Published on

Layers on top of TensorFlow

MNIST CNN - Rosetta Stone
Authors

CNN MNIST with sugar-coated TensorFlow

The following is just a collection of code samples for solving CNN MNIST (all using roughly the same network structure) using different layer-helpers on top of regular TensorFlow. In a review post (coming soon), I'll figure out which one makes the most sense to me (as someone who has previous used Theano/Lasagne).

Currently the list is :

  • Plain TensorFlow
  • Keras
  • TF-Slim
  • tf.contrib.learn aka skflow
  • PrettyTensor
  • TFlearn
  • TensorLayer
  • SugarTensor

If you have any suggestions about other sugar-coatings I should consider, please leave a comment.

Plain TensorFlow

When building a simple CNN for MNIST, raw TensorFlow will typically build some helper functions to make the process easier.

The following is from ddigiorg's repo :

sess = tf.InteractiveSession()

x  = tf.placeholder(tf.float32, shape=[None, 784])
y_ = tf.placeholder(tf.float32, shape=[None, 10])

# Helper functions
def weight_variable(shape):
  initial = tf.truncated_normal(shape, stddev=0.1)
  return tf.Variable(initial)

def bias_variable(shape):
  initial = tf.constant(0.1, shape=shape)
  return tf.Variable(initial)

def conv2d(x, W):
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

def max_pool_2x2(x):
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')


# Reshape input
x_image = tf.reshape(x, [-1,28,28,1])

#First Convolutional and Max Pool Layers
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])

h_conv1 = tf.nn.relu(conv2d(x, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)

#Second Convolutional and Max Pool Layers
W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])

h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

#Densely Connected Layer
W_fcl = weight_variable([7 * 7 * 64, 1024])
b_fcl = bias_variable([1024])

h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
h_fcl = tf.nn.relu(tf.matmul(h_pool2_flat, W_fcl) + b_fcl)

#Dropout
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fcl, keep_prob)

#Output Layer (Softmax)
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])

y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

#Train and Evaluate the Model
cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))

sess.run(tf.initialize_all_variables())
for i in range(20000):
    batch = mnist.train.next_batch(50)
    if i % 100 == 0:
        train_accuracy = accuracy.eval(feed_dict = {x:batch[0], y_: batch[1], keep_prob: 1.0})
        print("step %d, training accuracy %g" % (i, train_accuracy))
    train_step.run(feed_dict = {x: batch[0], y_: batch[1], keep_prob: 0.5})

print("test accuracy %g" % accuracy.eval(feed_dict = {x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))

A more interesting example that uses pure TensorFlow is Fast Style Transfer, or this thorough LSTM-based tutorial.

Keras

This is a popular frontend - particularly since it can use both Theano and TensorFlow as a backend.

The flip-side of this flexibility is that one loses the ability to dig into the backend's explicit representation : So adding custom layers is (roughly speaking) no longer possible.

One can see from the model=Sequential(); model.add( fn ) code that Keras has its own environment for formulating the computational graph, before handing it off to whichever backend is chosen.


from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten
from keras.layers import Convolution2D, MaxPooling2D
from keras.utils import np_utils
#from keras import backend as K  # Can detect 'theano' vs 'tensorflow'

X_train = X_train.reshape(X_train.shape[0], img_rows, img_cols, 1)
X_test = X_test.reshape(X_test.shape[0], img_rows, img_cols, 1)
input_shape = (img_rows, img_cols, 1)

# convert class vectors to binary class matrices
Y_train = np_utils.to_categorical(y_train, nb_classes)
Y_test  = np_utils.to_categorical(y_test, nb_classes)

model = Sequential()

model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1],
                        border_mode='valid',
                        input_shape=input_shape))
model.add(Activation('relu'))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1]))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size))
model.add(Dropout(0.25))

model.add(Flatten())
model.add(Dense(128))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(nb_classes))
model.add(Activation('softmax'))

model.compile(loss='categorical_crossentropy',
              optimizer='adadelta',
              metrics=['accuracy'])

model.fit(X_train, Y_train, batch_size=batch_size, nb_epoch=nb_epoch,
          verbose=1, validation_data=(X_test, Y_test))
score = model.evaluate(X_test, Y_test, verbose=0)

TF-Slim

This looks like 'just what the Doctor ordered', until you look into the seq2seq code, and find that the promising looking library has quite a few stubbed functions, which solely execute pass.

From a subdirectory of the main TensorFlow repo :

import tensorflow as tf

slim = tf.contrib.slim

def lenet(images, num_classes=10, is_training=False,
          dropout_keep_prob=0.5, prediction_fn=slim.softmax,
          scope='LeNet'):
  """Creates a variant of the LeNet model.
  Args:
    images: A batch of `Tensors` of size [batch_size, height, width, channels].
    scope: Optional variable_scope.
  Returns:
    logits: the pre-softmax activations, a tensor of size [batch_size, `num_classes`]
    end_points: a dictionary from components of the network to the corresponding activation.
  """
  end_points = {}

  with tf.variable_scope(scope, 'LeNet', [images, num_classes]):
    net = slim.conv2d(images, 32, [5, 5], scope='conv1')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool1')
    net = slim.conv2d(net, 64, [5, 5], scope='conv2')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool2')
    net = slim.flatten(net)
    end_points['Flatten'] = net

    net = slim.fully_connected(net, 1024, scope='fc3')
    net = slim.dropout(net, dropout_keep_prob, is_training=is_training,
                       scope='dropout3')
    logits = slim.fully_connected(net, num_classes, activation_fn=None,
                                  scope='fc4')

  end_points['Logits'] = logits
  end_points['Predictions'] = prediction_fn(logits, scope='Predictions')

  return logits, end_points

lenet.default_image_size = 28

def lenet_arg_scope(weight_decay=0.0):
  """Defines the default lenet argument scope.
  Args:
    weight_decay: The weight decay to use for regularizing the model.
  Returns:
    An `arg_scope` to use for the inception v3 model.
  """
  with slim.arg_scope( [slim.conv2d, slim.fully_connected],
      weights_regularizer=slim.l2_regularizer(weight_decay),
      weights_initializer=tf.truncated_normal_initializer(stddev=0.1),
      activation_fn=tf.nn.relu) as sc:
    return sc

# ...

labels = slim.one_hot_encoding(
    labels, dataset.num_classes - FLAGS.labels_offset)
batch_queue = slim.prefetch_queue.prefetch_queue(
    [images, labels], capacity=2 * deploy_config.num_clones)

# ...

# slim.learning.train( ... )

Also noteworthy : the number of pretrained models available in the slim hierarchy.

tf.contrib.learn aka skflow

Not to be confused with tflearn...

Head of their part of the TensorFlow repo.

Documentation at the main TensorFlow site.

Code from the skflow examples directory :

"""This showcases how simple it is to build image classification networks.
It follows description from this TensorFlow tutorial:
    https://www.tensorflow.org/versions/master/tutorials/mnist/pros/index.html#deep-mnist-for-experts
"""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

from sklearn import metrics
import tensorflow as tf
from tensorflow.contrib import learn

## Download and load MNIST data.
mnist = learn.datasets.load_dataset('mnist')


## Linear classifier.
feature_columns = learn.infer_real_valued_columns_from_input(mnist.train.images)
classifier = learn.LinearClassifier(
    feature_columns=feature_columns, n_classes=10)
classifier.fit(mnist.train.images, mnist.train.labels, batch_size=100,
               steps=1000)
score = metrics.accuracy_score(
    mnist.test.labels, classifier.predict(mnist.test.images))
print('Accuracy: {0:f}'.format(score))


## Convolutional network
def max_pool_2x2(tensor_in):
  return tf.nn.max_pool(
      tensor_in, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

def conv_model(X, y):
  # pylint: disable=invalid-name,missing-docstring
  # reshape X to 4d tensor with 2nd and 3rd dimensions being image width and
  # height final dimension being the number of color channels.
  X = tf.reshape(X, [-1, 28, 28, 1])

  # first conv layer will compute 32 features for each 5x5 patch
  with tf.variable_scope('conv_layer1'):
    h_conv1 = learn.ops.conv2d(X, n_filters=32, filter_shape=[5, 5],
                               bias=True, activation=tf.nn.relu)
    h_pool1 = max_pool_2x2(h_conv1)

  # second conv layer will compute 64 features for each 5x5 patch.
  with tf.variable_scope('conv_layer2'):
    h_conv2 = learn.ops.conv2d(h_pool1, n_filters=64, filter_shape=[5, 5],
                               bias=True, activation=tf.nn.relu)
    h_pool2 = max_pool_2x2(h_conv2)
    # reshape tensor into a batch of vectors
    h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])

  # densely connected layer with 1024 neurons.
  h_fc1 = learn.ops.dnn(
      h_pool2_flat, [1024], activation=tf.nn.relu, dropout=0.5)
  return learn.models.logistic_regression(h_fc1, y)

# Training and predicting.
classifier = learn.TensorFlowEstimator(
    model_fn=conv_model, n_classes=10, batch_size=100, steps=20000,
    learning_rate=0.001)
classifier.fit(mnist.train.images, mnist.train.labels)
score = metrics.accuracy_score(
    mnist.test.labels, classifier.predict(mnist.test.images))
print('Accuracy: {0:f}'.format(score))

PrettyTensor

The main code demonstrates the 'Fluent' style of function chaining, though there are some complaints about documentation.

Nice tutorial walk-through, with code in GitHub :

%matplotlib inline
import tensorflow as tf
import numpy as np

import prettytensor as pt

from tensorflow.examples.tutorials.mnist import input_data
data = input_data.read_data_sets('data/MNIST/', one_hot=True)

# initialisation is standard Tensorflow
x = tf.placeholder(tf.float32, shape=[None, img_size_flat], name='x')
x_image = tf.reshape(x, [-1, img_size, img_size, num_channels])
y_true = tf.placeholder(tf.float32, shape=[None, 10], name='y_true')
y_true_cls = tf.argmax(y_true, dimension=1)


# Now the PrettyTensor magic
x_pretty = pt.wrap(x_image)

with pt.defaults_scope(activation_fn=tf.nn.relu):
    y_pred, loss = x_pretty.\
        conv2d(kernel=5, depth=16, name='layer_conv1').\
        max_pool(kernel=2, stride=2).\
        conv2d(kernel=5, depth=36, name='layer_conv2').\
        max_pool(kernel=2, stride=2).\
        flatten().\
        fully_connected(size=128, name='layer_fc1').\
        softmax_classifier(class_count=10, labels=y_true)

# This is a helper function to extract mid-flow variables for inspection
def get_weights_variable(layer_name):
    # Retrieve an existing variable named 'weights' in the scope
    # with the given layer_name.
    # This is awkward because the TensorFlow function was
    # really intended for another purpose.

    with tf.variable_scope(layer_name, reuse=True):
        variable = tf.get_variable('weights')
    return variable

weights_conv1 = get_weights_variable(layer_name='layer_conv1')
weights_conv2 = get_weights_variable(layer_name='layer_conv2')

# rest of code as before...
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(loss)

TFlearn ( not to be confused with tflearn )

Documentation all online in regular Python Sphinx format.

Complete example from main TFlearn GitHub repo :

from __future__ import division, print_function, absolute_import

import tflearn
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.normalization import local_response_normalization
from tflearn.layers.estimator import regression

# Data loading and preprocessing
import tflearn.datasets.mnist as mnist
X, Y, testX, testY = mnist.load_data(one_hot=True)
X = X.reshape([-1, 28, 28, 1])
testX = testX.reshape([-1, 28, 28, 1])

# Building convolutional network
network = input_data(shape=[None, 28, 28, 1], name='input')
network = conv_2d(network, 32, 3, activation='relu', regularizer="L2")
network = max_pool_2d(network, 2)
network = local_response_normalization(network)
network = conv_2d(network, 64, 3, activation='relu', regularizer="L2")
network = max_pool_2d(network, 2)
network = local_response_normalization(network)
network = fully_connected(network, 128, activation='tanh')
network = dropout(network, 0.8)
network = fully_connected(network, 256, activation='tanh')
network = dropout(network, 0.8)
network = fully_connected(network, 10, activation='softmax')
network = regression(network, optimizer='adam', learning_rate=0.01,
                     loss='categorical_crossentropy', name='target')

# Training
model = tflearn.DNN(network, tensorboard_verbose=0)
model.fit({'input': X}, {'target': Y}, n_epoch=20,
           validation_set=({'input': testX}, {'target': testY}),
           snapshot_step=100, show_metric=True, run_id='convnet_mnist')

TensorLayer

Documentation all online in regular Python Sphinx format.

Partial example from the documentation itself :

network = tl.layers.InputLayer(x, name='input_layer')
network = tl.layers.Conv2dLayer(network,
                        act = tf.nn.relu,
                        shape = [5, 5, 1, 32],  # 32 features for each 5x5 patch
                        strides=[1, 1, 1, 1],
                        padding='SAME',
                        name ='cnn_layer1')     # output: (?, 28, 28, 32)
network = tl.layers.PoolLayer(network,
                        ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1],
                        padding='SAME',
                        pool = tf.nn.max_pool,
                        name ='pool_layer1',)   # output: (?, 14, 14, 32)
network = tl.layers.Conv2dLayer(network,
                        act = tf.nn.relu,
                        shape = [5, 5, 32, 64], # 64 features for each 5x5 patch
                        strides=[1, 1, 1, 1],
                        padding='SAME',
                        name ='cnn_layer2')     # output: (?, 14, 14, 64)
network = tl.layers.PoolLayer(network,
                        ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1],
                        padding='SAME',
                        pool = tf.nn.max_pool,
                        name ='pool_layer2',)   # output: (?, 7, 7, 64)
network = tl.layers.FlattenLayer(network, name='flatten_layer')
                                                # output: (?, 3136)
network = tl.layers.DropoutLayer(network, keep=0.5, name='drop1')
                                                # output: (?, 3136)
network = tl.layers.DenseLayer(network, n_units=256, act = tf.nn.relu, name='relu1')
                                                # output: (?, 256)
network = tl.layers.DropoutLayer(network, keep=0.5, name='drop2')
                                                # output: (?, 256)
network = tl.layers.DenseLayer(network, n_units=10,
                act = tf.identity, name='output_layer')
                                                # output: (?, 10)

y = network.outputs
y_op = tf.argmax(tf.nn.softmax(y), 1)
cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(y, y_))

cost = cost + tl.cost.maxnorm_regularizer(1.0)(network.all_params[0]) +
              tl.cost.maxnorm_regularizer(1.0)(network.all_params[2])

train_params = network.all_params
train_op = tf.train.AdamOptimizer(learning_rate, beta1=0.9, beta2=0.999,
    epsilon=1e-08, use_locking=False).minimize(cost, var_list=train_params)

# ...

feed_dict = {x: X_train_a, y_: y_train_a}
feed_dict.update( network.all_drop )
sess.run(train_op, feed_dict=feed_dict)

# ...

for batch in tl.iterate.minibatches(inputs, targets, batchsize, shuffle=False):
  print(batch)   # ... creates batches for Training
# ...

dp_dict = tl.utils.dict_to_one( network.all_drop )
feed_dict = {x: X_test_a, y_: y_test_a}
feed_dict.update(dp_dict)
err, ac = sess.run([cost, acc], feed_dict=feed_dict)

# ...

correct_prediction = tf.equal(tf.argmax(y, 1), y_)
acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

sugartensor

This jumped out because of the same author used it to reimplement Speech-to-Text-WaveNet.

Documentation is very limited (mainly read-the-code), but the code is in GitHub, and makes the interesting choice of putting itself on the TensorFlow tensor variable structure itself, with the prefix sg_ to avoid polluting the namespace. This is somewhere in between {clever, ingenious, convenient} and nasty (TBD). And it also means that these sugar methods can be chained (like prettytensor, without requiring special wrapping).

Code from main repository README :

import sugartensor as tf

# set log level to debug
tf.sg_verbosity(10)

# MNIST input tensor ( with QueueRunner )
data = tf.sg_data.Mnist()

# inputs
x = data.train.image
y = data.train.label

# create training graph
with tf.sg_context(act='relu', bn=True):
    logit = (x.sg_conv(dim=16).sg_pool()
             .sg_conv(dim=32).sg_pool()
             .sg_conv(dim=32).sg_pool()
             .sg_flatten()
             .sg_dense(dim=256)
             .sg_dense(dim=10, act='linear', bn=False))

# cross entropy loss with logit ( for training set )
loss = logit.sg_ce(target=y)

# accuracy evaluation ( for validation set )
acc = (logit.sg_reuse(input=data.valid.image).sg_softmax()
       .sg_accuracy(target=data.valid.label, name='val'))

# train
tf.sg_train(loss=loss, eval_metric=[acc], ep_size=data.train.num_batch, save_dir='asset/train/conv')