nn 0.1.0

A neural network library built on top of TensorFlow for quickly building deep learning models.

A neural network library built on top of TensorFlow for quickly building deep learning models.

Usage

nn.Tensor is the core data structure which is a wrapper for tf.Tensor and provides additional functionality. It can be created using the nn.tensor() function:

import nn

a = nn.tensor([1, 2, 3])
assert isinstance(a, nn.Tensor)
assert a.shape == (3, )

It supports method chaining:

c = a.square().sum()
assert c.numpy() == 14

and can be used with tf.Tensor objects:

import tensorflow as tf

b = tf.constant(2)
c = (a - b).square().sum()
assert c.numpy() == 2

It can also be used with high level APIs such as tf.keras:

model = nn.Sequential([
  nn.Dense(128, activation='relu'),
  nn.Dropout(0.2),
  nn.Dense(10)
])

y = model(x)
assert isinstance(y, nn.Tensor)

and to perform automatic differentiation and optimization:

optimizer = nn.Adam()
with nn.GradientTape() as tape:
    outputs = model(inputs)
    loss = (targets - outputs).square().mean()
grads = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))

To use it with ops that expect tf.Tensor objects as inputs, wrap the ops using nn.op():

mean = nn.op(tf.reduce_mean)
c = mean(a)
assert isinstance(c, nn.Tensor)

maximum = nn.op(tf.maximum, binary=True)
c = maximum(a, b)
assert isinstance(c, nn.Tensor)

or convert it to a tf.Tensor object using the tf() method or nn.tf() function:

b = a.tf()
assert isinstance(b, tf.Tensor)

b = nn.tf(a)
assert isinstance(b, tf.Tensor)

See more examples here.

Installation

Requirements:

• TensorFlow >= 2.0
• Python >= 3.6

Install from PyPI (recommended):

pip install nn

Alternatively, install from source:

git clone https://github.com/marella/nn.git
cd nn
pip install -e .

TensorFlow should be installed separately.

Testing

To run tests, install dependencies:

pip install -e .[tests]

and run:

pytest tests