Trainer.py 19.3 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
                 snapshot=1000,
63
                 validation_snapshot=2000,#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

            # Taking snapshot
180
            if step % self.snapshot == 0:            
181 182 183 184 185 186
                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
                                    # Learning rate
                                    learning_rate=None,
217
                                    prelogits=None
Tiago Pereira's avatar
Tiago Pereira committed
218 219
                                    ):

Tiago Pereira's avatar
Tiago Pereira committed
220 221
        """
        Prepare all the tensorflow variables before training.
222
        
Tiago Pereira's avatar
Tiago Pereira committed
223
        **Parameters**
224

Tiago Pereira's avatar
Tiago Pereira committed
225
            graph: Input graph for training
226

Tiago Pereira's avatar
Tiago Pereira committed
227
            optimizer: Solver
228

Tiago Pereira's avatar
Tiago Pereira committed
229
            loss: Loss function
230

Tiago Pereira's avatar
Tiago Pereira committed
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

240 241 242 243 244 245 246
        # TODO: SPECIFIC HACK FOR THE CENTER LOSS. I NEED TO FIND A CLEAN SOLUTION FOR THAT
        self.centers = None
        if prelogits is not None:
            tf.add_to_collection("prelogits", prelogits)
            self.predictor, self.centers = self.loss(self.graph, prelogits, self.label_ph)
        else:
            self.predictor = self.loss(self.graph, self.label_ph)
247
        
Tiago Pereira's avatar
Tiago Pereira committed
248 249
        self.optimizer_class = optimizer
        self.learning_rate = learning_rate
250
        self.global_step = tf.contrib.framework.get_or_create_global_step()
Tiago Pereira's avatar
Tiago Pereira committed
251

252 253 254 255
        # 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
256
        # Saving all the variables
257 258
        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
259

260
        self.summaries_train = self.create_general_summary(self.predictor, self.graph, self.label_ph)
261

262 263
        # SAving some variables
        tf.add_to_collection("global_step", self.global_step)
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
264

265 266
            
        tf.add_to_collection("graph", self.graph)
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
267
        
Tiago Pereira's avatar
Tiago Pereira committed
268
        tf.add_to_collection("predictor", self.predictor)
269

Tiago Pereira's avatar
Tiago Pereira committed
270 271
        tf.add_to_collection("data_ph", self.data_ph)
        tf.add_to_collection("label_ph", self.label_ph)
272

Tiago Pereira's avatar
Tiago Pereira committed
273 274
        tf.add_to_collection("optimizer", self.optimizer)
        tf.add_to_collection("learning_rate", self.learning_rate)
275

Tiago Pereira's avatar
Tiago Pereira committed
276
        tf.add_to_collection("summaries_train", self.summaries_train)
277

278 279 280 281
        # Appending histograms for each trainable variables
        for var in tf.trainable_variables():
            tf.summary.histogram(var.op.name, var)

282
        # Same business with the validation
283
        if self.validation_data_shuffler is not None:
284 285 286 287 288
            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

289
            if self.validate_with_embeddings:            
290
                self.validation_predictor = self.validation_graph
291
            else:            
292
                self.validation_predictor = self.loss(self.validation_graph, self.validation_label_ph)
293 294 295 296 297 298 299 300 301 302

            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
303

Tiago Pereira's avatar
Tiago Pereira committed
304
        # Creating the variables
305
        tf.local_variables_initializer().run(session=self.session)
Tiago Pereira's avatar
Tiago Pereira committed
306 307
        tf.global_variables_initializer().run(session=self.session)

308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325
    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)))
326
            
327
    def load_variables_from_external_model(self, checkpoint_path, var_list):
328 329 330 331 332
        """
        Load a set of variables from a given model and update them in the current one
        
        ** Parameters **
        
333
          checkpoint_path:
334 335 336 337 338 339 340 341 342 343 344 345 346
            Name of the tensorflow model to be loaded
          var_list:
            List of variables to be loaded. A tensorflow exception will be raised in case the variable does not exists
        
        """
        
        assert len(var_list)>0
        
        tf_varlist = []
        for v in var_list:
            tf_varlist += tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=v)

        saver = tf.train.Saver(tf_varlist)
347
        saver.restore(self.session, tf.train.latest_checkpoint(checkpoint_path))
348

349
    def create_network_from_file(self, file_name, clear_devices=True):
Tiago Pereira's avatar
Tiago Pereira committed
350
        """
Tiago Pereira's avatar
Tiago Pereira committed
351
        Bootstrap a graph from a checkpoint
Tiago Pereira's avatar
Tiago Pereira committed
352 353 354

         ** Parameters **

Tiago Pereira's avatar
Tiago Pereira committed
355
           file_name: Name of of the checkpoing
Tiago Pereira's avatar
Tiago Pereira committed
356
        """
357 358 359

        logger.info("Loading last checkpoint !!")
        self.load_checkpoint(file_name, clear_devices=True)
Tiago Pereira's avatar
Tiago Pereira committed
360 361

        # Loading training graph
Tiago Pereira's avatar
Tiago Pereira committed
362 363
        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
364 365 366 367 368 369 370

        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]
371
        self.summaries_train = tf.get_collection("summaries_train")[0]        
Tiago Pereira's avatar
Tiago Pereira committed
372 373
        self.global_step = tf.get_collection("global_step")[0]
        self.from_scratch = False
374 375
        
        # Loading the validation bits
376
        if self.validation_data_shuffler is not None:
377 378 379 380 381 382 383 384 385
            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
386 387
    def __del__(self):
        tf.reset_default_graph()
388 389 390

    def get_feed_dict(self, data_shuffler):
        """
391
        Given a data shuffler prepared the dictionary to be injected in the graph
392 393

        ** Parameters **
394 395

            data_shuffler: Data shuffler :py:class:`bob.learn.tensorflow.datashuffler.Base`
396

397
        """
398
        [data, labels] = data_shuffler.get_batch()
399

Tiago Pereira's avatar
Tiago Pereira committed
400 401
        feed_dict = {self.data_ph: data,
                     self.label_ph: labels}
402 403
        return feed_dict

404
    def fit(self, step):
405 406 407 408 409 410 411 412 413
        """
        Run one iteration (`forward` and `backward`)

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

        """

414
        if self.train_data_shuffler.prefetch:
415 416 417 418 419 420 421 422
            # TODO: SPECIFIC HACK FOR THE CENTER LOSS. I NEED TO FIND A CLEAN SOLUTION FOR THAT        
            if self.centers is None:            
                _, l, lr, summary = self.session.run([self.optimizer, self.predictor,
                                                      self.learning_rate, self.summaries_train])
            else:
                _, l, lr, summary, _ = self.session.run([self.optimizer, self.predictor,
                                                      self.learning_rate, self.summaries_train, self.centers])
            
423 424
        else:
            feed_dict = self.get_feed_dict(self.train_data_shuffler)
425
            _, l, lr, summary = self.session.run([self.optimizer, self.predictor,
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
426
                                                  self.learning_rate, self.summaries_train], feed_dict=feed_dict)
427

428 429
        logger.info("Loss training set step={0} = {1}".format(step, l))
        self.train_summary_writter.add_summary(summary, step)
430

431
    def compute_validation(self, step):
Tiago Pereira's avatar
Tiago Pereira committed
432 433 434 435 436 437 438 439 440 441
        """
        Computes the loss in the validation set

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

        """

442 443 444 445 446 447 448 449
        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
450

451 452
        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
453

454 455 456 457 458 459 460 461 462 463
    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

        """
464
        
465 466 467 468 469 470 471 472 473 474 475 476 477 478
        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)               


479
    def create_general_summary(self, average_loss, output, label):
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
480
        """
481
        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
482
        """
483
        # Train summary
484
        tf.summary.scalar('loss', average_loss)
485
        tf.summary.scalar('lr', self.learning_rate)        
486 487

        # Computing accuracy
488
        correct_prediction = tf.equal(tf.argmax(output, 1), label)
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
489
        
490 491
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        tf.summary.scalar('accuracy', accuracy)        
492
        return tf.summary.merge_all()
493

494
    def start_thread(self):
Tiago Pereira's avatar
Tiago Pereira committed
495
        """
496 497 498 499
        Start pool of threads for pre-fetching

        **Parameters**
          session: Tensorflow session
Tiago Pereira's avatar
Tiago Pereira committed
500
        """
501

502
        threads = []
503
        for n in range(self.train_data_shuffler.prefetch_threads):
504
            t = threading.Thread(target=self.load_and_enqueue, args=())
505 506 507 508
            t.daemon = True  # thread will close when parent quits
            t.start()
            threads.append(t)
        return threads
509

510
    def load_and_enqueue(self):
Tiago Pereira's avatar
Tiago Pereira committed
511
        """
512
        Injecting data in the place holder queue
513 514 515

        **Parameters**
          session: Tensorflow session
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
516

Tiago Pereira's avatar
Tiago Pereira committed
517
        """
518
        while not self.thread_pool.should_stop():
519
            [train_data, train_labels] = self.train_data_shuffler.get_batch()
520

521 522
            data_ph = self.train_data_shuffler("data", from_queue=False)
            label_ph = self.train_data_shuffler("label", from_queue=False)
523

524 525 526 527
            feed_dict = {data_ph: train_data,
                         label_ph: train_labels}

            self.session.run(self.train_data_shuffler.enqueue_op, feed_dict=feed_dict)
528

529