Trainer.py 12.4 KB
Newer Older
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
1
2
3
4
5
6
7
#!/usr/bin/env python
# vim: set fileencoding=utf-8 :
# @author: Tiago de Freitas Pereira <tiago.pereira@idiap.ch>
# @date: Tue 09 Aug 2016 15:25:22 CEST

import tensorflow as tf
from ..network import SequenceNetwork
8
9
10
import threading
import os
import bob.io.base
11
import bob.core
12
from ..analyzers import SoftmaxAnalizer
13
from tensorflow.core.framework import summary_pb2
14
import time
15
16
from bob.learn.tensorflow.datashuffler.OnlineSampling import OnLineSampling

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
17
18
#os.environ["CUDA_VISIBLE_DEVICES"] = "1,3,0,2"
os.environ["CUDA_VISIBLE_DEVICES"] = ""
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
19

20
logger = bob.core.log.setup("bob.learn.tensorflow")
21

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
22

23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
class Trainer(object):
    """
    One graph trainer.
    Use this trainer when your CNN is composed by one graph

    **Parameters**
      architecture: The architecture that you want to run. Should be a :py:class`bob.learn.tensorflow.network.SequenceNetwork`
      optimizer: One of the tensorflow optimizers https://www.tensorflow.org/versions/r0.10/api_docs/python/train.html
      use_gpu: Use GPUs in the training
      loss: Loss
      temp_dir: The output directory

      base_learning_rate: Initial learning rate
      weight_decay:
      convergence_threshold:

      iterations: Maximum number of iterations
      snapshot: Will take a snapshot of the network at every `n` iterations
      prefetch: Use extra Threads to deal with the I/O
      analizer: Neural network analizer :py:mod:`bob.learn.tensorflow.analyzers`
      verbosity_level:

    """
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
46
    def __init__(self,
47
48
                 architecture,
                 optimizer=tf.train.AdamOptimizer(),
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
49
50
                 use_gpu=False,
                 loss=None,
51
                 temp_dir="cnn",
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
52

53
                 # Learning rate
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
54
                 base_learning_rate=0.1,
55
                 weight_decay=0.9,
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
56
                 decay_steps=1000,
57

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
58
                 ###### training options ##########
59
                 convergence_threshold=0.01,
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
60
                 iterations=5000,
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
61
62
                 snapshot=500,
                 validation_snapshot=100,
63
                 prefetch=False,
64
65

                 ## Analizer
66
                 analizer=SoftmaxAnalizer(),
67

68
                 verbosity_level=2):
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
69

70
71
        if not isinstance(architecture, SequenceNetwork):
            raise ValueError("`architecture` should be instance of `SequenceNetwork`")
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
72
73

        self.architecture = architecture
74
        self.optimizer_class = optimizer
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
75
        self.use_gpu = use_gpu
76
77
78
79
80
        self.loss = loss
        self.temp_dir = temp_dir

        self.base_learning_rate = base_learning_rate
        self.weight_decay = weight_decay
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
81
        self.decay_steps = decay_steps
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
82
83
84

        self.iterations = iterations
        self.snapshot = snapshot
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
85
        self.validation_snapshot = validation_snapshot
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
86
        self.convergence_threshold = convergence_threshold
87
        self.prefetch = prefetch
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
88

89
90
91
92
93
94
95
96
        # Training variables used in the fit
        self.optimizer = None
        self.training_graph = None
        self.learning_rate = None
        self.training_graph = None
        self.train_data_shuffler = None
        self.summaries_train = None
        self.train_summary_writter = None
97
        self.thread_pool = None
98
99
100
101
102

        # Validation data
        self.validation_graph = None
        self.validation_summary_writter = None

103
104
105
106
107
        # Analizer
        self.analizer = analizer

        self.thread_pool = None
        self.enqueue_op = None
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
108
        self.global_step = None
109

110
111
        bob.core.log.set_verbosity_level(logger, verbosity_level)

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
112
113
114
    def __del__(self):
        tf.reset_default_graph()

115
    def compute_graph(self, data_shuffler, prefetch=False, name=""):
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
116
        """
117
118
        Computes the graph for the trainer.

119

120
121
122
        ** Parameters **

            data_shuffler: Data shuffler
123
            prefetch:
124
125
126
127
            name: Name of the graph
        """

        # Defining place holders
128
        if prefetch:
129
            [placeholder_data, placeholder_labels] = data_shuffler.get_placeholders_forprefetch(name=name)
130
131
132
133
134
135
136

            # Defining a placeholder queue for prefetching
            queue = tf.FIFOQueue(capacity=10,
                                 dtypes=[tf.float32, tf.int64],
                                 shapes=[placeholder_data.get_shape().as_list()[1:], []])

            # Fetching the place holders from the queue
137
            self.enqueue_op = queue.enqueue_many([placeholder_data, placeholder_labels])
138
139
140
141
142
143
144
            feature_batch, label_batch = queue.dequeue_many(data_shuffler.batch_size)

            # Creating the architecture for train and validation
            if not isinstance(self.architecture, SequenceNetwork):
                raise ValueError("The variable `architecture` must be an instance of "
                                 "`bob.learn.tensorflow.network.SequenceNetwork`")
        else:
145
            [feature_batch, label_batch] = data_shuffler.get_placeholders(name=name)
146
147
148
149
150
151
152
153
154

        # Creating graphs and defining the loss
        network_graph = self.architecture.compute_graph(feature_batch)
        graph = self.loss(network_graph, label_batch)

        return graph

    def get_feed_dict(self, data_shuffler):
        """
155
        Given a data shuffler prepared the dictionary to be injected in the graph
156
157
158
159

        ** Parameters **
            data_shuffler:

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
160
        """
161
162
        [data, labels] = data_shuffler.get_batch()
        [data_placeholder, label_placeholder] = data_shuffler.get_placeholders()
163
164
165
166
167

        feed_dict = {data_placeholder: data,
                     label_placeholder: labels}
        return feed_dict

168
169
170
171
172
173
174
175
176
177
    def fit(self, session, step):
        """
        Run one iteration (`forward` and `backward`)

        ** Parameters **
            session: Tensorflow session
            step: Iteration number

        """

178
        if self.prefetch:
179
180
            _, l, lr, summary = session.run([self.optimizer, self.training_graph,
                                             self.learning_rate, self.summaries_train])
181
182
183
184
185
        else:
            feed_dict = self.get_feed_dict(self.train_data_shuffler)
            _, l, lr, summary = session.run([self.optimizer, self.training_graph,
                                             self.learning_rate, self.summaries_train], feed_dict=feed_dict)

186
187
        logger.info("Loss training set step={0} = {1}".format(step, l))
        self.train_summary_writter.add_summary(summary, step)
188

189
    def compute_validation(self,  session, data_shuffler, step):
190
191
192
193
194
195
196
197
198
        """
        Computes the loss in the validation set

        ** Parameters **
            session: Tensorflow session
            data_shuffler: The data shuffler to be used
            step: Iteration number

        """
199
        # Opening a new session for validation
200
201
202
203
        self.validation_graph = self.compute_graph(data_shuffler, name="validation")
        feed_dict = self.get_feed_dict(data_shuffler)
        l = session.run(self.validation_graph, feed_dict=feed_dict)

204
205
206
        if self.validation_summary_writter is None:
            self.validation_summary_writter = tf.train.SummaryWriter(os.path.join(self.temp_dir, 'validation'), session.graph)

207
208
209
210
211
        summaries = []
        summaries.append(summary_pb2.Summary.Value(tag="loss", simple_value=float(l)))
        self.validation_summary_writter.add_summary(summary_pb2.Summary(value=summaries), step)
        logger.info("Loss VALIDATION set step={0} = {1}".format(step, l))

212
213
214
215
216
    def create_general_summary(self):
        """
        Creates a simple tensorboard summary with the value of the loss and learning rate
        """

217
218
219
220
221
        # Train summary
        tf.scalar_summary('loss', self.training_graph, name="train")
        tf.scalar_summary('lr', self.learning_rate, name="train")
        return tf.merge_all_summaries()

222
    def start_thread(self, session):
223
224
225
226
227
228
229
        """
        Start pool of threads for pre-fetching

        **Parameters**
          session: Tensorflow session
        """

230
        threads = []
231
232
        for n in range(3):
            t = threading.Thread(target=self.load_and_enqueue, args=(session,))
233
234
235
236
            t.daemon = True  # thread will close when parent quits
            t.start()
            threads.append(t)
        return threads
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
237

238
239
    def load_and_enqueue(self, session):
        """
240
        Injecting data in the place holder queue
241
242
243

        **Parameters**
          session: Tensorflow session
244
        """
245

246
        while not self.thread_pool.should_stop():
247
248
            [train_data, train_labels] = self.train_data_shuffler.get_batch()
            [train_placeholder_data, train_placeholder_labels] = self.train_data_shuffler.get_placeholders()
249

250
251
252
            feed_dict = {train_placeholder_data: train_data,
                         train_placeholder_labels: train_labels}

253
            session.run(self.enqueue_op, feed_dict=feed_dict)
254
255
256

    def train(self, train_data_shuffler, validation_data_shuffler=None):
        """
257
        Train the network
258
259
260
261
262
        """

        # Creating directory
        bob.io.base.create_directories_safe(self.temp_dir)
        self.train_data_shuffler = train_data_shuffler
263

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
264
265
266
        # Pickle the architecture to save
        self.architecture.pickle_net(train_data_shuffler.deployment_shape)

267
        # TODO: find an elegant way to provide this as a parameter of the trainer
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
268
        self.global_step = tf.Variable(0, trainable=False)
269
        self.learning_rate = tf.train.exponential_decay(
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
270
271
272
273
274
            learning_rate=self.base_learning_rate,  # Learning rate
            global_step=self.global_step,
            decay_steps=self.decay_steps,
            decay_rate=self.weight_decay,  # Decay step
            staircase=False
275
        )
276
        self.training_graph = self.compute_graph(train_data_shuffler, prefetch=self.prefetch, name="train")
277

278
        # Preparing the optimizer
279
        self.optimizer_class._learning_rate = self.learning_rate
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
280
281
        self.optimizer = self.optimizer_class.minimize(self.training_graph, global_step=self.global_step)

282
        # Train summary
283
        self.summaries_train = self.create_general_summary()
284
285

        logger.info("Initializing !!")
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
286

287
288
        config = tf.ConfigProto(log_device_placement=True)
        config.gpu_options.allow_growth = True
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
289
        with tf.Session(config=config) as session:
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
290
            tf.initialize_all_variables().run()
291

292
293
294
            if isinstance(train_data_shuffler, OnLineSampling):
                train_data_shuffler.set_feature_extractor(self.architecture, session=session)

295
            # Start a thread to enqueue data asynchronously, and hide I/O latency.
296
297
298
299
            if self.prefetch:
                self.thread_pool = tf.train.Coordinator()
                tf.train.start_queue_runners(coord=self.thread_pool)
                threads = self.start_thread(session)
300

301
            # TENSOR BOARD SUMMARY
302
            self.train_summary_writter = tf.train.SummaryWriter(os.path.join(self.temp_dir, 'train'), session.graph)
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
303
            for step in range(self.iterations):
304
305
306
307
308
309
310

                start = time.time()
                self.fit(session, step)
                end = time.time()
                summary = summary_pb2.Summary.Value(tag="elapsed_time", simple_value=float(end-start))
                self.train_summary_writter.add_summary(summary_pb2.Summary(value=[summary]), step)

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
311
312
                # Running validation
                if validation_data_shuffler is not None and step % self.validation_snapshot == 0:
313
                    self.compute_validation(session, validation_data_shuffler, step)
314

315
316
                    if self.analizer is not None:
                        self.validation_summary_writter.add_summary(self.analizer(
317
                             validation_data_shuffler, self.architecture, session), step)
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
318

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
319
320
321
322
323
324
325
                # Taking snapshot
                if step % self.snapshot == 0:
                    logger.info("Taking snapshot")
                    hdf5 = bob.io.base.HDF5File(os.path.join(self.temp_dir, 'model_snapshot{0}.hdf5'.format(step)), 'w')
                    self.architecture.save(hdf5)
                    del hdf5

326
327
328
329
330
            logger.info("Training finally finished")

            self.train_summary_writter.close()
            if validation_data_shuffler is not None:
                self.validation_summary_writter.close()
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
331

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
332
333
            # Saving the final network
            hdf5 = bob.io.base.HDF5File(os.path.join(self.temp_dir, 'model.hdf5'), 'w')
334
335
336
            self.architecture.save(hdf5)
            del hdf5

337
338
339
340
            if self.prefetch:
                # now they should definetely stop
                self.thread_pool.request_stop()
                self.thread_pool.join(threads)
341

342
            session.close() # For some reason the session is not closed after the context manager finishes