/
framework.py
149 lines (124 loc) · 5.21 KB
/
framework.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
# -*- coding: utf-8 -*-
import time
import os
import pdb
import torch
from torch.nn import DataParallel
import numpy as np
class Framework:
def __init__(self,
train_dataset=None,
dev_dataset=None,
test_dataset=None,
metric=None,
device=None,
opt=None):
self.train_dataset = train_dataset
self.dev_dataset = dev_dataset
self.test_dataset = test_dataset
self.metric = metric
self.device = device
self.opt = opt
def to_device(self, inputs):
for k in inputs.keys():
inputs[k] = inputs[k].to(self.device)
return inputs
def save_model(self, model, save_ckpt):
checkpoint = {'opt': self.opt}
if isinstance(model, DataParallel):
checkpoint['state_dict'] = model.module.state_dict()
else:
checkpoint['state_dict'] = model.state_dict()
torch.save(checkpoint, save_ckpt)
def load_model(self, model, load_ckpt):
if os.path.isfile(load_ckpt):
checkpoint = torch.load(load_ckpt)
print(f"Successfully loaded checkpoint : {load_ckpt}")
else:
raise Exception(f"No checkpoint found at {load_ckpt}")
load_state = checkpoint['state_dict']
model_state = model.state_dict()
for name, param in load_state.items():
if name not in model_state:
continue
if param.shape != model_state[name].shape:
print(f"In load model : {name} param shape not match")
continue
model_state[name].copy_(param)
def evaluate(self,
model,
eval_epoch,
evalN, K, Q,
mode="dev",
load_ckpt=None):
if load_ckpt is not None:
print(f"loading checkpint {load_ckpt}")
self.load_model(model, load_ckpt)
print(f"checkpoint {load_ckpt} loaded")
self.metric.reset()
model.to(self.device)
if mode == "dev":
eval_dataset = self.dev_dataset
elif mode == "test":
eval_dataset = self.test_dataset
elif mode == "train":
eval_dataset = self.train_dataset
model.eval()
for i in range(eval_epoch):
support_set, query_set, id2label = next(eval_dataset)
support_set, query_set = self.to_device(support_set), self.to_device(query_set)
loss, logits, pred = model(support_set, query_set, evalN, K, Q)
self.metric.update_state(pred, query_set['trigger_label'], id2label)
return self.metric.result()
def train(self,
model,
trainN, evalN, K, Q,
optimizer,
scheduler,
train_epoch,
eval_epoch,
eval_step,
load_ckpt=None,
save_ckpt=None):
if load_ckpt is not None:
print(f"loading checkpint {load_ckpt}")
self.load_model(model, load_ckpt)
print(f"checkpoint {load_ckpt} loaded")
model.to(self.device)
best_p = 0
best_r = 0
best_f1 = 0
best_epoch = 0
for epoch in range(train_epoch):
epoch_begin = time.time()
# train
model.train()
support_set, query_set, id2label = next(self.train_dataset)
support_set, query_set = self.to_device(support_set), self.to_device(query_set)
loss, logits, pred = model(support_set, query_set, trainN, K, Q)
loss = loss.mean()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 10)
optimizer.step()
optimizer.zero_grad()
scheduler.step()
epoch_end = time.time()
epoch_time = epoch_end - epoch_begin
remain_time_s = epoch_time * ((train_epoch-epoch) + np.ceil((train_epoch-epoch) / eval_step) * eval_epoch)
remain_time_h = remain_time_s / 3600
print(f"Epoch : {epoch}, loss : {loss:.4f}, time : {epoch_time:.4f}s, remain time : {remain_time_s:.4f}s ({remain_time_h:.2f}h)", end="\r")
# evaluate
if (epoch+1) % eval_step == 0:
eval_time = time.time()
p, r, f1 = self.evaluate(model, eval_epoch, evalN, K, Q)
print()
print(f"Evaluate result of epoch {epoch} - eval time : {time.time()-eval_time:.4f}s, P : {p:.6f}, R : {r:.6f}, F1 : {f1:.6f}")
if f1 >= best_f1:
self.save_model(model, save_ckpt)
best_p = p
best_r = r
best_f1 = f1
best_epoch = epoch
print(f"New best performance in epoch {epoch} - P: {best_p:.6f}, R: {best_r:.6f}, F1: {best_f1:.6f}")
else:
print(f"Current best performance - P: {best_p:.6f}, R: {best_r:.6f}, F1: {best_f1:.6f} in epoch {best_epoch}")