Trainer.py 17.6 KB
Newer Older
1 2 3 4 5 6
#!/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
7 8 9
import threading
import os
import bob.io.base
10
import bob.core
11
from ..analyzers import SoftmaxAnalizer
12
from tensorflow.core.framework import summary_pb2
13
import time
14
from bob.learn.tensorflow.datashuffler import OnlineSampling, TFRecord
15
from bob.learn.tensorflow.utils.session import Session
16
from bob.learn.tensorflow.utils import compute_embedding_accuracy
17
from .learning_rate import constant
18
import time
19

20 21 22 23 24
#logger = bob.core.log.setup("bob.learn.tensorflow")

import logging
logger = logging.getLogger("bob.learn")

25

26 27 28 29 30 31
class Trainer(object):
    """
    One graph trainer.
    Use this trainer when your CNN is composed by one graph

    **Parameters**
32

Tiago Pereira's avatar
Tiago Pereira committed
33 34
    train_data_shuffler:
      The data shuffler used for batching data for training
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
35

Tiago Pereira's avatar
Tiago Pereira committed
36
    iterations:
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
37
      Maximum number of iterations
38

Tiago Pereira's avatar
Tiago Pereira committed
39
    snapshot:
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
40
      Will take a snapshot of the network at every `n` iterations
41

Tiago Pereira's avatar
Tiago Pereira committed
42 43
    validation_snapshot:
      Test with validation each `n` iterations
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
44 45 46 47

    analizer:
      Neural network analizer :py:mod:`bob.learn.tensorflow.analyzers`

Tiago Pereira's avatar
Tiago Pereira committed
48 49 50
    temp_dir: str
      The output directory

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
51
    verbosity_level:
52 53

    """
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
54

55
    def __init__(self,
Tiago Pereira's avatar
Tiago Pereira committed
56
                 train_data_shuffler,
57
                 validation_data_shuffler=None,
58
                 validate_with_embeddings=False,
59

60 61
                 ###### training options ##########
                 iterations=5000,
62 63
                 snapshot=1000,
                 validation_snapshot=2000,
64
                 keep_checkpoint_every_n_hours=2,
65 66

                 ## Analizer
67
                 analizer=SoftmaxAnalizer(),
68

Tiago Pereira's avatar
Tiago Pereira committed
69 70
                 # Temporatu dir
                 temp_dir="cnn",
71

72
                 verbosity_level=2):
73

Tiago Pereira's avatar
Tiago Pereira committed
74
        self.train_data_shuffler = train_data_shuffler
75

76 77
        self.temp_dir = temp_dir

78 79
        self.iterations = iterations
        self.snapshot = snapshot
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
80
        self.validation_snapshot = validation_snapshot
81
        self.keep_checkpoint_every_n_hours = keep_checkpoint_every_n_hours
82

83 84 85
        # Training variables used in the fit
        self.summaries_train = None
        self.train_summary_writter = None
86
        self.thread_pool = None
87 88 89

        # Validation data
        self.validation_summary_writter = None
90
        self.summaries_validation = None
91
        self.validation_data_shuffler = validation_data_shuffler
92

93 94
        # Analizer
        self.analizer = analizer
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
95
        self.global_step = None
96

97
        self.session = None
98

Tiago Pereira's avatar
Tiago Pereira committed
99
        self.graph = None
100 101
        self.validation_graph = None
                
Tiago Pereira's avatar
Tiago Pereira committed
102
        self.loss = None
103
        
Tiago Pereira's avatar
Tiago Pereira committed
104
        self.predictor = None
105 106
        self.validation_predictor = None  
        self.validate_with_embeddings = validate_with_embeddings      
107
        
Tiago Pereira's avatar
Tiago Pereira committed
108 109
        self.optimizer_class = None
        self.learning_rate = None
110

Tiago Pereira's avatar
Tiago Pereira committed
111 112
        # Training variables used in the fit
        self.optimizer = None
113
        
Tiago Pereira's avatar
Tiago Pereira committed
114 115
        self.data_ph = None
        self.label_ph = None
116 117 118 119
        
        self.validation_data_ph = None
        self.validation_label_ph = None
        
Tiago Pereira's avatar
Tiago Pereira committed
120 121
        self.saver = None

122 123
        bob.core.log.set_verbosity_level(logger, verbosity_level)

Tiago Pereira's avatar
Tiago Pereira committed
124 125 126
        # Creating the session
        self.session = Session.instance(new=True).session
        self.from_scratch = True
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
        
    def train(self):
        """
        Train the network        
        Here we basically have the loop for that takes your graph and do a sequence of session.run
        """

        # Creating directories
        bob.io.base.create_directories_safe(self.temp_dir)
        logger.info("Initializing !!")

        # Loading a pretrained model
        if self.from_scratch:
            start_step = 0
        else:
            start_step = self.global_step.eval(session=self.session)

        # TODO: Put this back as soon as possible
        #if isinstance(train_data_shuffler, OnlineSampling):
        #    train_data_shuffler.set_feature_extractor(self.architecture, session=self.session)

        # Start a thread to enqueue data asynchronously, and hide I/O latency.        
        if self.train_data_shuffler.prefetch:
            self.thread_pool = tf.train.Coordinator()
            tf.train.start_queue_runners(coord=self.thread_pool, sess=self.session)
            # In case you have your own queue
            if not isinstance(self.train_data_shuffler, TFRecord):
                threads = self.start_thread()

        # Bootstrapping the summary writters
        self.train_summary_writter = tf.summary.FileWriter(os.path.join(self.temp_dir, 'train'), self.session.graph)
        if self.validation_data_shuffler is not None:
            self.validation_summary_writter = tf.summary.FileWriter(os.path.join(self.temp_dir, 'validation'),
                                                                    self.session.graph)

        ######################### Loop for #################
        for step in range(start_step, start_step+self.iterations):
            # Run fit in the graph
            start = time.time()
            self.fit(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)

            # Running validation
            if self.validation_data_shuffler is not None and step % self.validation_snapshot == 0:
174 175 176 177
                if self.validate_with_embeddings:
                    self.compute_validation_embeddings(step)
                else:
                    self.compute_validation(step)
178 179 180 181 182 183 184 185 186

            # Taking snapshot
            if step % self.snapshot == 0:
                logger.info("Taking snapshot")
                path = os.path.join(self.temp_dir, 'model_snapshot{0}.ckp'.format(step))
                self.saver.save(self.session, path, global_step=step)

        # Running validation for the last time
        if self.validation_data_shuffler is not None:
187 188 189 190 191
            if self.validate_with_embeddings:
                self.compute_validation_embeddings(step)
            else:
                self.compute_validation(step)
            
192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208
            
        logger.info("Training finally finished")

        self.train_summary_writter.close()
        if self.validation_data_shuffler is not None:
            self.validation_summary_writter.close()

        # Saving the final network
        path = os.path.join(self.temp_dir, 'model.ckp')
        self.saver.save(self.session, path)

        if self.train_data_shuffler.prefetch or isinstance(self.train_data_shuffler, TFRecord):
            # now they should definetely stop
            self.thread_pool.request_stop()
            #if not isinstance(self.train_data_shuffler, TFRecord):
            #    self.thread_pool.join(threads)        

Tiago Pereira's avatar
Tiago Pereira committed
209 210
    def create_network_from_scratch(self,
                                    graph,
211
                                    validation_graph=None,
Tiago Pereira's avatar
Tiago Pereira committed
212 213
                                    optimizer=tf.train.AdamOptimizer(),
                                    loss=None,
214

Tiago Pereira's avatar
Tiago Pereira committed
215 216 217 218
                                    # Learning rate
                                    learning_rate=None,
                                    ):

Tiago Pereira's avatar
Tiago Pereira committed
219 220
        """
        Prepare all the tensorflow variables before training.
221
        
Tiago Pereira's avatar
Tiago Pereira committed
222
        **Parameters**
223

Tiago Pereira's avatar
Tiago Pereira committed
224
            graph: Input graph for training
225

Tiago Pereira's avatar
Tiago Pereira committed
226
            optimizer: Solver
227

Tiago Pereira's avatar
Tiago Pereira committed
228
            loss: Loss function
229

Tiago Pereira's avatar
Tiago Pereira committed
230 231 232
            learning_rate: Learning rate
        """

233
        # Getting the pointer to the placeholders
234 235
        self.data_ph = self.train_data_shuffler("data", from_queue=True)
        self.label_ph = self.train_data_shuffler("label", from_queue=True)
236
                
Tiago Pereira's avatar
Tiago Pereira committed
237
        self.graph = graph
238
        self.loss = loss        
239

240 241 242
        # Attaching the loss in the graph
        self.predictor = self.loss(self.graph, self.label_ph)
        
Tiago Pereira's avatar
Tiago Pereira committed
243 244
        self.optimizer_class = optimizer
        self.learning_rate = learning_rate
245
        self.global_step = tf.contrib.framework.get_or_create_global_step()
Tiago Pereira's avatar
Tiago Pereira committed
246

247 248 249 250
        # Preparing the optimizer
        self.optimizer_class._learning_rate = self.learning_rate
        self.optimizer = self.optimizer_class.minimize(self.predictor, global_step=self.global_step)

Tiago Pereira's avatar
Tiago Pereira committed
251
        # Saving all the variables
252 253
        self.saver = tf.train.Saver(var_list=tf.global_variables() + tf.local_variables(), 
                                    keep_checkpoint_every_n_hours=self.keep_checkpoint_every_n_hours)
Tiago Pereira's avatar
Tiago Pereira committed
254

255
        self.summaries_train = self.create_general_summary(self.predictor, self.graph, self.label_ph)
256

257 258
        # SAving some variables
        tf.add_to_collection("global_step", self.global_step)
Tiago Pereira's avatar
Tiago Pereira committed
259 260
        tf.add_to_collection("graph", self.graph)
        tf.add_to_collection("predictor", self.predictor)
261

Tiago Pereira's avatar
Tiago Pereira committed
262 263
        tf.add_to_collection("data_ph", self.data_ph)
        tf.add_to_collection("label_ph", self.label_ph)
264

Tiago Pereira's avatar
Tiago Pereira committed
265 266
        tf.add_to_collection("optimizer", self.optimizer)
        tf.add_to_collection("learning_rate", self.learning_rate)
267

Tiago Pereira's avatar
Tiago Pereira committed
268
        tf.add_to_collection("summaries_train", self.summaries_train)
269

270
        # Same business with the validation
271
        if self.validation_data_shuffler is not None:
272 273 274 275 276
            self.validation_data_ph = self.validation_data_shuffler("data", from_queue=True)
            self.validation_label_ph = self.validation_data_shuffler("label", from_queue=True)

            self.validation_graph = validation_graph

277 278 279 280
            if self.validate_with_embeddings:
                self.validation_predictor = self.validation_graph
            else:
                self.validation_predictor = self.loss(self.validation_graph, self.validation_label_ph)
281 282 283 284 285 286 287 288 289 290

            self.summaries_validation = self.create_general_summary(self.validation_predictor, self.validation_graph, self.validation_label_ph)
            tf.add_to_collection("summaries_validation", self.summaries_validation)
            
            tf.add_to_collection("validation_graph", self.validation_graph)
            tf.add_to_collection("validation_data_ph", self.validation_data_ph)
            tf.add_to_collection("validation_label_ph", self.validation_label_ph)

            tf.add_to_collection("validation_predictor", self.validation_predictor)
            tf.add_to_collection("summaries_validation", self.summaries_validation)
Tiago Pereira's avatar
Tiago Pereira committed
291

Tiago Pereira's avatar
Tiago Pereira committed
292
        # Creating the variables
293
        tf.local_variables_initializer().run(session=self.session)
Tiago Pereira's avatar
Tiago Pereira committed
294 295
        tf.global_variables_initializer().run(session=self.session)

296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313
    def load_checkpoint(self, file_name, clear_devices=True):
        """
        Load a checkpoint

        ** Parameters **

           file_name:
                Name of the metafile to be loaded.
                If a directory is passed, the last checkpoint will be loaded

        """
        if os.path.isdir(file_name):
            checkpoint_path = tf.train.get_checkpoint_state(file_name).model_checkpoint_path
            self.saver = tf.train.import_meta_graph(checkpoint_path + ".meta", clear_devices=clear_devices)
            self.saver.restore(self.session, tf.train.latest_checkpoint(file_name))
        else:
            self.saver = tf.train.import_meta_graph(file_name, clear_devices=clear_devices)
            self.saver.restore(self.session, tf.train.latest_checkpoint(os.path.dirname(file_name)))
314

315
    def create_network_from_file(self, file_name, clear_devices=True):
Tiago Pereira's avatar
Tiago Pereira committed
316
        """
Tiago Pereira's avatar
Tiago Pereira committed
317
        Bootstrap a graph from a checkpoint
Tiago Pereira's avatar
Tiago Pereira committed
318 319 320

         ** Parameters **

Tiago Pereira's avatar
Tiago Pereira committed
321
           file_name: Name of of the checkpoing
Tiago Pereira's avatar
Tiago Pereira committed
322
        """
323 324 325

        logger.info("Loading last checkpoint !!")
        self.load_checkpoint(file_name, clear_devices=True)
Tiago Pereira's avatar
Tiago Pereira committed
326 327

        # Loading training graph
Tiago Pereira's avatar
Tiago Pereira committed
328 329
        self.data_ph = tf.get_collection("data_ph")[0]
        self.label_ph = tf.get_collection("label_ph")[0]
Tiago Pereira's avatar
Tiago Pereira committed
330 331 332 333 334 335 336

        self.graph = tf.get_collection("graph")[0]
        self.predictor = tf.get_collection("predictor")[0]

        # Loding other elements
        self.optimizer = tf.get_collection("optimizer")[0]
        self.learning_rate = tf.get_collection("learning_rate")[0]
337
        self.summaries_train = tf.get_collection("summaries_train")[0]        
Tiago Pereira's avatar
Tiago Pereira committed
338 339
        self.global_step = tf.get_collection("global_step")[0]
        self.from_scratch = False
340 341
        
        # Loading the validation bits
342
        if self.validation_data_shuffler is not None:
343 344 345 346 347 348 349 350 351
            self.summaries_validation = tf.get_collection("summaries_validation")[0]

            self.validation_graph = tf.get_collection("validation_graph")[0]
            self.validation_data_ph = tf.get_collection("validation_data_ph")[0]
            self.validation_label = tf.get_collection("validation_label_ph")[0]

            self.validation_predictor = tf.get_collection("validation_predictor")[0]
            self.summaries_validation = tf.get_collection("summaries_validation")[0]

Tiago Pereira's avatar
Tiago Pereira committed
352 353
    def __del__(self):
        tf.reset_default_graph()
354 355 356

    def get_feed_dict(self, data_shuffler):
        """
357
        Given a data shuffler prepared the dictionary to be injected in the graph
358 359

        ** Parameters **
360 361

            data_shuffler: Data shuffler :py:class:`bob.learn.tensorflow.datashuffler.Base`
362

363
        """
364
        [data, labels] = data_shuffler.get_batch()
365

Tiago Pereira's avatar
Tiago Pereira committed
366 367
        feed_dict = {self.data_ph: data,
                     self.label_ph: labels}
368 369
        return feed_dict

370
    def fit(self, step):
371 372 373 374 375 376 377 378 379
        """
        Run one iteration (`forward` and `backward`)

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

        """

380 381
        if self.train_data_shuffler.prefetch:
            _, l, lr, summary = self.session.run([self.optimizer, self.predictor,
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
382
                                                  self.learning_rate, self.summaries_train])
383 384
        else:
            feed_dict = self.get_feed_dict(self.train_data_shuffler)
385
            _, l, lr, summary = self.session.run([self.optimizer, self.predictor,
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
386
                                                  self.learning_rate, self.summaries_train], feed_dict=feed_dict)
387

388 389
        logger.info("Loss training set step={0} = {1}".format(step, l))
        self.train_summary_writter.add_summary(summary, step)
390

391
    def compute_validation(self, step):
Tiago Pereira's avatar
Tiago Pereira committed
392 393 394 395 396 397 398 399 400 401
        """
        Computes the loss in the validation set

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

        """

402 403 404 405 406 407 408 409
        if self.validation_data_shuffler.prefetch:
            l, lr, summary = self.session.run([self.validation_predictor,
                                               self.learning_rate, self.summaries_validation])
        else:
            feed_dict = self.get_feed_dict(self.validation_data_shuffler)
            l, lr, summary = self.session.run([self.validation_predictor,
                                               self.learning_rate, self.summaries_validation],
                                               feed_dict=feed_dict)
Tiago Pereira's avatar
Tiago Pereira committed
410

411 412
        logger.info("Loss VALIDATION set step={0} = {1}".format(step, l))
        self.validation_summary_writter.add_summary(summary, step)               
Tiago Pereira's avatar
Tiago Pereira committed
413

414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437
    def compute_validation_embeddings(self, step):
        """
        Computes the loss in the validation set with embeddings

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

        """
        if self.validation_data_shuffler.prefetch:
            embedding, labels = self.session.run([self.validation_predictor, self.validation_label_ph])
        else:
            feed_dict = self.get_feed_dict(self.validation_data_shuffler)
            embedding, labels = self.session.run([self.validation_predictor, self.validation_label_ph],
                                               feed_dict=feed_dict)
                                               
        accuracy = compute_embedding_accuracy(embedding, labels)
        
        summary = summary_pb2.Summary.Value(tag="accuracy", simple_value=accuracy)
        logger.info("VALIDATION Accuracy set step={0} = {1}".format(step, accuracy))
        self.validation_summary_writter.add_summary(summary_pb2.Summary(value=[summary]), step)               


438
    def create_general_summary(self, average_loss, output, label):
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
439
        """
440
        Creates a simple tensorboard summary with the value of the loss and learning rate
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
441
        """
442
        # Train summary
443
        tf.summary.scalar('loss', average_loss)
444
        tf.summary.scalar('lr', self.learning_rate)        
445 446 447 448 449

        # Computing accuracy
        correct_prediction = tf.equal(tf.argmax(output, 1), label)
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        tf.summary.scalar('accuracy', accuracy)        
450
        return tf.summary.merge_all()
451

452
    def start_thread(self):
Tiago Pereira's avatar
Tiago Pereira committed
453
        """
454 455 456 457
        Start pool of threads for pre-fetching

        **Parameters**
          session: Tensorflow session
Tiago Pereira's avatar
Tiago Pereira committed
458
        """
459

460
        threads = []
461
        for n in range(self.train_data_shuffler.prefetch_threads):
462
            t = threading.Thread(target=self.load_and_enqueue, args=())
463 464 465 466
            t.daemon = True  # thread will close when parent quits
            t.start()
            threads.append(t)
        return threads
467

468
    def load_and_enqueue(self):
Tiago Pereira's avatar
Tiago Pereira committed
469
        """
470
        Injecting data in the place holder queue
471 472 473

        **Parameters**
          session: Tensorflow session
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
474

Tiago Pereira's avatar
Tiago Pereira committed
475
        """
476
        while not self.thread_pool.should_stop():
477
            [train_data, train_labels] = self.train_data_shuffler.get_batch()
478

479 480
            data_ph = self.train_data_shuffler("data", from_queue=False)
            label_ph = self.train_data_shuffler("label", from_queue=False)
481

482 483 484 485
            feed_dict = {data_ph: train_data,
                         label_ph: train_labels}

            self.session.run(self.train_data_shuffler.enqueue_op, feed_dict=feed_dict)
486

487