-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
189 lines (163 loc) · 6.94 KB
/
trainer.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# @File : trainer.py
# @Author:
# @Date : 2021/11/1 15:45
# @Desc :
import copy
import time
import os
import torch
import numpy as np
from loguru import logger
from torch import optim
from torch.utils.data import DataLoader
from tqdm import tqdm, trange
from data_set import DataSet
from metrics import metrics_dict
class Trainer(object):
def __init__(self, model, dataset: DataSet, args):
self.model = model
self.dataset = dataset
self.behaviors = args.behaviors
self.topk = args.topk
self.metrics = args.metrics
self.learning_rate = args.lr
self.weight_decay = args.decay
self.batch_size = args.batch_size
self.test_batch_size = args.test_batch_size
self.min_epoch = args.min_epoch
self.epochs = args.epochs
self.model_path = args.model_path
self.model_name = args.model_name
self.device = args.device
self.TIME = args.TIME
self.optimizer = self.get_optimizer(self.model)
def get_optimizer(self, model):
optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()),
lr=self.learning_rate,
weight_decay=self.weight_decay)
return optimizer
def clear_parameter(self, model):
# for device in model.device_ids:
# model.module.storage_user_embeddings = None
# model.module.storage_item_embeddings = None
model.storage_user_embeddings = None
model.storage_item_embeddings = None
@logger.catch()
def train_model(self):
train_dataset_loader = DataLoader(dataset=self.dataset.behavior_dataset(),
batch_size=self.batch_size,
shuffle=True)
# best_result = np.zeros(len(self.topk) * len(self.metrics))
best_result = 0
best_dict = {}
best_epoch = 0
best_model = None
final_test = None
for epoch in range(self.epochs):
self.model.train()
test_metric_dict, validate_metric_dict = self._train_one_epoch(train_dataset_loader, epoch)
if validate_metric_dict is not None:
result = validate_metric_dict['hit@20']
# early stop
if result - best_result > 0:
final_test = test_metric_dict
best_result = result
best_dict = validate_metric_dict
best_model = copy.deepcopy(self.model)
best_epoch = epoch
if epoch - best_epoch > 4:
break
# save the best model
self.save_model(best_model)
logger.info(f"training end, best iteration %d, results: %s" %
(best_epoch + 1, best_dict.__str__()))
logger.info(f"final test result is: %s" % final_test.__str__())
def _train_one_epoch(self, behavior_dataset_loader, epoch):
start_time = time.time()
behavior_dataset_iter = (
tqdm(
enumerate(behavior_dataset_loader),
total=len(behavior_dataset_loader),
desc=f"\033[1;35m Train {epoch + 1:>5}\033[0m"
)
)
total_loss = 0.0
batch_no = 0
for batch_index, batch_data in behavior_dataset_iter:
start = time.time()
batch_data = batch_data.to(self.device)
self.optimizer.zero_grad()
loss = self.model(batch_data)
# loss = loss.sum()
loss.backward()
self.optimizer.step()
batch_no = batch_index + 1
total_loss += loss.item()
total_loss = total_loss / batch_no
epoch_time = time.time() - start_time
logger.info('epoch %d %.2fs Train loss is [%.4f] ' % (epoch + 1, epoch_time, total_loss))
self.clear_parameter(self.model)
# validate
start_time = time.time()
validate_metric_dict = self.evaluate(epoch, self.test_batch_size, self.dataset.validate_dataset(),
self.dataset.validation_interacts, self.dataset.validation_gt_length)
epoch_time = time.time() - start_time
logger.info(
f"validate %d cost time %.2fs, result: %s " % (epoch + 1, epoch_time, validate_metric_dict.__str__()))
# test
start_time = time.time()
test_metric_dict = self.evaluate(epoch, self.test_batch_size, self.dataset.test_dataset(),
self.dataset.test_interacts, self.dataset.test_gt_length)
epoch_time = time.time() - start_time
logger.info(
f"test %d cost time %.2fs, result: %s " % (epoch + 1, epoch_time, test_metric_dict.__str__()))
return test_metric_dict, validate_metric_dict
@logger.catch()
@torch.no_grad()
def evaluate(self, epoch, test_batch_size, dataset, gt_interacts, gt_length):
data_loader = DataLoader(dataset=dataset, batch_size=test_batch_size)
self.model.eval()
start_time = time.time()
iter_data = (
tqdm(
enumerate(data_loader),
total=len(data_loader),
desc=f"\033[1;35mEvaluate \033[0m"
)
)
topk_list = []
train_items = self.dataset.train_behavior_dict[self.behaviors[-1]]
for batch_index, batch_data in iter_data:
batch_data = batch_data.to(self.device)
start = time.time()
# scores = self.model.module.full_predict(batch_data)
scores = self.model.full_predict(batch_data)
for index, user in enumerate(batch_data):
user_score = scores[index]
items = train_items.get(str(user.item()), None)
if items is not None:
user_score[items] = -np.inf
_, topk_idx = torch.topk(user_score, max(self.topk), dim=-1)
gt_items = gt_interacts[str(user.item())]
mask = np.isin(topk_idx.to('cpu'), gt_items)
topk_list.append(mask)
topk_list = np.array(topk_list)
metric_dict = self.calculate_result(topk_list, gt_length)
return metric_dict
def calculate_result(self, topk_list, gt_len):
result_list = []
for metric in self.metrics:
metric_fuc = metrics_dict[metric.lower()]
result = metric_fuc(topk_list, gt_len)
result_list.append(result)
result_list = np.stack(result_list, axis=0).mean(axis=1)
metric_dict = {}
for topk in self.topk:
for metric, value in zip(self.metrics, result_list):
key = '{}@{}'.format(metric, topk)
metric_dict[key] = np.round(value[topk - 1], 4)
return metric_dict
def save_model(self, model):
torch.save(model.state_dict(), os.path.join(self.model_path, self.model_name + self.TIME + '.pth'))