Module tf

Bring in all of the public TensorFlow interface.

#Classes: tf.Graph() '''A TensorFlow computation, represented as a dataflow graph.''' tf.device() '''A TensorFlow class to define the device (cpu/gpu) to use ''' tf.Operation() '''Represents a graph node that performs computation on tensors.''' tf.Session()'''A class for running TensorFlow operations.''' #Functions: '''Returns x + y element-wise.''' a = tf.add(x, y) '''Return a tensor with the same shape and contents as input. (name: A name for the operation (optional))''' b = tf.identity(a, name=None) '''Returns the index with the largest value across axes of a tensor example: for a 2D tensor, axis=1 refers to rows''' c = tf.argmax(input, axis=None,name=None) '''Computes square root of x element-wise.''' d = tf.sqrt(x, name=None)

Module tf.data

This module contains different class used to manipulate data. Different files format are supported. It also introduces special classes fitted to TensorFlow for encoding/decoding data.

# Load the training data into two NumPy arrays, for example using `np.load()`. with np.load("/var/data/training_data.npy") as data: features = data["features"] labels = data["labels"] # Assume that each row of `features` corresponds to the same row as `labels`. assert features.shape[0] == labels.shape[0] dataset = tf.data.Dataset.from_tensor_slices((features, labels))

If you choose tfrecord format for encoding your data, you can use the following way to read (parse) your file(s):

# It accepts one or more filenames. filenames = ["/var/data/file1.tfrecord", "/var/data/file2.tfrecord"] dataset = tf.data.TFRecordDataset(filenames) #Apply a transformation function to your data dataset = dataset.map(func)

If your data is contained in text files, you can use the following way to read (parse) it:

# It accepts one or more filenames. filenames = ["/var/data/file1.txt", "/var/data/file2.txt"] dataset = tf.data.TextLineDataset(filenames)

In order to iterate over the dataset, TensorFlow provides the Iterator class:

# The returned iterator will be in an uninitialized state, and you must run the iterator.initializer operation before using it: iterator = dataset.make_initializable_iterator() tf.Session().run(iterator.initializer) #Or use one_shot iterator that will be automatically initialized: iterator = dataset.make_one_shot_iterator()

Module tf.nn

This module introduces basic functions for computer vision and sequential data (ex : text)

# Build a graph:
input_img = tf.placeholder(tf.float32, [batch_size, height, width, channels])
filter_img = tf.Variable(tf.random_uniform([7,7,3,target_channels]))

#strides: , padding can take the values "SAME" to maintain  the same dimension of the input or "VALID"
net = tf.nn.conv2d(input_img, filter_img, strides=[1,2,2,1], padding="SAME")

More functions are available on the following link : https://www.tensorflow.org/api_docs/python/tf/nn

Module tf.metrics

In order to evaluate your model, TensorFlow provides you with different functions that you can call during the computation graph. For each target response, we compare it with the value predicted by the model:

accuracy = tf.metrics.accuracy(labels,predictions)

Calculates how often predictions matches labels. The accuracy function creates two local variables, total and count that are used to compute the frequency with which predictions matches labels. This frequency is ultimately returned as accuracy: an idempotent operation that simply divides total by count.

precision = tf.metrics.precision(labels,predictions)

recall = tf.metrics.recall(labels,predictions)

Most of the functions provided can be applied for a classification task More functions are available on the following link : https://www.tensorflow.org/api_docs/python/tf/metrics

Module tf.train

This module introduces basic classes and functions to monitor your training. In order to train your model, you need to use an optimizer that will handle the backpropagation algorithm. TensorFlow provides many implemantations of famous optimizers:

#We suppose that we have defined the loss function
your_learning_rate = value
optimizer = tf.train.AdamOptimizer(
            learning_rate = your_learning_rate
            )

#Define the training operation:
train_op = optimizer.minimize(loss)

Below a list of the different optimizers, that you can find in the following link: https://www.tensorflow.org/api_docs/python/tf/train

tf.train.AdadeltaOptimizer #M. D. Zeiler
tf.train.AdagradDAOptimizer #John Duchi, Elad Hazan, Yoram Singer
tf.train.MomentumOptimizer

The tf.train module also come with different Session types, that will manage running your operation graph, and provides live information about your training like the number of step, the actual value of your loss function and the actual value(s) of the metric(s) you have chosen. Below an exaustive list:

tf.train.MonitoredSession 
tf.train.SessionManager

Finally, the tf.train module regroup classes and functions to define distributed training over a cluster, and provides tools to monitor it.

Introduction

Import

Module tf

Module tf.data

Module tf.nn

Module tf.metrics

Module tf.train