/
model.py
403 lines (344 loc) · 14.6 KB
/
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
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
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
from . import display
from . import losses as losses_module
from . import metrics as metrics_module
from . import train as train_module
from .backend import tf
from .callbacks import CallbackList
from .utils import guarantee_initialized_variables, timing
class Model(object):
"""The ``Model`` class trains a ``Map`` on a ``Data``.
Args:
data: ``deepxde.data.Data`` instance.
net: ``deepxde.maps.Map`` instance.
"""
def __init__(self, data, net):
self.data = data
self.net = net
self.optimizer = None
self.batch_size = None
self.losses = None
self.totalloss = None
self.train_op = None
self.metrics = None
self.sess = None
self.saver = None
self.train_state = TrainState()
self.losshistory = LossHistory()
self.stop_training = False
self.callbacks = None
def close(self):
self._close_tfsession()
@timing
def compile(
self,
optimizer,
lr=None,
loss="MSE",
metrics=None,
decay=None,
loss_weights=None,
):
"""Configures the model for training.
Args:
optimizer: String. Name of optimizer.
lr: A Tensor or a floating point value. The learning rate.
loss: String (name of objective function) or objective function.
metrics: List of metrics to be evaluated by the model during training.
decay: Tuple. Name and parameters of decay to the initial learning rate. One of the following options:
- `inverse time decay <https://www.tensorflow.org/api_docs/python/tf/compat/v1/train/inverse_time_decay>`_: ("inverse time", decay_steps, decay_rate)
- `cosine decay <https://www.tensorflow.org/api_docs/python/tf/compat/v1/train/cosine_decay>`_: ("cosine", decay_steps, alpha)
loss_weights: A list specifying scalar coefficients (Python floats)
to weight the loss contributions. The loss value that will be minimized by the model
will then be the weighted sum of all individual losses,
weighted by the loss_weights coefficients.
"""
print("Compiling model...")
if not self.net.built:
self.net.build()
self._open_tfsession()
self.optimizer = optimizer
loss = losses_module.get(loss)
self.losses = self.data.losses(self.net.targets, self.net.outputs, loss, self)
if self.net.regularizer is not None:
self.losses.append(tf.losses.get_regularization_loss())
self.losses = tf.convert_to_tensor(self.losses)
if loss_weights is not None:
self.losses *= loss_weights
self.losshistory.set_loss_weights(loss_weights)
self.totalloss = tf.reduce_sum(self.losses)
self.train_op = train_module.get_train_op(
self.totalloss, self.optimizer, lr=lr, decay=decay
)
metrics = metrics or []
self.metrics = [metrics_module.get(m) for m in metrics]
@timing
def train(
self,
epochs=None,
batch_size=None,
display_every=1000,
uncertainty=False,
disregard_previous_best=False,
callbacks=None,
model_restore_path=None,
model_save_path=None,
print_model=False,
):
"""Trains the model for a fixed number of epochs (iterations on a dataset).
Args:
epochs: Integer. Number of epochs to train the model.
batch_size: Integer or ``None``. Not fully supported yet.
display_every: Integer. Print the loss and metrics every this steps.
uncertainty: Boolean. If ``True``, use Monte-Carlo Dropout to estimate uncertainty.
disregard_previous_best: If ``True``, disregard the previous saved best model.
callbacks: List of ``deepxde.callbacks.Callback`` instances.
List of callbacks to apply during training.
model_restore_path: String. Path where parameters were previously saved.
See ``save_path`` in `tf.train.Saver.restore <https://www.tensorflow.org/api_docs/python/tf/compat/v1/train/Saver#restore>`_.
model_save_path: String. Prefix of filenames created for the checkpoint.
See ``save_path`` in `tf.train.Saver.save <https://www.tensorflow.org/api_docs/python/tf/compat/v1/train/Saver#save>`_.
print_model: If ``True``, print the values of all variables.
"""
self.batch_size = batch_size
self.callbacks = CallbackList(callbacks=callbacks)
self.callbacks.set_model(self)
if disregard_previous_best:
self.train_state.disregard_best()
if self.train_state.step == 0:
print("Initializing variables...")
self.sess.run(tf.global_variables_initializer())
else:
guarantee_initialized_variables(self.sess)
if model_restore_path is not None:
print("Restoring model from {} ...".format(model_restore_path))
self.saver.restore(self.sess, model_restore_path)
print("Training model...\n")
self.stop_training = False
self.train_state.set_data_train(*self.data.train_next_batch(self.batch_size))
self.train_state.set_data_test(*self.data.test())
self._test(uncertainty)
self.callbacks.on_train_begin()
if train_module.is_scipy_opts(self.optimizer):
self._train_scipy(display_every, uncertainty)
else:
if epochs is None:
raise ValueError("No epochs for {}.".format(self.optimizer))
self._train_sgd(epochs, display_every, uncertainty)
self.callbacks.on_train_end()
print("")
display.training_display.summary(self.train_state)
if print_model:
self._print_model()
if model_save_path is not None:
self.save(model_save_path, verbose=1)
return self.losshistory, self.train_state
def evaluate(self, x, y, callbacks=None):
"""Returns the loss values & metrics values for the model in test mode.
"""
raise NotImplementedError(
"Model.evaluate to be implemented. Alternatively, use Model.predict."
)
@timing
def predict(self, x, operator=None, callbacks=None):
"""Generates output predictions for the input samples.
"""
print("Predicting...")
self.callbacks = CallbackList(callbacks=callbacks)
self.callbacks.set_model(self)
self.callbacks.on_predict_begin()
if operator is None:
y = self.sess.run(
self.net.outputs, feed_dict=self.net.feed_dict(False, False, 2, x)
)
else:
y = self.sess.run(
operator(self.net.inputs, self.net.outputs),
feed_dict=self.net.feed_dict(False, False, 2, x),
)
self.callbacks.on_predict_end()
return y
def _open_tfsession(self):
if self.sess is not None:
return
tfconfig = tf.ConfigProto()
tfconfig.gpu_options.allow_growth = True
self.sess = tf.Session(config=tfconfig)
self.saver = tf.train.Saver(max_to_keep=None)
self.train_state.set_tfsession(self.sess)
def _close_tfsession(self):
self.sess.close()
def _train_sgd(self, epochs, display_every, uncertainty):
for i in range(epochs):
self.callbacks.on_epoch_begin()
self.callbacks.on_batch_begin()
self.train_state.set_data_train(
*self.data.train_next_batch(self.batch_size)
)
self.sess.run(
self.train_op,
feed_dict=self.net.feed_dict(
True, True, 0, self.train_state.X_train, self.train_state.y_train
),
)
self.train_state.epoch += 1
self.train_state.step += 1
if self.train_state.step % display_every == 0 or i + 1 == epochs:
self._test(uncertainty)
self.callbacks.on_batch_end()
self.callbacks.on_epoch_end()
if self.stop_training:
break
def _train_scipy(self, display_every, uncertainty):
def loss_callback(loss_train):
self.train_state.epoch += 1
self.train_state.step += 1
self.train_state.loss_train = loss_train
self.train_state.loss_test = None
self.train_state.metrics_test = None
self.losshistory.append(
self.train_state.step, self.train_state.loss_train, None, None
)
if self.train_state.step % display_every == 0:
display.training_display(self.train_state)
self.train_state.set_data_train(*self.data.train_next_batch(self.batch_size))
self.train_op.minimize(
self.sess,
feed_dict=self.net.feed_dict(
True, True, 0, self.train_state.X_train, self.train_state.y_train
),
fetches=[self.losses],
loss_callback=loss_callback,
)
self._test(uncertainty)
def _test(self, uncertainty):
self.train_state.loss_train, self.train_state.y_pred_train = self.sess.run(
[self.losses, self.net.outputs],
feed_dict=self.net.feed_dict(
False, False, 0, self.train_state.X_train, self.train_state.y_train
),
)
if uncertainty:
# TODO: support multi outputs
losses, y_preds = [], []
for _ in range(1000):
loss_one, y_pred_test_one = self.sess.run(
[self.losses, self.net.outputs],
feed_dict=self.net.feed_dict(
False, True, 1, self.train_state.X_test, self.train_state.y_test
),
)
losses.append(loss_one)
y_preds.append(y_pred_test_one)
self.train_state.loss_test = np.mean(losses, axis=0)
self.train_state.y_pred_test = np.mean(y_preds, axis=0)
self.train_state.y_std_test = np.std(y_preds, axis=0)
else:
self.train_state.loss_test, self.train_state.y_pred_test = self.sess.run(
[self.losses, self.net.outputs],
feed_dict=self.net.feed_dict(
False, False, 1, self.train_state.X_test, self.train_state.y_test
),
)
if isinstance(self.net.targets, (list, tuple)):
self.train_state.metrics_test = [
m(self.train_state.y_test[i], self.train_state.y_pred_test[i])
for m in self.metrics
for i in range(len(self.net.targets))
]
else:
self.train_state.metrics_test = [
m(self.train_state.y_test, self.train_state.y_pred_test)
for m in self.metrics
]
self.train_state.update_best()
self.losshistory.append(
self.train_state.step,
self.train_state.loss_train,
self.train_state.loss_test,
self.train_state.metrics_test,
)
display.training_display(self.train_state)
def _print_model(self):
variables_names = [v.name for v in tf.trainable_variables()]
values = self.sess.run(variables_names)
for k, v in zip(variables_names, values):
print("Variable: {}, Shape: {}".format(k, v.shape))
print(v)
def save(self, save_path, verbose=0):
if verbose > 0:
print(
"Epoch {}: saving model to {}-{} ...\n".format(
self.train_state.epoch, save_path, self.train_state.epoch
)
)
self.saver.save(self.sess, save_path, global_step=self.train_state.epoch)
def restore(self, save_path, verbose=0):
if verbose > 0:
print("Restoring model from {} ...\n".format(save_path))
self.saver.restore(self.sess, save_path)
class TrainState(object):
def __init__(self):
self.epoch, self.step = 0, 0
self.sess = None
# Data
self.X_train, self.y_train = None, None
self.X_test, self.y_test = None, None
# Results of current step
self.y_pred_train = None
self.loss_train, self.loss_test = None, None
self.y_pred_test, self.y_std_test = None, None
self.metrics_test = None
# The best results correspond to the min train loss
self.best_step = 0
self.best_loss_train, self.best_loss_test = np.inf, np.inf
self.best_y, self.best_ystd = None, None
self.best_metrics = None
def set_tfsession(self, sess):
self.sess = sess
def set_data_train(self, X_train, y_train):
self.X_train, self.y_train = X_train, y_train
def set_data_test(self, X_test, y_test):
self.X_test, self.y_test = X_test, y_test
def update_best(self):
if self.best_loss_train > np.sum(self.loss_train):
self.best_step = self.step
self.best_loss_train = np.sum(self.loss_train)
self.best_loss_test = np.sum(self.loss_test)
self.best_y, self.best_ystd = self.y_pred_test, self.y_std_test
self.best_metrics = self.metrics_test
def disregard_best(self):
self.best_loss_train = np.inf
def packed_data(self):
def merge_values(values):
if values is None:
return None
return np.hstack(values) if isinstance(values, (list, tuple)) else values
X_train = merge_values(self.X_train)
y_train = merge_values(self.y_train)
X_test = merge_values(self.X_test)
y_test = merge_values(self.y_test)
best_y = merge_values(self.best_y)
best_ystd = merge_values(self.best_ystd)
return X_train, y_train, X_test, y_test, best_y, best_ystd
class LossHistory(object):
def __init__(self):
self.steps = []
self.loss_train = []
self.loss_test = []
self.metrics_test = []
self.loss_weights = 1
def set_loss_weights(self, loss_weights):
self.loss_weights = loss_weights
def append(self, step, loss_train, loss_test, metrics_test):
self.steps.append(step)
self.loss_train.append(loss_train)
if loss_test is None:
loss_test = self.loss_test[-1]
if metrics_test is None:
metrics_test = self.metrics_test[-1]
self.loss_test.append(loss_test)
self.metrics_test.append(metrics_test)