Permalink
7a2c3a2 Jul 17, 2017
245 lines (177 sloc) 11.9 KB

Getting Started with TensorFlow™ in DIGITS

Table of Contents

Enabling Support For TensorFlow In DIGITS

DIGITS will automatically enable support for TensorFlow if it detects that TensorFlow-gpu is installed in the system. This is done by a line of python code that attempts to import tensorflow to see if it actually imports.

If DIGITS cannot enable tensorflow, a message will be printed in the console saying: TensorFlow support is disabled

Selecting TensorFlow When Creating A Model In DIGITS

Click on the "TensorFlow" tab on the model creation page

Select TensorFlow

Defining A TensorFlow Model In DIGITS

To define a TensorFlow model in DIGITS, you need to write a python class that follows this basic template

class UserModel(Tower):

    @model_propertyOther TensorFlow Tools in DIGITS
    def inference(self):
        # Your code here
        return model

    @model_property#with tf.variable_scope(digits.GraphKeys.MODEL, reuse=None):
    def loss(self):
        # Your code here
        return loss

For example, this is what it looks like for LeNet-5, a model that was created for the classification of hand written digits by Yann Lecun:

class UserModel(Tower):

    @model_property
    def inference(self):
        x = tf.reshape(self.x, shape=[-1, self.input_shape[0], self.input_shape[1], self.input_shape[2]])
        # scale (divide by MNIST std)
        x = x * 0.0125
        with slim.arg_scope([slim.conv2d, slim.fully_connected],
                            weights_initializer=tf.contrib.layers.xavier_initializer(),
                            weights_regularizer=slim.l2_regularizer(0.0005) ):
            model = slim.conv2d(x, 20, [5, 5], padding='VALID', scope='conv1')
            model = slim.max_pool2d(model, [2, 2], padding='VALID', scope='pool1')
            model = slim.conv2d(model, 50, [5, 5], padding='VALID', scope='conv2')
            model = slim.max_pool2d(model, [2, 2], padding='VALID', scope='pool2')
            model = slim.flatten(model)
            model = slim.fully_connected(model, 500, scope='fc1')
            model = slim.dropout(model, 0.5, is_training=self.is_training, scope='do1')
            model = slim.fully_connected(model, self.nclasses, activation_fn=None, scope='fc2')
            return model

    @model_property
    def loss(self):
        loss = digits.classification_loss(self.inference, self.y)
        accuracy = digits.classification_accuracy(self.inference, self.y)
        self.summaries.append(tf.summary.scalar(accuracy.op.name, accuracy))
        return loss

The properties inference and loss must be defined and the class must be called UserModel and it must inherit Tower. This is how DIGITS will interact with the python code.

Provided Properties

Properties that are accessible through self

Property name Type Description
nclasses number Number of classes (for classification datasets). For other type of datasets, this is undefined
input_shape Tensor Shape (1D Tensor) of the first input Tensor. For image data, this is set to height, width, and channels accessible by [0], [1], and [2] respectively.
is_training boolean Whether this is a training graph
is_inference boolean Whether this graph is created for inference/testing
x Tensor The input node, with the shape of [N, H, W, C]
y Tensor The label, [N] for scalar labels, [N, H, W, C] otherwise. Defined only if self.is_training is True

Internal Properties

These properties are in the UserModel class written by the user

Property name Return Type Description
__init()__ None The constructor for the UserModel class
inference() Tensor Called during training and inference
loss() Tensor Called during training to determine the loss and variables to train

Tensors

The network are fed with TensorFlow Tensor objects that are in [N, H, W, C] format.

Other TensorFlow Tools in DIGITS

DIGITS provides a few useful tools to help with your development with TensorFlow.

Provided Helpful Functions

DIGITS provides a few helpful functions to help you with creating the model. Here are the functions we provide inside the digits class

Function Name Parameters Description
classification_loss pred - the images to be classified
y - the labels
Used for classification training to calculate the loss of image classification
mse_loss lhs - left hand tensor
rhs - right hand tensor
Used for calculating the mean square loss between 2 tensors
constrastive_loss lhs - left hand tensor
rhs - right hand tensor
y - labels
Calculates the contrastive loss with respect to the Caffe definition
classification_accuracy pred - the image to be classified
y - the labels
Used to measure how accurate the classification task is
nhwc_to_nchw x - the tensor to transpose Transpose the tensor that was originally NHWC format to NCHW. The tensor must be a degree of 4
nchw_to_nhwc x - the tensor to transpose Transpose the tensor that was originally NCHW format to NHWC. The tensor must be a degree of 4
hwc_to_chw x - the tensor to transpose Transpose the tensor that was originally HWC format to CHW. The tensor must be a degree of 3
chw_to_hwc x - the tensor to transpose Transpose the tensor that was originally CHW format to HWC. The tensor must be a degree of 3
bgr_to_rgb x - the tensor to transform Transform the tensor that was originally in BGR channels to RGB.
rgb_to_bgr x - the tensor to transform Transform the tensor that was originally in RGB channels to BGR.

Visualization With TensorBoard

TensorBoard

TensorBoard is a visualization tools provided by TensorFlow to see the graph of your neural network. DIGITS provides easy access to TensorBoard network visualization for your network while creating it. This can be accessed by clicking on the Visualize button under Custom Network as seen in the image below.

Visualize TensorBoard

If there is something wrong with the network model, DIGITS will automatically provide with you the stacktrace and the error message to help you understand where the problem is.

You can also spin up the full Tensorboard server while your model is training with the command

$ tensorboard --logdir <job_dir>/tb/

where <job_dir> is the directory where them model is being trained at, which can be found here:

Job Dir

Afterwards, you can open up the Tensorboard page by going to http://localhost:6006

Or you can click the Tensorboard link under Visualization

Visualize Button

To know more about how TensorBoard works, its official documentation is availabile in the official tensorflow documentaton

Examples

Simple Auto-Encoder Network

The following network is a simple auto encoder to demostate the structure of how to use tensorflow in DIGITS. An auto encoder is a 2 part network that basically acts as a compression mechanism. The first part will try to compress an image to a size smaller than original while the second part will try to decompress the compressed representation created by the compression network.

class UserModel(Tower):

    @model_property
    def inference(self):

        # the order for input shape is [0] -> H, [1] -> W, [2] -> C
        # this is because tensorflow's default order is NHWC
        model = tf.reshape(self.x, shape=[-1, self.input_shape[0], self.input_shape[1], self.input_shape[2]])
        image_dim = self.input_shape[0] * self.input_shape[1]

        with slim.arg_scope([slim.fully_connected], 
                        weights_initializer=tf.contrib.layers.xavier_initializer(),
                        weights_regularizer=slim.l2_regularizer(0.0005)):

            # first we reshape the images to something
            model = tf.reshape(_x, shape=[-1, image_dim])

            # encode the image
            model = slim.fully_connected(model, 300, scope='fc1')
            model = slim.fully_connected(model, 50, scope='fc2')

            # decode the image
            model = slim.fully_connected(model, 300, scope='fc3')
            model = slim.fully_connected(model, image_dim, activation_fn=None, scope='fc4')

            # form it back to the original
            model = tf.reshape(model, shape=[-1, self.input_shape[0], self.input_shape[1], self.input_shape[2]])
            
            return model

    @model_property
    def loss(self):

        # In an autoencoder, we compare the encoded and then decoded image with the original
        original = tf.reshape(self.x, shape=[-1, self.input_shape[0], self.input_shape[1], self.input_shape[2]])

        # self.inference is called to get the processed image
        model = self.inference
        loss = digits.mse_loss(original, model)

        return loss

Freezing Variables in Pre-Trained Models by Renaming

The following is a demonstration of how to specifying which weights we would like to use for training. This works best if we are using a pre-trained model. This is applicable for fine tuning a model.

When you originally train a model, tensorflow will save the variables with their specified names. When you reload the model to retrain it, tensorflow will simutainously reload all those variables and mark them available to retrain if they are specified in the model definition. When you change the name of the variables in the model, tensorflow will then know to not train that variable and thus "freezes" it.

class UserModel(Tower):

    @model_property
    def inference(self):

        model = construct_model()
        """code to construct the network omitted"""

        # assuming the original model have weight2 and bias2 variables
        # in here, we renamed them by adding the suffix _not_in_use
        # this tells TensorFlow that these variables in the pre-trained model should
        # not be retrained and it should be frozen
        # If we would like to freeze a weight, all we have to do is just rename it
        self.weights = {
            'weight1': tf.get_variable('weight1', [5, 5, self.input_shape[2], 20], initializer=tf.contrib.layers.xavier_initializer()),
            'weight2': tf.get_variable('weight2_not_in_use', [5, 5, 20, 50], initializer=tf.contrib.layers.xavier_initializer())
        }

        self.biases = {
            'bias1': tf.get_variable('bias1', [20], initializer=tf.constant_initializer(0.0)),
            'bias2': tf.get_variable('bias2_not_in_use', [50], initializer=tf.constant_initializer(0.0))
        }

        return model

    @model_property
    def loss(self):
        loss = calculate_loss()
        """code to calculate loss omitted"""
        return loss