-
Notifications
You must be signed in to change notification settings - Fork 44
/
train.py
234 lines (182 loc) · 8.79 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
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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import torch
import torch.optim as optim
import numpy as np
from tqdm import tqdm
from .model import gen_model_dir
import os
from functools import partial
from collections import deque, defaultdict
import logging
logger = logging.getLogger(__name__)
OPTIMIZER_FILE = "optimizer.pt"
# Decay functions to be used with lr_scheduler
def lr_decay_noam(hparams):
return lambda t: (
10.0 * hparams.hidden_size**-0.5 * min(
(t + 1) * hparams.learning_rate_warmup_steps**-1.5, (t + 1)**-0.5))
def lr_decay_exp(hparams):
return lambda t: hparams.learning_rate_falloff ** t
# Map names to lr decay functions
lr_decay_map = {
'noam': lr_decay_noam,
'exp': lr_decay_exp
}
def compute_num_params(model):
"""
Computes number of trainable and non-trainable parameters
"""
sizes = [(np.array(p.data.size()).prod(), int(p.requires_grad)) for p in model.parameters()]
return sum(map(lambda t: t[0]*t[1],sizes)), sum(map(lambda t: t[0]*(1 - t[1]),sizes))
class Trainer(object):
"""
Class to handle training in a task-agnostic way
"""
def __init__(self, task_name, model, hparams, train_iter, evaluator):
"""
Parameters:
task_name: Name of the task
model: Model instance (derived from model.Model)
hparams: Instance of HParams
train_iter: An instance of torchtext.data.Iterator
evaluator: Instance of evalutation.Evaluator that will
run metrics on the validation dataset
"""
self.task_name = task_name
self.model = model
self.hparams = hparams
self.evaluator = evaluator
self.train_iter = train_iter
# Disable repetitions
self.train_iter.repeat = False
model_params = filter(lambda p: p.requires_grad, model.parameters())
# TODO: Add support for other optimizers
self.optimizer = optim.Adam(
model_params,
betas=(hparams.optimizer_adam_beta1, hparams.optimizer_adam_beta2),
lr=hparams.learning_rate)
self.opt_path = os.path.join(gen_model_dir(task_name, model.__class__),
OPTIMIZER_FILE)
# If model is loaded from a checkpoint restore optimizer also
if int(model.iterations) > 0:
self.optimizer.load_state_dict(torch.load(self.opt_path))
self.lr_scheduler_step = self.lr_scheduler_epoch = None
# Set up learing rate decay scheme
if hparams.learning_rate_decay is not None:
if '_' not in hparams.learning_rate_decay:
raise ValueError("Malformed learning_rate_decay")
lrd_scheme, lrd_range = hparams.learning_rate_decay.split('_')
if lrd_scheme not in lr_decay_map:
raise ValueError("Unknown lr decay scheme {}".format(lrd_scheme))
lrd_func = lr_decay_map[lrd_scheme]
lr_scheduler = optim.lr_scheduler.LambdaLR(
self.optimizer,
lrd_func(hparams),
last_epoch=int(self.model.iterations) or -1
)
# For each scheme, decay can happen every step or every epoch
if lrd_range == 'epoch':
self.lr_scheduler_epoch = lr_scheduler
elif lrd_range == 'step':
self.lr_scheduler_step = lr_scheduler
else:
raise ValueError("Unknown lr decay range {}".format(lrd_range))
# Display number of parameters
logger.info('Parameters: {}(trainable), {}(non-trainable)'.format(*compute_num_params(self.model)))
def _get_early_stopping_criteria(self, early_stopping):
es = early_stopping.split('_')
if len(es) != 3:
raise ValueError('Malformed early stopping criteria')
best_type, window, metric = es
logger.info('Early stopping for {} value of validation {} after {} epochs'
.format(best_type, metric, window))
if best_type == 'lowest':
best_fn = partial(min, key=lambda item: item[0])
elif best_type == 'highest':
best_fn = partial(max, key=lambda item: item[0])
else:
raise ValueError('Unknown best type {}'.format(best_type))
return best_fn, int(window), metric
def train(self, num_epochs, early_stopping=None, save=True):
"""
Run the training loop for given number of epochs. The model
is evaluated at the end of every epoch and saved as well
Parameters:
num_epochs: Total number of epochs to run
early_stopping: A string indicating how to perform early stopping
Should be of the form lowest/highest_n_metric where:
lowest/highest: Track lowest or highest values
n: The window size within which to track best
metric: Name of the metric to track. Should be available
in the dict returned by evaluator
save: Save model every epoch if true
Returns:
Tuple of best checkpoint number and metrics array (for plotting etc)
"""
all_metrics = defaultdict(list)
best_iteration = 0
if early_stopping:
if not save:
raise ValueError('save should be True for early stopping')
if self.evaluator is None:
raise ValueError('early stopping requires an eval function')
best_fn, best_window, best_metric_name = self._get_early_stopping_criteria(early_stopping)
tracking = deque([], best_window + 1)
for epoch in range(num_epochs):
self.train_iter.init_epoch()
epoch_loss = 0
count = 0
logger.info('Epoch %d (%d)'%(epoch + 1, int(self.model.iterations)))
prog_iter = tqdm(self.train_iter, leave=False)
for batch in prog_iter:
# Train mode
self.model.train()
self.optimizer.zero_grad()
loss, _ = self.model.loss(batch)
loss.backward()
self.optimizer.step()
if self.lr_scheduler_step:
self.lr_scheduler_step.step()
epoch_loss += loss.item()
count += 1
self.model.iterations += 1
# Display loss
prog_iter.set_description('Training')
prog_iter.set_postfix(loss=(epoch_loss/count))
if self.lr_scheduler_epoch:
self.lr_scheduler_epoch.step()
train_loss = epoch_loss/count
all_metrics['train_loss'].append(train_loss)
logger.info('Train Loss: {:3.5f}'.format(train_loss))
best_iteration = int(self.model.iterations)
# Run evaluation
if self.evaluator:
eval_metrics = self.evaluator.evaluate(self.model)
if not isinstance(eval_metrics, dict):
raise ValueError('eval_fn should return a dict of metrics')
# Display eval metrics
logger.info('Validation metrics: ')
logger.info(', '.join(['{}={:3.5f}'.format(k, v) for k,v in eval_metrics.items()]))
# Append metrics
for k, v in eval_metrics.items():
all_metrics[k].append(v)
# Handle early stopping
tracking.append((eval_metrics[best_metric_name], int(self.model.iterations), epoch))
logger.debug('Epoch {} Tracking: {}'.format(epoch, tracking))
if epoch >= best_window:
# Get the best value of metric in the window
best_metric, best_iteration, best_epoch = best_fn(tracking)
if tracking[0][1] == best_iteration:
# The best value has gone outside the desired window
# hence stop
logger.info('Early stopping at iteration {}, epoch {}, {}={:3.5f}'
.format(best_iteration, best_epoch, best_metric_name, best_metric))
# Update the file time of that checkpoint file to latest
self.model.set_latest(self.task_name, best_iteration)
break
if save:
self.model.save(self.task_name)
torch.save(self.optimizer.state_dict(), self.opt_path)
return best_iteration, all_metrics