-
Notifications
You must be signed in to change notification settings - Fork 10
/
train.py
204 lines (159 loc) · 5.9 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
import pandas as pd
import numpy as np
import torch
import os
import glob
from datetime import datetime, timedelta
from tqdm import tqdm
from sklearn.metrics import roc_auc_score, accuracy_score, mean_squared_error
if torch.cuda.is_available():
torch.set_default_tensor_type(torch.cuda.FloatTensor)
def model_train(
fold,
model,
accelerator,
opt,
train_loader,
valid_loader,
test_loader,
config,
n_gpu,
early_stop=True,
):
train_losses = []
avg_train_losses = []
best_valid_auc = 0
logs_df = pd.DataFrame()
num_epochs = config["train_config"]["num_epochs"]
model_name = config["model_name"]
data_name = config["data_name"]
train_config = config["train_config"]
log_path = train_config["log_path"]
now = (datetime.now() + timedelta(hours=9)).strftime("%Y%m%d-%H%M%S") # KST time
token_cnts = 0
label_sums = 0
for i in range(1, num_epochs + 1):
for batch in tqdm(train_loader):
opt.zero_grad()
model.train()
out_dict = model(batch)
if n_gpu > 1:
loss, token_cnt, label_sum = model.module.loss(batch, out_dict)
else:
loss, token_cnt, label_sum = model.loss(batch, out_dict)
accelerator.backward(loss)
token_cnts += token_cnt
label_sums += label_sum
if train_config["max_grad_norm"] > 0.0:
torch.nn.utils.clip_grad_norm_(
model.parameters(), max_norm=train_config["max_grad_norm"]
)
opt.step()
train_losses.append(loss.item())
print("token_cnts", token_cnts, "label_sums", label_sums)
total_preds = []
total_trues = []
with torch.no_grad():
for batch in valid_loader:
model.eval()
out_dict = model(batch)
pred = out_dict["pred"].flatten()
true = out_dict["true"].flatten()
mask = true > -1
pred = pred[mask]
true = true[mask]
total_preds.append(pred)
total_trues.append(true)
total_preds = torch.cat(total_preds).squeeze(-1).detach().cpu().numpy()
total_trues = torch.cat(total_trues).squeeze(-1).detach().cpu().numpy()
train_loss = np.average(train_losses)
avg_train_losses.append(train_loss)
valid_auc = roc_auc_score(y_true=total_trues, y_score=total_preds)
path = os.path.join("saved_model", model_name, data_name)
if not os.path.isdir(path):
os.makedirs(path)
if valid_auc > best_valid_auc:
path = os.path.join(
os.path.join("saved_model", model_name, data_name), "params_*"
)
for _path in glob.glob(path):
os.remove(_path)
best_valid_auc = valid_auc
best_epoch = i
torch.save(
{"epoch": i, "model_state_dict": model.state_dict(),},
os.path.join(
os.path.join("saved_model", model_name, data_name),
"params_{}".format(str(best_epoch)),
),
)
if i - best_epoch > 10:
break
# clear lists to track next epochs
train_losses = []
valid_losses = []
total_preds, total_trues = [], []
# evaluation on test dataset
with torch.no_grad():
for batch in test_loader:
model.eval()
out_dict = model(batch)
pred = out_dict["pred"].flatten()
true = out_dict["true"].flatten()
mask = true > -1
pred = pred[mask]
true = true[mask]
total_preds.append(pred)
total_trues.append(true)
total_preds = torch.cat(total_preds).squeeze(-1).detach().cpu().numpy()
total_trues = torch.cat(total_trues).squeeze(-1).detach().cpu().numpy()
test_auc = roc_auc_score(y_true=total_trues, y_score=total_preds)
print(
"Fold {}:\t Epoch {}\t\tTRAIN LOSS: {:.5f}\tVALID AUC: {:.5f}\tTEST AUC: {:.5f}".format(
fold, i, train_loss, valid_auc, test_auc
)
)
checkpoint = torch.load(
os.path.join(
os.path.join("saved_model", model_name, data_name),
"params_{}".format(str(best_epoch)),
)
)
model.load_state_dict(checkpoint["model_state_dict"])
total_preds, total_trues = [], []
total_q_embeds, total_qr_embeds = [], []
# evaluation on test dataset
with torch.no_grad():
for batch in test_loader:
model.eval()
out_dict = model(batch)
pred = out_dict["pred"].flatten()
true = out_dict["true"].flatten()
mask = true > -1
pred = pred[mask]
true = true[mask]
total_preds.append(pred)
total_trues.append(true)
total_preds = torch.cat(total_preds).squeeze(-1).detach().cpu().numpy()
total_trues = torch.cat(total_trues).squeeze(-1).detach().cpu().numpy()
auc = roc_auc_score(y_true=total_trues, y_score=total_preds)
acc = accuracy_score(y_true=total_trues >= 0.5, y_pred=total_preds >= 0.5)
rmse = np.sqrt(mean_squared_error(y_true=total_trues, y_pred=total_preds))
print(
"Best Model\tTEST AUC: {:.5f}\tTEST ACC: {:5f}\tTEST RMSE: {:5f}".format(
auc, acc, rmse
)
)
logs_df = logs_df.append(
pd.DataFrame(
{"EarlyStopEpoch": best_epoch, "auc": auc, "acc": acc, "rmse": rmse},
index=[0],
),
sort=False,
)
log_out_path = os.path.join(log_path, data_name)
os.makedirs(log_out_path, exist_ok=True)
logs_df.to_csv(
os.path.join(log_out_path, "{}_{}.csv".format(model_name, now)), index=False
)
return auc, acc, rmse