-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
367 lines (304 loc) · 17.5 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
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
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
import os
from utils.data import get_dataloader
import config.const as const_util
import utils.Context.ContextManager as CM
import torch
import torch.optim as optim
import copy
import pickle
import numpy as np
import random
from time import time
import torch.nn as nn
from utils.topk import TopK_custom, set_device_for_topk
class Trainer(object):
def __init__(self, flags_obj, cm, dm, new_config=None):
self.cm = cm #context manager
self.dm = dm #dataset manager
self.flags_obj = flags_obj
self.recommender = CM.set_recommender(flags_obj, cm.workspace, dm, new_config)
self.device = CM.set_device(self.flags_obj)
self.recommender.transfer_model(self.device)
self.lr_rec = self.flags_obj.lr_rec
self.lr_label = self.flags_obj.lr_label
self.set_train_dataloader()
self.set_valid_dataloader()
self.set_test_dataloader()
set_device_for_topk(self.device)
self.topk_model = TopK_custom(k=10, epsilon=1e-4).to(self.device)
if self.flags_obj.dataset_name == 'kuaishou':
if self.flags_obj.normalization == 1:
def normalize_play_time(tensor):
mask = tensor <= 50.0
y = torch.zeros_like(tensor)
y[mask] = tensor[mask] / 50.0 * 0.2
if len(mask) < len(tensor):
y[~mask] = torch.min( (tensor[~mask]-50.0) / (800.0-50.0) * 0.8 + 0.2, torch.tensor([1.0]).to(self.device))[0]
return y
def normalize_duration(tensor):
mask = tensor <= 235.0
y = torch.zeros_like(tensor)
y[mask] = tensor[mask] / 235.0 * 0.2
if len(mask) < len(tensor):
y[~mask] = torch.min( (tensor[~mask]-235.0) / (3600.0-235.0) * 0.8 + 0.2, torch.tensor([1.0]).to(self.device))[0]
return y
else:
def normalize_play_time(tensor):
return tensor
def normalize_duration(tensor):
return tensor
elif self.flags_obj.dataset_name == 'wechat':
if self.flags_obj.normalization == 1:
def normalize_play_time(tensor):
mask = tensor <= 51.3
y = torch.zeros_like(tensor)
y[mask] = tensor[mask] / 51.3 * 0.2
if len(mask) < len(tensor):
y[~mask] = torch.min( (tensor[~mask]-51.3) / (8514.0-51.3) * 0.8 + 0.2, torch.tensor([1.0]).to(self.device))[0]
return y
def normalize_duration(tensor):
mask = tensor <= 59.0
y = torch.zeros_like(tensor)
y[mask] = tensor[mask] / 59.0 * 0.2
if len(mask) < len(tensor):
y[~mask] = torch.min( (tensor[~mask]-59.0) / (10275.0-59.0) * 0.8 + 0.2, torch.tensor([1.0]).to(self.device))[0]
return y
else:
def normalize_play_time(tensor):
return tensor
def normalize_duration(tensor):
return tensor
self.normalize_duration = normalize_duration
self.normalize_play_time = normalize_play_time
def set_train_dataloader(self):
self.train_dataset = self.recommender.get_dataset(const_util.train_file, self.dm, True)
self.train_dataloader = get_dataloader(
data_set = self.train_dataset,
bs = self.dm.batch_size,
collate_fn = self.train_dataset.collate_func,
shuffle = True
)
def set_valid_dataloader(self, k=10):
self.valid_dataset = self.recommender.get_dataset(const_util.valid_file, self.dm, False)
self.valid_user_counts = self.valid_dataset.sampler.record.groupby('user_id').size().to_dict()
self.valid_user_counts= {key: value for key, value in self.valid_user_counts.items() if value > k}
self.valid_dataloader = get_dataloader(
data_set = self.valid_dataset,
bs = self.dm.batch_size,
collate_fn = self.valid_dataset.collate_func,
shuffle = False
)
def set_test_dataloader(self, k=10):
self.test_dataset = self.recommender.get_dataset(const_util.test_file, self.dm, False)
self.test_user_counts = self.test_dataset.sampler.record.groupby('user_id').size().to_dict()
self.test_user_counts= {key: value for key, value in self.test_user_counts.items() if value > k}
self.test_dataloader = get_dataloader(
data_set = self.test_dataset,
bs = self.dm.batch_size,
collate_fn = self.test_dataset.collate_func,
shuffle = False
)
def test(self, k=10, user_counts=None, dataset=None):
with torch.no_grad():
model = self.recommender.model
result_test_k = dict()
test_user_num = len(user_counts)
test_user_weight = list(user_counts.values())
temp = sum(test_user_weight)
test_user_weight = [x / temp for x in test_user_weight]
for u in list(user_counts.keys()):
items_info, feedbacks_info, scores = self.get_score_test(u, model, dataset=dataset)
play_time = feedbacks_info[:, 0].float()
duration = feedbacks_info[:, 1].float()
lfc = feedbacks_info[:, 2].float()
if len(play_time) > k:
_, top_indices = torch.topk(scores, k=k)
pos_weight = torch.log2(torch.arange(2, k+2)).to(self.device)
duration_result = torch.std(duration[top_indices]).cpu().numpy()
top_play_time_pos = torch.dot(pos_weight, torch.topk(play_time, k=k)[0])
top_play_time_pos_by_score = torch.dot(pos_weight, play_time[top_indices])
play_time_pos_result = (top_play_time_pos_by_score / top_play_time_pos).cpu().numpy() if top_play_time_pos > 0 else np.array(1.0)
top_lfc_pos = torch.dot(pos_weight, torch.topk(lfc, k=k)[0])
top_lfc_pos_by_score = torch.dot(pos_weight, lfc[top_indices])
lfc_pos_result = (top_lfc_pos_by_score / top_lfc_pos).cpu().numpy() if top_lfc_pos > 0 else np.array(1.0)
result_test_k[u] = [duration_result,
play_time_pos_result,
lfc_pos_result]
temp = [v[0] for v in list(result_test_k.values())]
duration_average_result = np.sum(temp) / test_user_num
duration_weighted_result = sum([test_user_weight[i] * temp[i] for i in range(len(test_user_weight))])
print("top@{} {} duration std:{}".format(k, 'average', duration_average_result))
print("top@{} {} duration std:{}".format(k, 'weighted', duration_weighted_result))
temp = [v[1] for v in list(result_test_k.values())]
play_time_pos_average_result = np.sum(temp) / test_user_num
play_time_pos_weighted_result = sum([test_user_weight[i] * temp[i] for i in range(len(test_user_weight))])
print("top@{} {} play time ratio:{}".format(k, 'average', play_time_pos_average_result))
print("top@{} {} play time ratio:{}".format(k, 'weighted', play_time_pos_weighted_result))
temp = [v[2] for v in list(result_test_k.values())]
lfc_pos_average_result = np.sum(temp) / test_user_num
lfc_pos_weighted_result = sum([test_user_weight[i] * temp[i] for i in range(len(test_user_weight))])
print("top@{} {} lfc ratio:{}".format(k, 'average', lfc_pos_average_result))
print("top@{} {} lfc ratio:{}".format(k, 'weighted', lfc_pos_weighted_result))
result = [duration_average_result, duration_weighted_result,
play_time_pos_average_result, play_time_pos_weighted_result,
lfc_pos_average_result, lfc_pos_weighted_result]
return result
def get_score_test(self, user, model, params_updated=None, dataset=None):
sample = dataset.get_user_batch_final(user)
sample = [[k.to(self.device) for k in i] if type(i) == list else i.to(self.device) for i in sample]
input_data, feedbacks = sample[:-1], sample[3]
scores = model.forward(input_data=input_data, params_updated=params_updated) # shape: [B]
items_info = sample[1]
feedbacks_info = sample[3]
return items_info, feedbacks_info, scores
def train_and_test(self):
train_loss = [] #store every training loss
valid_metric = [] #store every validation metric
not_better = 0
best_metric = 0
save_name = ''
for name_str, name_val in self.recommender.model_config.items():
save_name += name_str + '_' + str(name_val) + '_'
for epoch in range(self.flags_obj.epochs):
loss = self.train_one_epoch(epoch, self.train_dataloader)
print("TRAIN")
print("train loss:{}".format(loss))
train_loss.append(loss)
if epoch % 1 == 0:
print("VALIDATE")
valid_result = self.test(k=10, user_counts=self.valid_user_counts, dataset=self.valid_dataset)
valid_loss = self.get_loss(self.valid_dataloader)
print("valid loss:", valid_loss)
valid_metric.append(valid_result[3])
print("TEST")
test_result = self.test(k=10, user_counts=self.test_user_counts, dataset=self.test_dataset)
test_loss = self.get_loss(self.test_dataloader)
print("test loss:", test_loss)
if valid_result[3] > best_metric + 1e-4:
not_better = 0
best_metric = valid_result[3]
torch.save(self.recommender.model.state_dict(), './workspace/ckpt/rec_' + self.flags_obj.extra_name + '_' + save_name + '.pth')
torch.save(self.recommender.labeler.state_dict(), './workspace/ckpt/labeler_' + self.flags_obj.extra_name + '_' + save_name + '.pth')
else:
not_better += 1
if not_better >= 5:
print("end at epoch {}".format(epoch))
print("train loss", train_loss)
print("valid metric", valid_metric)
break
print('best result:')
self.recommender.model.load_state_dict(torch.load('./workspace/ckpt/rec_' + self.flags_obj.extra_name + '_' + save_name + '.pth'))
self.recommender.labeler.load_state_dict(torch.load('./workspace/ckpt/labeler_' + self.flags_obj.extra_name + '_' + save_name + '.pth'))
valid_result = self.test(k=10, user_counts=self.valid_user_counts, dataset=self.valid_dataset)
test_result = self.test(k=10, user_counts=self.test_user_counts, dataset=self.test_dataset)
def train_one_epoch(self, epoch, dataloader):
self.recommender.model.train()
self.recommender.labeler.train()
loss_train = 0
for step, sample in enumerate(dataloader):
#t0 = time()
sample = [[k.to(self.device) for k in i] if type(i) == list else i.to(self.device) for i in sample]
#t1 = time()
#print("Time: sample to device ", t1-t0)
model_assume = copy.deepcopy(self.recommender.model).to(self.device)
for param in model_assume.parameters():
param.grad = None
input_data, feedbacks = sample[:-1], sample[-1]
labels = self.recommender.labeler(feedbacks).squeeze(-1)
loss_assume = model_assume.forward(input_data, labels)
grads = torch.autograd.grad(loss_assume, model_assume.parameters(), create_graph=True, retain_graph=True)
params_assume_updated = list()
for i, param in enumerate(model_assume.parameters()):
param_updated = param - self.lr_rec * grads[i]
params_assume_updated.append(param_updated)
#t2 = time()
#print("Time: substep1 ", t2-t1)
valid_user_num = len(self.valid_user_counts)
sampled_users = random.sample(list(self.valid_user_counts.keys()), int(valid_user_num * self.flags_obj.sample_ratio))
result = dict()
for u in sampled_users:
result[u] = self.get_score_test(u, model_assume, params_assume_updated, self.valid_dataset)
total_play_time, total_duration, total_lfc= self.get_metric(result=result, k=10)
total_play_time = torch.stack(total_play_time)
total_duration = torch.stack(total_duration)
total_lfc = torch.stack(total_lfc)
metric_play_time, metric_duration, metric_lfc = sum(total_play_time), sum(total_duration), sum(total_lfc)
grads_alpha_play_time = torch.autograd.grad(metric_play_time, self.recommender.labeler.parameters(), retain_graph=True)
grads_alpha_duration = torch.autograd.grad(metric_duration, self.recommender.labeler.parameters(), retain_graph=True)
grads_alpha_lfc = torch.autograd.grad(metric_lfc, self.recommender.labeler.parameters(), retain_graph=True)
grads_alpha_play_time_fc = self.flatten_and_concatenate(grads_alpha_play_time)
grads_alpha_duration_fc = self.flatten_and_concatenate(grads_alpha_duration)
grads_alpha_lfc_fc = self.flatten_and_concatenate(grads_alpha_lfc)
if self.flags_obj.adapt == 0:
weight = torch.tensor([1/3, 1/3, 1/3], dtype=torch.float32)
elif self.flags_obj.disable > 0:
weight = torch.ones(3, dtype=torch.float32)
weight[self.flags_obj.disable - 1] = 0
else:
weight = torch.stack([metric_play_time, metric_duration, metric_lfc])
while torch.min(weight) < 1:
weight = weight * 10
weight = torch.exp(-self.flags_obj.adapt * weight)
weight = weight / torch.sum(weight)
total_metric = metric_play_time * weight[0] + metric_duration * weight[1] + metric_lfc * weight[2]
for i, param in enumerate(self.recommender.labeler.parameters()):
param.data = param.data + self.lr_label * (grads_alpha_play_time[i] * weight[0] + grads_alpha_duration[i] * weight[1] + grads_alpha_lfc[i] * weight[2])
#t3 = time()
#print("Time: substep2 ", t3-t2)
#self.rec_optimizer.zero_grad()
with torch.no_grad():
labels = self.recommender.labeler(feedbacks).squeeze(-1)
loss = self.recommender.model.forward(input_data, labels)
loss_train += loss.detach().cpu().numpy()
grads_final = torch.autograd.grad(loss, self.recommender.model.parameters())
for i, param in enumerate(self.recommender.model.parameters()):
param.data = param.data - self.lr_rec * grads_final[i]
#t4 = time()
#print("Time: substep3 ", t4-t3)
print("epoch:{}, step:{}, loss_assume:{}".format(epoch, step, loss_assume.data))
print("metric_play_time:{}, metric_duration:{}, metric_lfc:{}".format(metric_play_time, metric_duration, metric_lfc))
print("weight:{}, total_metric:{}".format(weight.data.cpu().numpy(), total_metric))
print("loss:{}".format(loss))
print("\n")
return loss_train / step
def get_loss(self, dataloader):
with torch.no_grad():
model = self.recommender.model
labeler = self.recommender.labeler
result = 0
for step, sample in enumerate(dataloader):
sample = [[k.to(self.device) for k in i] if type(i) == list else i.to(self.device) for i in sample]
input_data, feedbacks = sample[:-1], sample[-1]
labels = labeler(feedbacks).squeeze(-1)
loss = model.forward(input_data, labels)
result += loss.detach().cpu().numpy()
return result / step
def flatten_and_concatenate(self, tensor_list):
flat_tensor_list = []
for tensor in tensor_list:
flat_tensor = torch.flatten(tensor)
flat_tensor_list.append(flat_tensor)
concatenated_tensor = torch.cat(flat_tensor_list, dim=0)
return concatenated_tensor
def get_metric(self, result, k=10):
total_play_time = []
total_duration = []
total_lfc = []
for user, info in result.items():
items_info = info[0] # id,duration,tag
feedbacks_info = info[1] # play_time, duration, like follow comment
play_time = feedbacks_info[:, 0].float()
duration = feedbacks_info[:, 1].float()
lfc = feedbacks_info[:, 2].float()
scores = info[-1]
N = items_info.shape[0]
if N < k:
continue
is_topk = self.topk_model(scores.unsqueeze(dim=0))[0]
is_topk = torch.sum(torch.sum(is_topk, dim=0), dim=1)
play_time = self.normalize_play_time(play_time)
duration = self.normalize_duration(duration)
total_play_time.append(torch.dot(is_topk, play_time)/k)
total_duration.append(torch.std(is_topk * duration))
total_lfc.append(torch.dot(is_topk, lfc)/k)
return total_play_time, total_duration, total_lfc