Trainer.py 13.5 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
from bob.learn.tensorflow.datashuffler.OnlineSampling import OnLineSampling
16
from .learning_rate import constant
17

18
logger = bob.core.log.setup("bob.learn.tensorflow")
19

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
20

21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
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
44
    def __init__(self,
45
46
                 architecture,
                 optimizer=tf.train.AdamOptimizer(),
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
47
48
                 use_gpu=False,
                 loss=None,
49
                 temp_dir="cnn",
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
50

51
                 # Learning rate
52
                 learning_rate=constant(),
53

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
54
                 ###### training options ##########
55
                 convergence_threshold=0.01,
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
56
                 iterations=5000,
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
57
58
                 snapshot=500,
                 validation_snapshot=100,
59
                 prefetch=False,
60
61

                 ## Analizer
62
                 analizer=SoftmaxAnalizer(),
63

64
65
66
                 ### Pretrained model
                 model_from_file="",

67
                 verbosity_level=2):
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
68

69
70
        if not isinstance(architecture, SequenceNetwork):
            raise ValueError("`architecture` should be instance of `SequenceNetwork`")
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
71
72

        self.architecture = architecture
73
        self.optimizer_class = optimizer
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
74
        self.use_gpu = use_gpu
75
76
77
        self.loss = loss
        self.temp_dir = temp_dir

78
79
80
81
        #self.base_learning_rate = base_learning_rate
        self.learning_rate = learning_rate
        #self.weight_decay = weight_decay
        #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
        # Training variables used in the fit
        self.optimizer = None
        self.training_graph = None
        self.train_data_shuffler = None
        self.summaries_train = None
        self.train_summary_writter = None
95
        self.thread_pool = None
96
97
98
99
100

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

101
102
103
104
105
        # Analizer
        self.analizer = analizer

        self.thread_pool = None
        self.enqueue_op = None
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
106
        self.global_step = None
107

108
109
        self.model_from_file = model_from_file

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="", training=True):
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

        # Creating graphs and defining the loss
148
        network_graph = self.architecture.compute_graph(feature_batch, training=training)
149
150
151
152
153
154
        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
        feed_dict = self.get_feed_dict(data_shuffler)
        l = session.run(self.validation_graph, feed_dict=feed_dict)

203
204
205
        if self.validation_summary_writter is None:
            self.validation_summary_writter = tf.train.SummaryWriter(os.path.join(self.temp_dir, 'validation'), session.graph)

206
207
208
209
210
        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))

211
212
213
214
215
    def create_general_summary(self):
        """
        Creates a simple tensorboard summary with the value of the loss and learning rate
        """

216
217
218
219
220
        # 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()

221
    def start_thread(self, session):
222
223
224
225
226
227
228
        """
        Start pool of threads for pre-fetching

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

229
        threads = []
230
231
        for n in range(3):
            t = threading.Thread(target=self.load_and_enqueue, args=(session,))
232
233
234
235
            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
236

237
238
    def load_and_enqueue(self, session):
        """
239
        Injecting data in the place holder queue
240
241
242

        **Parameters**
          session: Tensorflow session
243
        """
244

245
        while not self.thread_pool.should_stop():
246
247
            [train_data, train_labels] = self.train_data_shuffler.get_batch()
            [train_placeholder_data, train_placeholder_labels] = self.train_data_shuffler.get_placeholders()
248

249
250
251
            feed_dict = {train_placeholder_data: train_data,
                         train_placeholder_labels: train_labels}

252
            session.run(self.enqueue_op, feed_dict=feed_dict)
253

254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
    def create_graphs(self, train_data_shuffler, validation_data_shuffler):

        # Creating train graph
        self.training_graph = self.compute_graph(train_data_shuffler, prefetch=self.prefetch, name="train")
        tf.add_to_collection("training_graph", self.training_graph)

        # Creating inference graph
        self.architecture.compute_inference_placeholder(train_data_shuffler.deployment_shape)
        self.architecture.compute_inference_graph()
        tf.add_to_collection("inference_placeholder", self.architecture.inference_placeholder)
        tf.add_to_collection("inference_graph", self.architecture.inference_graph)

        if validation_data_shuffler is not None:
            # Creating validation graph
            self.validation_graph = self.compute_graph(validation_data_shuffler, name="validation", training=False)
            tf.add_to_collection("validation_graph", self.validation_graph)

271
272
    def train(self, train_data_shuffler, validation_data_shuffler=None):
        """
273
        Train the network
274
275
276
277
278
        """

        # Creating directory
        bob.io.base.create_directories_safe(self.temp_dir)
        self.train_data_shuffler = train_data_shuffler
279

280
        logger.info("Initializing !!")
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
281

282
283
        config = tf.ConfigProto(log_device_placement=True)
        config.gpu_options.allow_growth = True
284

285
286
        # Pickle the architecture to save
        self.architecture.pickle_net(train_data_shuffler.deployment_shape)
287

288
        with tf.Session(config=config) as session:
289

290
291
292
            # Loading a pretrained model
            if self.model_from_file != "":
                logger.info("Loading pretrained model from {0}".format(self.model_from_file))
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
                saver = self.architecture.load(session, self.model_from_file)
                self.training_graph = tf.get_collection("training_graph")[0]
                self.optimizer = tf.get_collection("optimizer")[0]
                self.learning_rate = tf.get_collection("learning_rate")[0]
            else:
                self.create_graphs(train_data_shuffler, validation_data_shuffler)

                # TODO: find an elegant way to provide this as a parameter of the trainer
                self.global_step = tf.Variable(0, trainable=False)

                # Preparing the optimizer
                self.optimizer_class._learning_rate = self.learning_rate
                self.optimizer = self.optimizer_class.minimize(self.training_graph, global_step=self.global_step)
                tf.add_to_collection("optimizer", self.optimizer)
                tf.add_to_collection("learning_rate", self.learning_rate)

                # Train summary
                self.summaries_train = self.create_general_summary()

                tf.initialize_all_variables().run()

                # Original tensorflow saver object
                saver = tf.train.Saver(var_list=tf.all_variables())
316

317
318
319
            if isinstance(train_data_shuffler, OnLineSampling):
                train_data_shuffler.set_feature_extractor(self.architecture, session=session)

320
            # Start a thread to enqueue data asynchronously, and hide I/O latency.
321
322
323
324
            if self.prefetch:
                self.thread_pool = tf.train.Coordinator()
                tf.train.start_queue_runners(coord=self.thread_pool)
                threads = self.start_thread(session)
325

326
            # TENSOR BOARD SUMMARY
327
            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
328
            for step in range(self.iterations):
329
330
331
332
333
334
335

                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
336
337
                # Running validation
                if validation_data_shuffler is not None and step % self.validation_snapshot == 0:
338
                    self.compute_validation(session, validation_data_shuffler, step)
339

340
341
                    if self.analizer is not None:
                        self.validation_summary_writter.add_summary(self.analizer(
342
                             validation_data_shuffler, self.architecture, session), step)
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
343

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
344
345
346
                # Taking snapshot
                if step % self.snapshot == 0:
                    logger.info("Taking snapshot")
347
348
                    path = os.path.join(self.temp_dir, 'model_snapshot{0}.ckp'.format(step))
                    self.architecture.save(session, saver, path)
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
349

350
351
352
353
354
            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
355

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
356
            # Saving the final network
357
358
            path = os.path.join(self.temp_dir, 'model.ckp')
            self.architecture.save(session, saver, path)
359

360
361
362
363
            if self.prefetch:
                # now they should definetely stop
                self.thread_pool.request_stop()
                self.thread_pool.join(threads)