Vietnamese/tst2012.vi"
INPUT_LANG_TEST_DATA_PATH = "data/IWSLT15_English_Vietnamese/tst2013.en"
OUTPUT_LANG_TEST_DATA_PATH = "data/IWSLT15_English_Vietnamese/tst2013.vi"
INPUT_LANG = "en"
OUTPUT_LANG = "vi"
INPUT_LANG_WORD2IDX_PATH = f"vocab/{INPUT_LANG}_vocab.json"
OUTPUT_LANG_WORD2IDX_PATH = f"vocab/{OUTPUT_LANG}_vocab.json"
SAVE_MODEL_PATH = "models/seq2seq.pth"
# hyper-parameters
SEED = 42
DEVICE = "cuda"
VERBOSE = True
UNK_ID = 0
BOS_ID = 1
EOS_ID = 2
PAD_ID = 3
BATCH_SIZE = 32
REVERSE_SOURCE_LANG = True
BUCKET_SEQUENCING_PERCENTILE = 100
ENCODER_EMBEDDING_DIM = DECODER_EMBEDDING_DIM = 500
ENCODER_HIDDEN_SIZE = DECODER_HIDDEN_SIZE = 500
ENCODER_NUM_LAYERS = DECODER_NUM_LAYERS = 2
ENCODER_DROPOUT = DECODER_DROPOUT = 0.2
N_EPOCH = 12
LEARNING_RATE = 1
TRAIN_EVAL_FREQ = 50 # number of batches
# print params
if VERBOSE:
print("### PARAMETERS ###")
print()
print(f"INPUT_LANG_TRAIN_DATA_PATH: {INPUT_LANG_TRAIN_DATA_PATH}")
print(f"OUTPUT_LANG_TRAIN_DATA_PATH: {OUTPUT_LANG_TRAIN_DATA_PATH}")
print(f"INPUT_LANG_VAL_DATA_PATH: {INPUT_LANG_VAL_DATA_PATH}")
print(f"OUTPUT_LANG_VAL_DATA_PATH: {OUTPUT_LANG_VAL_DATA_PATH}")
print(f"INPUT_LANG_TEST_DATA_PATH: {INPUT_LANG_TEST_DATA_PATH}")
print(f"OUTPUT_LANG_TEST_DATA_PATH: {OUTPUT_LANG_TEST_DATA_PATH}")
print()
print(f"INPUT_LANG: {INPUT_LANG}")
print(f"OUTPUT_LANG: {OUTPUT_LANG}")
print(f"INPUT_LANG_WORD2IDX_PATH: {INPUT_LANG_WORD2IDX_PATH}")
print(f"OUTPUT_LANG_WORD2IDX_PATH: {OUTPUT_LANG_WORD2IDX_PATH}")
print()
print(f"SAVE_MODEL_PATH: {SAVE_MODEL_PATH}")
print()
print(f"SEED: {SEED}")
print(f"DEVICE: {DEVICE}")
print()
print(f"UNK_ID: {UNK_ID}")
print(f"BOS_ID: {BOS_ID}")
print(f"EOS_ID: {EOS_ID}")
print(f"PAD_ID: {PAD_ID}")
print()
print(f"BATCH_SIZE: {BATCH_SIZE}")
print(f"REVERSE_SOURCE_LANG: {REVERSE_SOURCE_LANG}")
print(f"BUCKET_SEQUENCING_PERCENTILE: {BUCKET_SEQUENCING_PERCENTILE}")
print()
print(f"ENCODER/DECODER_EMBEDDING_DIM: {ENCODER_EMBEDDING_DIM}")
print(f"ENCODER/DECODER_HIDDEN_SIZE: {ENCODER_HIDDEN_SIZE}")
print(f"ENCODER/DECODER_NUM_LAYERS: {ENCODER_NUM_LAYERS}")
print(f"ENCODER/DECODER_DROPOUT: {ENCODER_DROPOUT}")
print()
print(f"N_EPOCH: {N_EPOCH}")
print(f"LEARNING_RATE: {LEARNING_RATE}")
print(f"TRAIN_EVAL_FREQ: {TRAIN_EVAL_FREQ}")
print()
# seed and device
set_global_seed(SEED)
device = torch.device(DEVICE)
# language
input_language = Language(
language=INPUT_LANG,
path_to_word2idx=INPUT_LANG_WORD2IDX_PATH,
unk_id=UNK_ID,
bos_id=BOS_ID,
eos_id=EOS_ID,
)
output_language = Language(
language=OUTPUT_LANG,
path_to_word2idx=OUTPUT_LANG_WORD2IDX_PATH,
unk_id=UNK_ID,
bos_id=BOS_ID,
eos_id=EOS_ID,
)
# data
train_dataset = WMTDataset(
input_lang_data_path=INPUT_LANG_TRAIN_DATA_PATH,
output_lang_data_path=OUTPUT_LANG_TRAIN_DATA_PATH,
input_language=input_language,
output_language=output_language,
reverse_source_lang=REVERSE_SOURCE_LANG,
verbose=VERBOSE,
)
val_dataset = WMTDataset(
input_lang_data_path=INPUT_LANG_VAL_DATA_PATH,
output_lang_data_path=OUTPUT_LANG_VAL_DATA_PATH,
input_language=input_language,
output_language=output_language,
reverse_source_lang=REVERSE_SOURCE_LANG,
verbose=VERBOSE,
)
test_dataset = WMTDataset(
input_lang_data_path=INPUT_LANG_TEST_DATA_PATH,
output_lang_data_path=OUTPUT_LANG_TEST_DATA_PATH,
input_language=input_language,
output_language=output_language,
reverse_source_lang=REVERSE_SOURCE_LANG,
verbose=VERBOSE,
)
train_collator = WMTCollator(
input_lang_pad_id=PAD_ID,
output_lang_pad_id=PAD_ID,
percentile=BUCKET_SEQUENCING_PERCENTILE,
)
test_collator = WMTCollator( # same for val_loader
input_lang_pad_id=PAD_ID,
output_lang_pad_id=PAD_ID,
percentile=100,
)
train_loader = DataLoader(
dataset=train_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
collate_fn=train_collator,
)
val_loader = DataLoader(
dataset=val_dataset,
batch_size=1,
shuffle=False,
collate_fn=test_collator,
)
test_loader = DataLoader(
dataset=test_dataset,
batch_size=1,
shuffle=False,
collate_fn=test_collator,
)
# model
model = Seq2SeqModel(
encoder_num_embeddings=len(input_language.word2idx),
encoder_embedding_dim=ENCODER_EMBEDDING_DIM,
encoder_hidden_size=ENCODER_HIDDEN_SIZE,
encoder_num_layers=ENCODER_NUM_LAYERS,
encoder_dropout=ENCODER_DROPOUT,
decoder_num_embeddings=len(output_language.word2idx),
decoder_embedding_dim=DECODER_EMBEDDING_DIM,
decoder_hidden_size=DECODER_HIDDEN_SIZE,
decoder_num_layers=DECODER_NUM_LAYERS,
decoder_dropout=DECODER_DROPOUT,
).to(device)
if VERBOSE:
print(f"model number of parameters: {sum(p.numel() for p in model.parameters())}")
# criterion, optimizer, scheduler
criterion = nn.CrossEntropyLoss(
ignore_index=PAD_ID,
)
optimizer = optim.SGD(
model.parameters(),
lr=LEARNING_RATE,
)
scheduler = optim.lr_scheduler.MultiStepLR( # hardcoded
optimizer,
milestones=[8, 9, 10, 11],
gamma=0.5,
)
# train
train(
model=model,
trainloader=train_loader,
valloader=val_loader,
testloader=test_loader,
criterion=criterion,
optimizer=optimizer,
scheduler=scheduler,
n_epoch=N_EPOCH,
train_eval_freq=TRAIN_EVAL_FREQ,
device=device,
verbose=VERBOSE,
)
# save
os.makedirs("models", exist_ok=True)
torch.save(model.state_dict(), SAVE_MODEL_PATH)