Trainer.py 11.6 KB
Newer Older
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

17 18
#os.environ["CUDA_VISIBLE_DEVICES"] = "1,2,3,0"
os.environ["CUDA_VISIBLE_DEVICES"] = ""
19

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

    """
46
    def __init__(self,
47 48
                 architecture,
                 optimizer=tf.train.AdamOptimizer(),
49 50
                 use_gpu=False,
                 loss=None,
51
                 temp_dir="cnn",
52

53 54 55 56
                 # Learning rate
                 base_learning_rate=0.001,
                 weight_decay=0.9,

57
                 ###### training options ##########
58
                 convergence_threshold=0.01,
59
                 iterations=5000,
60 61
                 snapshot=100,
                 prefetch=False,
62 63

                 ## Analizer
64
                 analizer=SoftmaxAnalizer(),
65

66
                 verbosity_level=2):
67

68 69
        if not isinstance(architecture, SequenceNetwork):
            raise ValueError("`architecture` should be instance of `SequenceNetwork`")
70 71

        self.architecture = architecture
72
        self.optimizer_class = optimizer
73
        self.use_gpu = use_gpu
74 75 76 77 78
        self.loss = loss
        self.temp_dir = temp_dir

        self.base_learning_rate = base_learning_rate
        self.weight_decay = weight_decay
79 80 81 82

        self.iterations = iterations
        self.snapshot = snapshot
        self.convergence_threshold = convergence_threshold
83
        self.prefetch = prefetch
84

85 86 87 88 89 90 91 92
        # 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
93
        self.thread_pool = None
94 95 96 97 98

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

99 100 101 102 103 104
        # Analizer
        self.analizer = analizer

        self.thread_pool = None
        self.enqueue_op = None

105 106
        bob.core.log.set_verbosity_level(logger, verbosity_level)

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
107 108 109
    def __del__(self):
        tf.reset_default_graph()

110
    def compute_graph(self, data_shuffler, prefetch=False, name=""):
111
        """
112 113
        Computes the graph for the trainer.

114

115 116 117
        ** Parameters **

            data_shuffler: Data shuffler
118
            prefetch:
119 120 121 122
            name: Name of the graph
        """

        # Defining place holders
123
        if prefetch:
124
            [placeholder_data, placeholder_labels] = data_shuffler.get_placeholders_forprefetch(name=name)
125 126 127 128 129 130 131

            # 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
132
            self.enqueue_op = queue.enqueue_many([placeholder_data, placeholder_labels])
133 134 135 136 137 138 139
            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:
140
            [feature_batch, label_batch] = data_shuffler.get_placeholders(name=name)
141 142 143 144 145 146 147 148 149

        # 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):
        """
150
        Given a data shuffler prepared the dictionary to be injected in the graph
151 152 153 154

        ** Parameters **
            data_shuffler:

155
        """
156 157
        [data, labels] = data_shuffler.get_batch()
        [data_placeholder, label_placeholder] = data_shuffler.get_placeholders()
158 159 160 161 162

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

163 164 165 166 167 168 169 170 171 172
    def fit(self, session, step):
        """
        Run one iteration (`forward` and `backward`)

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

        """

173
        if self.prefetch:
174 175
            _, l, lr, summary = session.run([self.optimizer, self.training_graph,
                                             self.learning_rate, self.summaries_train])
176 177 178 179 180
        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)

181 182
        logger.info("Loss training set step={0} = {1}".format(step, l))
        self.train_summary_writter.add_summary(summary, step)
183

184
    def compute_validation(self,  session, data_shuffler, step):
185 186 187 188 189 190 191 192 193
        """
        Computes the loss in the validation set

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

        """
194
        # Opening a new session for validation
195 196 197 198
        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)

199 200 201
        if self.validation_summary_writter is None:
            self.validation_summary_writter = tf.train.SummaryWriter(os.path.join(self.temp_dir, 'validation'), session.graph)

202 203 204 205 206
        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))

207 208 209 210 211
    def create_general_summary(self):
        """
        Creates a simple tensorboard summary with the value of the loss and learning rate
        """

212 213 214 215 216
        # 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()

217
    def start_thread(self, session):
218 219 220 221 222 223 224
        """
        Start pool of threads for pre-fetching

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

225
        threads = []
226 227
        for n in range(3):
            t = threading.Thread(target=self.load_and_enqueue, args=(session,))
228 229 230 231
            t.daemon = True  # thread will close when parent quits
            t.start()
            threads.append(t)
        return threads
232

233 234
    def load_and_enqueue(self, session):
        """
235
        Injecting data in the place holder queue
236 237 238

        **Parameters**
          session: Tensorflow session
239
        """
240

241
        while not self.thread_pool.should_stop():
242 243
            [train_data, train_labels] = self.train_data_shuffler.get_batch()
            [train_placeholder_data, train_placeholder_labels] = self.train_data_shuffler.get_placeholders()
244

245 246 247
            feed_dict = {train_placeholder_data: train_data,
                         train_placeholder_labels: train_labels}

248
            session.run(self.enqueue_op, feed_dict=feed_dict)
249 250 251

    def train(self, train_data_shuffler, validation_data_shuffler=None):
        """
252
        Train the network
253 254 255 256 257
        """

        # Creating directory
        bob.io.base.create_directories_safe(self.temp_dir)
        self.train_data_shuffler = train_data_shuffler
258

259
        # TODO: find an elegant way to provide this as a parameter of the trainer
260
        self.learning_rate = tf.train.exponential_decay(
261 262 263 264 265 266
            self.base_learning_rate,  # Learning rate
            train_data_shuffler.batch_size,
            train_data_shuffler.n_samples,
            self.weight_decay  # Decay step
        )

267
        self.training_graph = self.compute_graph(train_data_shuffler, prefetch=self.prefetch, name="train")
268

269
        # Preparing the optimizer
270
        self.optimizer_class._learning_rate = self.learning_rate
271 272
        #self.optimizer = self.optimizer_class.minimize(self.training_graph, global_step=tf.Variable(0))
        self.optimizer = self.optimizer_class.minimize(self.training_graph)
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
273

274
        # Train summary
275
        self.summaries_train = self.create_general_summary()
276 277

        logger.info("Initializing !!")
278
        # Training
279
        hdf5 = bob.io.base.HDF5File(os.path.join(self.temp_dir, 'model.hdf5'), 'w')
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
280

281 282
        config = tf.ConfigProto(log_device_placement=True)
        config.gpu_options.allow_growth = True
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
283
        with tf.Session(config=config) as session:
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
284

285
            tf.initialize_all_variables().run()
286

287 288 289
            if isinstance(train_data_shuffler, OnLineSampling):
                train_data_shuffler.set_feature_extractor(self.architecture, session=session)

290
            # Start a thread to enqueue data asynchronously, and hide I/O latency.
291 292 293 294
            if self.prefetch:
                self.thread_pool = tf.train.Coordinator()
                tf.train.start_queue_runners(coord=self.thread_pool)
                threads = self.start_thread(session)
295

296
            # TENSOR BOARD SUMMARY
297
            self.train_summary_writter = tf.train.SummaryWriter(os.path.join(self.temp_dir, 'train'), session.graph)
298
            for step in range(self.iterations):
299 300 301 302 303 304 305

                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)

306
                if validation_data_shuffler is not None and step % self.snapshot == 0:
307
                    self.compute_validation(session, validation_data_shuffler, step)
308

309 310
                    if self.analizer is not None:
                        self.validation_summary_writter.add_summary(self.analizer(
311
                             validation_data_shuffler, self.architecture, session), step)
312

313 314 315 316 317
            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
318

319 320 321
            self.architecture.save(hdf5)
            del hdf5

322 323 324 325
            if self.prefetch:
                # now they should definetely stop
                self.thread_pool.request_stop()
                self.thread_pool.join(threads)
326

327
            session.close() # For some reason the session is not closed after the context manager finishes