Migration from numpy to tensorflow

23 The steps below provide guidance on migrating numpy code to Tensorflow. This is useful when wanting to run optimization problems while making use of parallelization & GPUs.

Import TensorFlow

import tensorflow as tf

Convert Variables to TensorFlow Variables

Only floats should be converted to Tensorflow variables. Ints work well with native tensorflow types

# Numpy
x = np.array([1.0, 2.0, 3.0])
 
# TensorFlow
x = tf.Variable([1.0, 2.0, 3.0])

Convert Operations to TensorFlow Operations

Replace numpy operations with Tensorflow operations.

# Numpy
y = np.square(x)
 
# TensorFlow
y = tf.square(x)

Use TensorFlow Optimizers

Use TensorFlow optimizers instead of custom optimization algorithms.

# Numpy
learning_rate = 0.01
for i in range(1000):
    gradients = compute_gradients()  # Custom gradient computation
    x.assign(x - learning_rate * gradients)
 
# TensorFlow
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01)
train_op = optimizer.minimize(loss)  # Assuming you have a loss function defined
 
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for i in range(1000):
        sess.run(train_op)

Compute Loss

Define your loss function using TensorFlow operations.

# Numpy
def compute_loss(x, y):
    return np.sum((x - y)**2)
 
# TensorFlow
def compute_loss(x, y):
    return tf.reduce_sum(tf.square(x - y))

Run Operations in a Session

TensorFlow operations need to be run within a session.

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    # Run your computations here

Ensure GPU Usage

To make use of GPUs, ensure you have the necessary drivers installed and TensorFlow is built with GPU support. You can specify the device explicitly using with tf.device('/gpu:0').

# Checking GPU availibility
tf.config.list_physical_devices('GPU')
 
# Ensuring the GPU is used
with tf.device('/gpu:0'):
    x = tf.Variable([1.0, 2.0, 3.0])
    y = tf.square(x)