-
Notifications
You must be signed in to change notification settings - Fork 147
/
rank_model.py
285 lines (258 loc) · 12.5 KB
/
rank_model.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
# =========================================================================
# Copyright (C) 2024. The FuxiCTR Library. All rights reserved.
# Copyright (C) 2022. Huawei Technologies Co., Ltd. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# =========================================================================
import torch.nn as nn
import numpy as np
import torch
import os, sys
import logging
from fuxictr.metrics import evaluate_metrics
from fuxictr.pytorch.torch_utils import get_device, get_optimizer, get_loss, get_regularizer
from fuxictr.utils import Monitor, not_in_whitelist
from tqdm import tqdm
class BaseModel(nn.Module):
def __init__(self,
feature_map,
model_id="BaseModel",
task="binary_classification",
gpu=-1,
monitor="AUC",
save_best_only=True,
monitor_mode="max",
early_stop_patience=2,
eval_steps=None,
embedding_regularizer=None,
net_regularizer=None,
reduce_lr_on_plateau=True,
**kwargs):
super(BaseModel, self).__init__()
self.device = get_device(gpu)
self._monitor = Monitor(kv=monitor)
self._monitor_mode = monitor_mode
self._early_stop_patience = early_stop_patience
self._eval_steps = eval_steps # None default, that is evaluating every epoch
self._save_best_only = save_best_only
self._embedding_regularizer = embedding_regularizer
self._net_regularizer = net_regularizer
self._reduce_lr_on_plateau = reduce_lr_on_plateau
self._verbose = kwargs["verbose"]
self.feature_map = feature_map
self.output_activation = self.get_output_activation(task)
self.model_id = model_id
self.model_dir = os.path.join(kwargs["model_root"], feature_map.dataset_id)
self.checkpoint = os.path.abspath(os.path.join(self.model_dir, self.model_id + ".model"))
self.validation_metrics = kwargs["metrics"]
def compile(self, optimizer, loss, lr):
self.optimizer = get_optimizer(optimizer, self.parameters(), lr)
self.loss_fn = get_loss(loss)
def regularization_loss(self):
reg_term = 0
if self._embedding_regularizer or self._net_regularizer:
emb_reg = get_regularizer(self._embedding_regularizer)
net_reg = get_regularizer(self._net_regularizer)
for _, module in self.named_modules():
for p_name, param in module.named_parameters():
if param.requires_grad:
if p_name in ["weight", "bias"]:
if type(module) == nn.Embedding:
if self._embedding_regularizer:
for emb_p, emb_lambda in emb_reg:
reg_term += (emb_lambda / emb_p) * torch.norm(param, emb_p) ** emb_p
else:
if self._net_regularizer:
for net_p, net_lambda in net_reg:
reg_term += (net_lambda / net_p) * torch.norm(param, net_p) ** net_p
return reg_term
def compute_loss(self, return_dict, y_true):
loss = self.loss_fn(return_dict["y_pred"], y_true, reduction='mean')
loss += self.regularization_loss()
return loss
def reset_parameters(self):
def reset_default_params(m):
# initialize nn.Linear/nn.Conv1d layers by default
if type(m) in [nn.Linear, nn.Conv1d]:
nn.init.xavier_normal_(m.weight)
if m.bias is not None:
m.bias.data.fill_(0)
def reset_custom_params(m):
# initialize layers with customized reset_parameters
if hasattr(m, 'reset_custom_params'):
m.reset_custom_params()
self.apply(reset_default_params)
self.apply(reset_custom_params)
def get_inputs(self, inputs, feature_source=None):
X_dict = dict()
for feature in inputs.keys():
if feature in self.feature_map.labels:
continue
spec = self.feature_map.features[feature]
if spec["type"] == "meta":
continue
if feature_source and not_in_whitelist(spec["source"], feature_source):
continue
X_dict[feature] = inputs[feature].to(self.device)
return X_dict
def get_labels(self, inputs):
""" Please override get_labels() when using multiple labels!
"""
labels = self.feature_map.labels
y = inputs[labels[0]].to(self.device)
return y.float().view(-1, 1)
def get_group_id(self, inputs):
return inputs[self.feature_map.group_id]
def model_to_device(self):
self.to(device=self.device)
def lr_decay(self, factor=0.1, min_lr=1e-6):
for param_group in self.optimizer.param_groups:
reduced_lr = max(param_group["lr"] * factor, min_lr)
param_group["lr"] = reduced_lr
return reduced_lr
def fit(self, data_generator, epochs=1, validation_data=None,
max_gradient_norm=10., **kwargs):
self.valid_gen = validation_data
self._max_gradient_norm = max_gradient_norm
self._best_metric = np.Inf if self._monitor_mode == "min" else -np.Inf
self._stopping_steps = 0
self._steps_per_epoch = len(data_generator)
self._stop_training = False
self._total_steps = 0
self._batch_index = 0
self._epoch_index = 0
if self._eval_steps is None:
self._eval_steps = self._steps_per_epoch
logging.info("Start training: {} batches/epoch".format(self._steps_per_epoch))
logging.info("************ Epoch=1 start ************")
for epoch in range(epochs):
self._epoch_index = epoch
self.train_epoch(data_generator)
if self._stop_training:
break
else:
logging.info("************ Epoch={} end ************".format(self._epoch_index + 1))
logging.info("Training finished.")
logging.info("Load best model: {}".format(self.checkpoint))
self.load_weights(self.checkpoint)
def checkpoint_and_earlystop(self, logs, min_delta=1e-6):
monitor_value = self._monitor.get_value(logs)
if (self._monitor_mode == "min" and monitor_value > self._best_metric - min_delta) or \
(self._monitor_mode == "max" and monitor_value < self._best_metric + min_delta):
self._stopping_steps += 1
logging.info("Monitor({})={:.6f} STOP!".format(self._monitor_mode, monitor_value))
if self._reduce_lr_on_plateau:
current_lr = self.lr_decay()
logging.info("Reduce learning rate on plateau: {:.6f}".format(current_lr))
else:
self._stopping_steps = 0
self._best_metric = monitor_value
if self._save_best_only:
logging.info("Save best model: monitor({})={:.6f}"\
.format(self._monitor_mode, monitor_value))
self.save_weights(self.checkpoint)
if self._stopping_steps >= self._early_stop_patience:
self._stop_training = True
logging.info("********* Epoch={} early stop *********".format(self._epoch_index + 1))
if not self._save_best_only:
self.save_weights(self.checkpoint)
def eval_step(self):
logging.info('Evaluation @epoch {} - batch {}: '.format(self._epoch_index + 1, self._batch_index + 1))
val_logs = self.evaluate(self.valid_gen, metrics=self._monitor.get_metrics())
self.checkpoint_and_earlystop(val_logs)
self.train()
def train_step(self, batch_data):
self.optimizer.zero_grad()
return_dict = self.forward(batch_data)
y_true = self.get_labels(batch_data)
loss = self.compute_loss(return_dict, y_true)
loss.backward()
nn.utils.clip_grad_norm_(self.parameters(), self._max_gradient_norm)
self.optimizer.step()
return loss
def train_epoch(self, data_generator):
self._batch_index = 0
train_loss = 0
self.train()
if self._verbose == 0:
batch_iterator = data_generator
else:
batch_iterator = tqdm(data_generator, disable=False, file=sys.stdout)
for batch_index, batch_data in enumerate(batch_iterator):
self._batch_index = batch_index
self._total_steps += 1
loss = self.train_step(batch_data)
train_loss += loss.item()
if self._total_steps % self._eval_steps == 0:
logging.info("Train loss: {:.6f}".format(train_loss / self._eval_steps))
train_loss = 0
self.eval_step()
if self._stop_training:
break
def evaluate(self, data_generator, metrics=None):
self.eval() # set to evaluation mode
with torch.no_grad():
y_pred = []
y_true = []
group_id = []
if self._verbose > 0:
data_generator = tqdm(data_generator, disable=False, file=sys.stdout)
for batch_data in data_generator:
return_dict = self.forward(batch_data)
y_pred.extend(return_dict["y_pred"].data.cpu().numpy().reshape(-1))
y_true.extend(self.get_labels(batch_data).data.cpu().numpy().reshape(-1))
if self.feature_map.group_id is not None:
group_id.extend(self.get_group_id(batch_data).numpy().reshape(-1))
y_pred = np.array(y_pred, np.float64)
y_true = np.array(y_true, np.float64)
group_id = np.array(group_id) if len(group_id) > 0 else None
if metrics is not None:
val_logs = self.evaluate_metrics(y_true, y_pred, metrics, group_id)
else:
val_logs = self.evaluate_metrics(y_true, y_pred, self.validation_metrics, group_id)
logging.info('[Metrics] ' + ' - '.join('{}: {:.6f}'.format(k, v) for k, v in val_logs.items()))
return val_logs
def predict(self, data_generator):
self.eval() # set to evaluation mode
with torch.no_grad():
y_pred = []
if self._verbose > 0:
data_generator = tqdm(data_generator, disable=False, file=sys.stdout)
for batch_data in data_generator:
return_dict = self.forward(batch_data)
y_pred.extend(return_dict["y_pred"].data.cpu().numpy().reshape(-1))
y_pred = np.array(y_pred, np.float64)
return y_pred
def evaluate_metrics(self, y_true, y_pred, metrics, group_id=None):
return evaluate_metrics(y_true, y_pred, metrics, group_id)
def save_weights(self, checkpoint):
torch.save(self.state_dict(), checkpoint)
def load_weights(self, checkpoint):
self.to(self.device)
state_dict = torch.load(checkpoint, map_location="cpu")
self.load_state_dict(state_dict)
def get_output_activation(self, task):
if task == "binary_classification":
return nn.Sigmoid()
elif task == "regression":
return nn.Identity()
else:
raise NotImplementedError("task={} is not supported.".format(task))
def count_parameters(self, count_embedding=True):
total_params = 0
for name, param in self.named_parameters():
if not count_embedding and "embedding" in name:
continue
if param.requires_grad:
total_params += param.numel()
logging.info("Total number of parameters: {}.".format(total_params))