-
Notifications
You must be signed in to change notification settings - Fork 12
/
train.py
269 lines (233 loc) · 12.3 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
257
258
259
260
261
262
263
264
265
266
267
268
269
# -*- coding: utf-8 -*-
import torch
import torch.nn as nn
from torch.optim.lr_scheduler import StepLR
from sklearn import metrics
import numpy as np
import argparse
import os
import sys
import time
import re
from .models import ModelBiLSTM
from .dataloader import SignalFeaData2
from .dataloader import clear_linecache
from .utils.process_utils import display_args
from .utils.process_utils import str2bool
from .utils.constants_torch import use_cuda
def train(args):
"""
:param args: train_sample_file, valid_sample_file, hyperparameters
:return: directory contains model_params_checkpoint_file
"""
total_start = time.time()
# torch.manual_seed(args.seed)
print("[main] train starts..")
if use_cuda:
print("GPU is available!")
else:
print("GPU is not available!")
print("reading data..")
train_dataset = SignalFeaData2(args.train_file)
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True)
valid_dataset = SignalFeaData2(args.valid_file)
valid_loader = torch.utils.data.DataLoader(dataset=valid_dataset,
batch_size=args.batch_size,
shuffle=False)
model_dir = args.model_dir
if model_dir != "/":
model_dir = os.path.abspath(model_dir).rstrip("/")
if not os.path.exists(model_dir):
os.makedirs(model_dir)
else:
model_regex = re.compile(r"" + args.model_type + "\.b\d+_s\d+_epoch\d+\.ckpt*")
for mfile in os.listdir(model_dir):
if model_regex.match(mfile):
os.remove(model_dir + "/" + mfile)
model_dir += "/"
model = ModelBiLSTM(args.seq_len, args.signal_len, args.layernum1, args.layernum2, args.class_num,
args.dropout_rate, args.hid_rnn,
args.n_vocab, args.n_embed, str2bool(args.is_base), str2bool(args.is_signallen),
args.model_type)
if use_cuda:
model = model.cuda()
if args.init_model is not None:
print("loading pre-trained model: {}".format(args.init_model))
para_dict = torch.load(args.init_model) if use_cuda else torch.load(args.init_model,
map_location=torch.device('cpu'))
model_dict = model.state_dict()
model_dict.update(para_dict)
model.load_state_dict(model_dict)
# Loss and optimizer
weight_rank = torch.from_numpy(np.array([1, args.pos_weight])).float()
if use_cuda:
weight_rank = weight_rank.cuda()
criterion = nn.CrossEntropyLoss(weight=weight_rank)
if args.optim_type == "Adam":
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
elif args.optim_type == "RMSprop":
optimizer = torch.optim.RMSprop(model.parameters(), lr=args.lr)
elif args.optim_type == "SGD":
optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=0.8)
elif args.optim_type == "Ranger":
# use Ranger optimizer
# refer to https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
# needs python>=3.6
try:
from .utils.ranger2020 import Ranger
except ImportError:
raise ImportError("please check if ranger2020.py is in utils/ dir!")
optimizer = Ranger(model.parameters(), lr=args.lr)
else:
raise ValueError("optim_type is not right!")
scheduler = StepLR(optimizer, step_size=args.lr_decay_step, gamma=args.lr_decay)
# Train the model
total_step = len(train_loader)
print("total_step: {}".format(total_step))
curr_best_accuracy = 0
model.train()
# train at most max_epoch_num epochs
for epoch in range(args.max_epoch_num):
curr_best_accuracy_epoch = 0
no_best_model = True
tlosses = []
start = time.time()
for i, sfeatures in enumerate(train_loader):
_, kmer, base_means, base_stds, base_signal_lens, signals, labels = sfeatures
if use_cuda:
kmer = kmer.cuda()
base_means = base_means.cuda()
base_stds = base_stds.cuda()
base_signal_lens = base_signal_lens.cuda()
signals = signals.cuda()
labels = labels.cuda()
# Forward pass
outputs, logits = model(kmer, base_means, base_stds, base_signal_lens, signals)
loss = criterion(outputs, labels)
tlosses.append(loss.detach().item())
# Backward and optimize
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5)
optimizer.step()
if (i + 1) % args.step_interval == 0 or (i + 1) == total_step:
model.eval()
with torch.no_grad():
vlosses, vlabels_total, vpredicted_total = [], [], []
for vi, vsfeatures in enumerate(valid_loader):
_, vkmer, vbase_means, vbase_stds, vbase_signal_lens, vsignals, vlabels = vsfeatures
if use_cuda:
vkmer = vkmer.cuda()
vbase_means = vbase_means.cuda()
vbase_stds = vbase_stds.cuda()
vbase_signal_lens = vbase_signal_lens.cuda()
vsignals = vsignals.cuda()
vlabels = vlabels.cuda()
voutputs, vlogits = model(vkmer, vbase_means, vbase_stds, vbase_signal_lens, vsignals)
vloss = criterion(voutputs, vlabels)
_, vpredicted = torch.max(vlogits.data, 1)
if use_cuda:
vlabels = vlabels.cpu()
vpredicted = vpredicted.cpu()
vlosses.append(vloss.item())
vlabels_total += vlabels.tolist()
vpredicted_total += vpredicted.tolist()
v_accuracy = metrics.accuracy_score(vlabels_total, vpredicted_total)
v_precision = metrics.precision_score(vlabels_total, vpredicted_total)
v_recall = metrics.recall_score(vlabels_total, vpredicted_total)
if v_accuracy > curr_best_accuracy_epoch:
curr_best_accuracy_epoch = v_accuracy
if curr_best_accuracy_epoch > curr_best_accuracy - 0.0002:
torch.save(model.state_dict(),
model_dir + args.model_type + '.b{}_s{}_epoch{}.ckpt'.format(args.seq_len,
args.signal_len,
epoch + 1))
if curr_best_accuracy_epoch > curr_best_accuracy:
curr_best_accuracy = curr_best_accuracy_epoch
no_best_model = False
time_cost = time.time() - start
print('Epoch [{}/{}], Step [{}/{}], TrainLoss: {:.4f}; '
'ValidLoss: {:.4f}, '
'Accuracy: {:.4f}, Precision: {:.4f}, Recall: {:.4f}, '
'curr_epoch_best_accuracy: {:.4f}; Time: {:.2f}s'
.format(epoch + 1, args.max_epoch_num, i + 1, total_step, np.mean(tlosses),
np.mean(vlosses), v_accuracy, v_precision, v_recall,
curr_best_accuracy_epoch, time_cost))
tlosses = []
start = time.time()
sys.stdout.flush()
model.train()
scheduler.step()
if no_best_model and epoch >= args.min_epoch_num - 1:
print("early stop!")
break
endtime = time.time()
clear_linecache()
print("[main] train costs {} seconds, "
"best accuracy: {}".format(endtime - total_start,
curr_best_accuracy))
def main():
parser = argparse.ArgumentParser("")
parser.add_argument('--train_file', type=str, required=True)
parser.add_argument('--valid_file', type=str, required=True)
parser.add_argument('--model_dir', type=str, required=True)
# model input
parser.add_argument('--model_type', type=str, default="both_bilstm",
choices=["both_bilstm", "seq_bilstm", "signal_bilstm"],
required=False,
help="type of model to use, 'both_bilstm', 'seq_bilstm' or 'signal_bilstm', "
"'both_bilstm' means to use both seq and signal bilstm, default: both_bilstm")
parser.add_argument('--seq_len', type=int, default=13, required=False,
help="len of kmer. default 13")
parser.add_argument('--signal_len', type=int, default=16, required=False,
help="the number of signals of one base to be used in deepsignal_plant, default 16")
# model param
parser.add_argument('--layernum1', type=int, default=3,
required=False, help="lstm layer num for combined feature, default 3")
parser.add_argument('--layernum2', type=int, default=1,
required=False, help="lstm layer num for seq feature (and for signal feature too), default 1")
parser.add_argument('--class_num', type=int, default=2, required=False)
parser.add_argument('--dropout_rate', type=float, default=0.5, required=False)
parser.add_argument('--n_vocab', type=int, default=16, required=False,
help="base_seq vocab_size (15 base kinds from iupac)")
parser.add_argument('--n_embed', type=int, default=4, required=False,
help="base_seq embedding_size")
parser.add_argument('--is_base', type=str, default="yes", required=False,
help="is using base features in seq model, default yes")
parser.add_argument('--is_signallen', type=str, default="yes", required=False,
help="is using signal length feature of each base in seq model, default yes")
# BiLSTM model param
parser.add_argument('--hid_rnn', type=int, default=256, required=False,
help="BiLSTM hidden_size for combined feature")
# model training
parser.add_argument('--optim_type', type=str, default="Adam", choices=["Adam", "RMSprop", "SGD",
"Ranger"],
required=False, help="type of optimizer to use, 'Adam' or 'SGD' or 'RMSprop' or 'Ranger', "
"default Adam")
parser.add_argument('--batch_size', type=int, default=512, required=False)
parser.add_argument('--lr', type=float, default=0.001, required=False)
parser.add_argument('--lr_decay', type=float, default=0.1, required=False)
parser.add_argument('--lr_decay_step', type=int, default=2, required=False)
parser.add_argument("--max_epoch_num", action="store", default=10, type=int,
required=False, help="max epoch num, default 10")
parser.add_argument("--min_epoch_num", action="store", default=5, type=int,
required=False, help="min epoch num, default 5")
parser.add_argument('--step_interval', type=int, default=100, required=False)
parser.add_argument('--pos_weight', type=float, default=1.0, required=False)
parser.add_argument('--init_model', type=str, default=None, required=False,
help="file path of pre-trained model parameters to load before training")
# parser.add_argument('--seed', type=int, default=1234,
# help='random seed')
# else
parser.add_argument('--tmpdir', type=str, default="/tmp", required=False)
args = parser.parse_args()
print("[main] start..")
total_start = time.time()
display_args(args)
train(args)
endtime = time.time()
print("[main] costs {} seconds".format(endtime - total_start))
if __name__ == '__main__':
main()