TensorLayer is a Deep Learning (DL) and Reinforcement Learning (RL) library extended from Google TensorFlow. It provides popular DL and RL modules that can be easily customized and assembled for tackling real-world machine learning problems.

Why TensorLayer

TensorLayer grow out from a need to combine the power of TensorFlow with the right building modules for deep neural networks. According to our years of research and practical experiences of tackling real-world machine learning problems, we come up with three design goals for TensorLayer:

• Simplicity: We make TensorLayer easy to work with by providing mass tutorials that can be deployed and run through in minutes. A TensorFlow user may find it easier to bootstrap with the simple, high-level APIs provided by TensorLayer, and then deep dive into their implementation details if need.
• Flexibility: Developing an effective DL algorithm for a specific domain typically requires careful tunings from many aspects. Without the loss of simplicity, TensorLayer allows users to customize their modules by manipulating the native APIs of TensorFlow (e.g., training parameters, iteration control and tensor components).
• Performance: TensorLayer aims to provide zero-cost abstraction for TensorFlow. With its first-class support for TensorFlow, it can easily run on either heterogeneous platforms or multiple computation nodes without compromise in performance.

A frequent question regarding TensorLayer is that why do we develop this library instead of leveraging existing ones like Keras and Tflearn. TensorLayer differentiates with those with its pursuits for flexibility and performance. A DL user may find it comfortable to bootstrap with Keras and Tflearn. These libraries provide high-level abstractions and hide engine implementation details from users. Though good for using, it becomes hard to tune and modify from the bottom, which is necessary when addressing domain-specific problems (i.e., one model does not fit all). Nevertheless, flexibility does not always come with the loss of performance. TensorLayer allows seamless distributed and heterogeneous deployment with its first-class support for the TensorFlow runtime. Besides, TensorLayer can merge with other libraries, such as TF-Slim.

Installation

TensorLayer has install prerequisites including TensorFlow, numpy and matplotlib. For GPU support, CUDA and cuDNN are required. Please check here for detailed instructions.

If you already had the pre-requisites ready, the simplest way to install TensorLayer in your python program is:

[for stable version] pip install tensorlayer==1.2.2
[for master version] pip install git+https://github.com/zsdonghao/tensorlayer.git

Documentation

The documentation [Online] [PDF] [Epub] [HTML] describes the usages of TensorLayer APIs. It is also a self-contained document that walks through different types of deep neural networks, reinforcement learning and their applications in Natural Language Processing (NLP) problems.

We have included the corresponding modularized implementations of Google TensorFlow Deep Learning tutorial, so you can read the TensorFlow tutorial [en] [cn] along with our document.

Chinese documentation is also available.

Your First Program

The first program trains a multi-layer perception network to solve the MNIST problem. We use the well-known scikit-style functions such as fit() and test(). The program is self-explained.

import tensorflow as tf
import tensorlayer as tl

sess = tf.InteractiveSession()

# Prepare data
X_train, y_train, X_val, y_val, X_test, y_test = tl.files.load_mnist_dataset(shape=(-1,784))

# Define placeholder
x = tf.placeholder(tf.float32, shape=[None, 784], name='x')
y_ = tf.placeholder(tf.int64, shape=[None, ], name='y_')

# Define the neural network structure
network = tl.layers.InputLayer(x, name='input_layer')
network = tl.layers.DropoutLayer(network, keep=0.8, name='drop1')
network = tl.layers.DenseLayer(network, n_units=800, act = tf.nn.relu, name='relu1')
network = tl.layers.DropoutLayer(network, keep=0.5, name='drop2')
network = tl.layers.DenseLayer(network, n_units=800, act = tf.nn.relu, name='relu2')
network = tl.layers.DropoutLayer(network, keep=0.5, name='drop3')

# The softmax is implemented internally in tl.cost.cross_entropy(y, y_) to
# speed up computation, so we use identity here.
# see tf.nn.sparse_softmax_cross_entropy_with_logits()
network = tl.layers.DenseLayer(network, n_units=10, act = tf.identity, name='output_layer')
                                
# Define cost function and metric.
y = network.outputs
cost = tl.cost.cross_entropy(y, y_)
correct_prediction = tf.equal(tf.argmax(y, 1), y_)
acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
y_op = tf.argmax(tf.nn.softmax(y), 1)

# Define the optimizer
train_params = network.all_params
train_op = tf.train.AdamOptimizer(learning_rate=0.0001, beta1=0.9, beta2=0.999,
                            epsilon=1e-08, use_locking=False).minimize(cost, var_list=train_params)

# Initialize all variables
sess.run(tf.initialize_all_variables())

# Print network information
network.print_params()
network.print_layers()

# Train the network, we recommend to use tl.iterate.minibatches()
tl.utils.fit(sess, network, train_op, cost, X_train, y_train, x, y_,
            acc=acc, batch_size=500, n_epoch=500, print_freq=5,
            X_val=X_val, y_val=y_val, eval_train=False)

# Evaluation
tl.utils.test(sess, network, acc, X_test, y_test, x, y_, batch_size=None, cost=cost)

# Save the network to .npz file
tl.files.save_npz(network.all_params , name='model.npz')

sess.close()

We provide many helper functions (like fit() , test()) that is similar to Keras to facilitate your development; however, if you want to obtain a fine-grain control over the model or its training process, you can use TensorFlow’s methods like sess.run() in your program directly (tutorial_mnist.py provides more details about this). Many more DL and RL examples can be found here.

Contribution Guideline

TensorLayer is a major ongoing research project in Data Science Institute, Imperial College London. The goal of the project is to develop a compositional language while complex learning systems can be build through composition of neural network modules. The whole development is now participated by numerous contributors here.