How do you make TensorFlow + Keras fast with a TFRecord dataset?

Question

What is an example of how to use a TensorFlow TFRecord with a Keras Model and tf.session.run() while keeping the dataset in tensors w/ queue runners?

Below is a snippet that works but it needs the following improvements:

• Use the Model API
• specify an Input()
• Load a dataset from a TFRecord
• Run through a dataset in parallel (such as with a queuerunner)

Here is the snippet, there are several TODO lines indicating what is needed:

from keras.models import Model
import tensorflow as tf
from keras import backend as K
from keras.layers import Dense, Input
from keras.objectives import categorical_crossentropy
from tensorflow.examples.tutorials.mnist import input_data

sess = tf.Session()
K.set_session(sess)

# Can this be done more efficiently than placeholders w/ TFRecords?
img = tf.placeholder(tf.float32, shape=(None, 784))
labels = tf.placeholder(tf.float32, shape=(None, 10))

# TODO: Use Input() 
x = Dense(128, activation='relu')(img)
x = Dense(128, activation='relu')(x)
preds = Dense(10, activation='softmax')(x)
# TODO: Construct model = Model(input=inputs, output=preds)

loss = tf.reduce_mean(categorical_crossentropy(labels, preds))

# TODO: handle TFRecord data, is it the same?
mnist_data = input_data.read_data_sets('MNIST_data', one_hot=True)

train_step = tf.train.GradientDescentOptimizer(0.5).minimize(loss)

sess.run(tf.global_variables_initializer())

# TODO remove default, add queuerunner
with sess.as_default():
    for i in range(1000):
        batch = mnist_data.train.next_batch(50)
        train_step.run(feed_dict={img: batch[0],
                                  labels: batch[1]})
    print(loss.eval(feed_dict={img:    mnist_data.test.images, 
                               labels: mnist_data.test.labels}))

Why is this question relevant?

• For high performance training without going back to python
  • no TFRecord to numpy to tensor conversions
• Keras will soon be part of tensorflow
• Demonstrate how Keras Model() classes can accept tensors for input data correctly.

Here is some starter information for a semantic segmentation problem example:

• example unet Keras model unet.py, happens to be for semantic segmentation.
• Keras + Tensorflow Blog Post
• An attempt at running the unet model a tf session with TFRecords and a Keras model (not working)
• Code to create the TFRecords: tf_records.py
• An attempt at running the unet model a tf session with TFRecords and a Keras model is in densenet_fcn.py (not working)

Answer 1

I don't use tfrecord dataset format so won't argue on the pros and cons, but I got interested in extending Keras to support the same.

github.com/indraforyou/keras_tfrecord is the repository. Will briefly explain the main changes.

Dataset creation and loading

data_to_tfrecord and read_and_decode here takes care of creating tfrecord dataset and loading the same. Special care must be to implement the read_and_decode otherwise you will face cryptic errors during training.

Initialization and Keras model

Now both tf.train.shuffle_batch and Keras Input layer returns tensor. But the one returned by tf.train.shuffle_batch don't have metadata needed by Keras internally. As it turns out, any tensor can be easily turned into a tensor with keras metadata by calling Input layer with tensor param.

So this takes care of initialization:

x_train_, y_train_ = ktfr.read_and_decode('train.mnist.tfrecord', one_hot=True, n_class=nb_classes, is_train=True)

x_train_batch, y_train_batch = K.tf.train.shuffle_batch([x_train_, y_train_],
                                                batch_size=batch_size,
                                                capacity=2000,
                                                min_after_dequeue=1000,
                                                num_threads=32) # set the number of threads here

x_train_inp = Input(tensor=x_train_batch)

Now with x_train_inp any keras model can be developed.

Training (simple)

Lets say train_out is the output tensor of your keras model. You can easily write a custom training loop on the lines of:

loss = tf.reduce_mean(categorical_crossentropy(y_train_batch, train_out))
train_op = tf.train.GradientDescentOptimizer(0.01).minimize(loss)


# sess.run(tf.global_variables_initializer())
sess.run(tf.initialize_all_variables())

with sess.as_default():
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)

    try:
      step = 0
      while not coord.should_stop():
        start_time = time.time()

        _, loss_value = sess.run([train_op, loss], feed_dict={K.learning_phase(): 0})

        duration = time.time() - start_time

        if step % 100 == 0:
          print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value,
                                                     duration))
        step += 1
    except tf.errors.OutOfRangeError:
      print('Done training for %d epochs, %d steps.' % (FLAGS.num_epochs, step))
    finally:
      coord.request_stop()

    coord.join(threads)
    sess.close()

Training (keras style)

One of the features of keras that makes it so lucrative is its generalized training mechanism with the callback functions.

But to support tfrecords type training there are several changes that are need in the fit function

• running the queue threads
• no feeding in batch data through feed_dict
• supporting validation becomes tricky as the validation data will also be coming in through another tensor an different model needs to be internally created with shared upper layers and validation tensor fed in by other tfrecord reader.

But all this can be easily supported by another flag parameter. What makes things messing are the keras features sample_weight and class_weight they are used to weigh each sample and weigh each class. For this in compile() keras creates placeholders (here) and placeholders are also implicitly created for the targets (here) which is not needed in our case the labels are already fed in by tfrecord readers. These placeholders needs to be fed in during session run which is unnecessary in our cae.

So taking into account these changes, compile_tfrecord(here) and fit_tfrecord(here) are the extension of compile and fit and shares say 95% of the code.

They can be used in the following way:

import keras_tfrecord as ktfr

train_model = Model(input=x_train_inp, output=train_out)
ktfr.compile_tfrecord(train_model, optimizer='rmsprop', loss='categorical_crossentropy', out_tensor_lst=[y_train_batch], metrics=['accuracy'])

train_model.summary()

ktfr.fit_tfrecord(train_model, X_train.shape[0], batch_size, nb_epoch=3)
train_model.save_weights('saved_wt.h5')

You are welcome to improve on the code and pull requests.

Answer 2

Update 2018-08-29 this is now directly supported in keras, see the following example:

https://github.com/keras-team/keras/blob/master/examples/mnist_tfrecord.py

Original Answer:

TFRecords are supported by using an external loss. Here are the key lines constructing an external loss:

# tf yield ops that supply dataset images and labels
x_train_batch, y_train_batch = read_and_decode_recordinput(...)

# create a basic cnn
x_train_input = Input(tensor=x_train_batch)
x_train_out = cnn_layers(x_train_input)

model = Model(inputs=x_train_input, outputs=x_train_out)
loss = keras.losses.categorical_crossentropy(y_train_batch, x_train_out)
model.add_loss(loss)

model.compile(optimizer='rmsprop', loss=None)

Here is an example for Keras 2. It works after applying the small patch #7060:

'''MNIST dataset with TensorFlow TFRecords.

Gets to 99.25% test accuracy after 12 epochs
(there is still a lot of margin for parameter tuning).
'''
import os
import copy
import time

import numpy as np

import tensorflow as tf
from tensorflow.python.ops import data_flow_ops
from keras import backend as K
from keras.models import Model
from keras.layers import Dense
from keras.layers import Dropout
from keras.layers import Flatten
from keras.layers import Input
from keras.layers import Conv2D
from keras.layers import MaxPooling2D
from keras.callbacks import EarlyStopping
from keras.callbacks import TensorBoard
from keras.objectives import categorical_crossentropy
from keras.utils import np_utils
from keras.utils.generic_utils import Progbar
from keras import callbacks as cbks
from keras import optimizers, objectives
from keras import metrics as metrics_module

from keras.datasets import mnist

if K.backend() != 'tensorflow':
    raise RuntimeError('This example can only run with the '
                       'TensorFlow backend for the time being, '
                       'because it requires TFRecords, which '
                       'are not supported on other platforms.')


def images_to_tfrecord(images, labels, filename):
    def _int64_feature(value):
        return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))

    def _bytes_feature(value):
        return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

    """ Save data into TFRecord """
    if not os.path.isfile(filename):
        num_examples = images.shape[0]

        rows = images.shape[1]
        cols = images.shape[2]
        depth = images.shape[3]

        print('Writing', filename)
        writer = tf.python_io.TFRecordWriter(filename)
        for index in range(num_examples):
            image_raw = images[index].tostring()
            example = tf.train.Example(features=tf.train.Features(feature={
                'height': _int64_feature(rows),
                'width': _int64_feature(cols),
                'depth': _int64_feature(depth),
                'label': _int64_feature(int(labels[index])),
                'image_raw': _bytes_feature(image_raw)}))
            writer.write(example.SerializeToString())
        writer.close()
    else:
        print('tfrecord %s already exists' % filename)


def read_and_decode_recordinput(tf_glob, one_hot=True, classes=None, is_train=None,
                                batch_shape=[1000, 28, 28, 1], parallelism=1):
    """ Return tensor to read from TFRecord """
    print 'Creating graph for loading %s TFRecords...' % tf_glob
    with tf.variable_scope("TFRecords"):
        record_input = data_flow_ops.RecordInput(
            tf_glob, batch_size=batch_shape[0], parallelism=parallelism)
        records_op = record_input.get_yield_op()
        records_op = tf.split(records_op, batch_shape[0], 0)
        records_op = [tf.reshape(record, []) for record in records_op]
        progbar = Progbar(len(records_op))

        images = []
        labels = []
        for i, serialized_example in enumerate(records_op):
            progbar.update(i)
            with tf.variable_scope("parse_images", reuse=True):
                features = tf.parse_single_example(
                    serialized_example,
                    features={
                        'label': tf.FixedLenFeature([], tf.int64),
                        'image_raw': tf.FixedLenFeature([], tf.string),
                    })
                img = tf.decode_raw(features['image_raw'], tf.uint8)
                img.set_shape(batch_shape[1] * batch_shape[2])
                img = tf.reshape(img, [1] + batch_shape[1:])

                img = tf.cast(img, tf.float32) * (1. / 255) - 0.5

                label = tf.cast(features['label'], tf.int32)
                if one_hot and classes:
                    label = tf.one_hot(label, classes)

                images.append(img)
                labels.append(label)

        images = tf.parallel_stack(images, 0)
        labels = tf.parallel_stack(labels, 0)
        images = tf.cast(images, tf.float32)

        images = tf.reshape(images, shape=batch_shape)

        # StagingArea will store tensors
        # across multiple steps to
        # speed up execution
        images_shape = images.get_shape()
        labels_shape = labels.get_shape()
        copy_stage = data_flow_ops.StagingArea(
            [tf.float32, tf.float32],
            shapes=[images_shape, labels_shape])
        copy_stage_op = copy_stage.put(
            [images, labels])
        staged_images, staged_labels = copy_stage.get()

        return images, labels


def save_mnist_as_tfrecord():
    (X_train, y_train), (X_test, y_test) = mnist.load_data()
    X_train = X_train[..., np.newaxis]
    X_test = X_test[..., np.newaxis]
    images_to_tfrecord(images=X_train, labels=y_train, filename='train.mnist.tfrecord')
    images_to_tfrecord(images=X_test, labels=y_test, filename='test.mnist.tfrecord')


def cnn_layers(x_train_input):
    x = Conv2D(32, (3, 3), activation='relu', padding='valid')(x_train_input)
    x = Conv2D(64, (3, 3), activation='relu')(x)
    x = MaxPooling2D(pool_size=(2, 2))(x)
    x = Dropout(0.25)(x)
    x = Flatten()(x)
    x = Dense(128, activation='relu')(x)
    x = Dropout(0.5)(x)
    x_train_out = Dense(classes,
                        activation='softmax',
                        name='x_train_out')(x)
    return x_train_out


sess = tf.Session()
K.set_session(sess)

save_mnist_as_tfrecord()

batch_size = 100
batch_shape = [batch_size, 28, 28, 1]
epochs = 3000
classes = 10
parallelism = 10

x_train_batch, y_train_batch = read_and_decode_recordinput(
    'train.mnist.tfrecord',
    one_hot=True,
    classes=classes,
    is_train=True,
    batch_shape=batch_shape,
    parallelism=parallelism)

x_test_batch, y_test_batch = read_and_decode_recordinput(
    'test.mnist.tfrecord',
    one_hot=True,
    classes=classes,
    is_train=True,
    batch_shape=batch_shape,
    parallelism=parallelism)


x_batch_shape = x_train_batch.get_shape().as_list()
y_batch_shape = y_train_batch.get_shape().as_list()

x_train_input = Input(tensor=x_train_batch, batch_shape=x_batch_shape)
x_train_out = cnn_layers(x_train_input)
y_train_in_out = Input(tensor=y_train_batch, batch_shape=y_batch_shape, name='y_labels')
cce = categorical_crossentropy(y_train_batch, x_train_out)
train_model = Model(inputs=[x_train_input], outputs=[x_train_out])
train_model.add_loss(cce)

train_model.compile(optimizer='rmsprop',
                    loss=None,
                    metrics=['accuracy'])
train_model.summary()

tensorboard = TensorBoard()

# tensorboard disabled due to Keras bug
train_model.fit(batch_size=batch_size,
                epochs=epochs)  # callbacks=[tensorboard])

train_model.save_weights('saved_wt.h5')

K.clear_session()

# Second Session, pure Keras
(X_train, y_train), (X_test, y_test) = mnist.load_data()
X_train = X_train[..., np.newaxis]
X_test = X_test[..., np.newaxis]
x_test_inp = Input(batch_shape=(None,) + (X_test.shape[1:]))
test_out = cnn_layers(x_test_inp)
test_model = Model(inputs=x_test_inp, outputs=test_out)

test_model.load_weights('saved_wt.h5')
test_model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
test_model.summary()

loss, acc = test_model.evaluate(X_test, np_utils.to_categorical(y_test), classes)
print('\nTest accuracy: {0}'.format(acc))

I've also been working to improve the support for TFRecords in the following issue and pull request:

• #6928 Yield Op support: High Performance Large Datasets via TFRecords, and RecordInput
• #7102 Keras Input Tensor API Design Proposal

Finally, it is possible to use tf.contrib.learn.Experiment to train Keras models in TensorFlow.