TensorBuilder is a TensorFlow-based library that enables you to easily create complex neural networks using functional programming.
##Import
For demonstration purposes we will import right now everything we will need for the rest of the exercises like this
from tensorbuilder.api import *
import tensorflow as tfbut you can also import just what you need from the tensorbuilder module.
With the T object you can create quick math-like lambdas using any operator, this lets you write things like
x, b = tf.placeholder('float'), tf.placeholder('float')
f = (T + b) / (T + 10) #lambda x: (x + b) / (x + 10)
y = f(x)
assert "div" in y.nameUse function composition with the >> operator to improve readability
x, w, b = tf.placeholder('float', [None, 5]), tf.placeholder('float', [5, 3]), tf.placeholder('float', [3])
f = T.matmul(w) >> T + b >> T.sigmoid()
y = f(x)
assert "Sigmoid" in y.nameAny function from the tf and nn modules is a method from the T object, as before you can use the >> operator or you can chain them to produce complex functions
x, w, b = tf.placeholder('float', [None, 5]), tf.placeholder('float', [5, 3]), tf.placeholder('float', [3])
f = T.matmul(w).add(b).sigmoid()
y = f(x)
assert "Sigmoid" in y.nameYou can use functions from the tf.contrib.layers module via the T.layers property. Here we will use Pipe to apply a value directly to an expression:
x = tf.placeholder('float', [None, 5])
y = Pipe(
x,
T.layers.fully_connected(64, activation_fn=tf.nn.sigmoid) # sigmoid layer 64
.layers.fully_connected(32, activation_fn=tf.nn.tanh) # tanh layer 32
.layers.fully_connected(16, activation_fn=None) # linear layer 16
.layers.fully_connected(8, activation_fn=tf.nn.relu) # relu layer 8
)
assert "Relu" in y.nameHowever, since it is such a common task to build fully_connected layers using the different functions from the tf.nn module, we've (dynamically) create all combination of these as their own methods so you con rewrite the previous as
x = tf.placeholder('float', [None, 5])
y = Pipe(
x,
T.sigmoid_layer(64) # sigmoid layer 64
.tanh_layer(32) # tanh layer 32
.linear_layer(16) # linear layer 16
.relu_layer(8) # relu layer 8
)
assert "Relu" in y.nameThe latter is much more compact, English readable, and reduces a lot of noise.
Coming soon!
Coming soon!
Coming soon!
Coming soon!
Tensor Builder assumes you have a working tensorflow installation. We don't include it in the requirements.txt since the installation of tensorflow varies depending on your setup.
pip install tensorbuilder
For the latest development version
pip install git+https://github.com/cgarciae/tensorbuilder.git@develop
Create neural network with a [5, 10, 3] architecture with a softmax output layer and a tanh hidden layer through a Builder and then get back its tensor:
import tensorflow as tf
from tensorbuilder import T
x = tf.placeholder(tf.float32, shape=[None, 5])
keep_prob = tf.placeholder(tf.float32)
h = T.Pipe(
x,
T.tanh_layer(10) # tanh(x * w + b)
.dropout(keep_prob) # dropout(x, keep_prob)
.softmax_layer(3) # softmax(x * w + b)
)Comming Soon!
Comming Soon!
Comming Soon!
Next is an example with all the features of TensorBuilder including the DSL, branching and scoping. It creates a branched computation where each branch is executed on a different device. All branches are then reduced to a single layer, but the computation is the branched again to obtain both the activation function and the trainer.
import tensorflow as tf
from tensorbuilder import T
x = placeholder(tf.float32, shape=[None, 10])
y = placeholder(tf.float32, shape=[None, 5])
[activation, trainer] = T.Pipe(
x,
[
T.With( tf.device("/gpu:0"):
T.relu_layer(20)
)
,
T.With( tf.device("/gpu:1"):
T.sigmoid_layer(20)
)
,
T.With( tf.device("/cpu:0"):
T.tanh_layer(20)
)
],
T.linear_layer(5),
[
T.softmax() # activation
,
T
.softmax_cross_entropy_with_logits(y) # loss
.minimize(tf.train.AdamOptimizer(0.01)) # trainer
]
)