Cover image for Neural Network from Scratch Using Tensorflow

Neural Network from Scratch Using Tensorflow

lankinen profile image Lankinen Updated on ・2 min read

In this article I show how to build a neural network from scratch. The example is simple and short to make it easier to understand but I haven’t took any shortcuts to hide details.

Looking for PyTorch version of this same tutorial? Go here.

import tensorflow as tf
import matplotlib.pyplot as plt

First we create some random data. x is just 1-D tensor and the model will predict one value y.

x = tf.Variable([[1.,2.]])

CONSOLE: TensorShape([1, 2])

y = 5.

The parameters are initialized using normal distribution where mean is 0 and variance 1.

def initalize_parameters(size, variance=1.0):
    return tf.Variable((tf.random.normal(size) * variance))

first_layer_output_size = 3

weights_1 = initalize_parameters((x.shape[1], first_layer_output_size))

CONSOLE: <tf.Variable 'Variable:0' shape=(2, 3) dtype=float32, 
                [ 0.0535108 ,  1.1256728 ,  0.19349864],
                [-0.8206305 ,  1.8411716 , -0.18347588]],   

bias_1 = initalize_parameters([1])

CONSOLE: <tf.Variable 'Variable:0' shape=(1,) dtype=float32,
          numpy=array([-1.7967013], dtype=float32)>

weights_2 = initalize_parameters((first_layer_output_size,1))

CONSOLE: <tf.Variable 'Variable:0' shape=(3, 1) dtype=float32,
               [-1.3771404 ],
               [-0.59087867]], dtype=float32)>

bias_2 = initalize_parameters([1])

CONSOLE: <tf.Variable 'Variable:0' shape=(1,) dtype=float32, 
          numpy=array([-0.93876433], dtype=float32)>

The neural network contains two linear functions and one non-linear function between them.

def simple_neural_network(xb):
    # linear (1,2 @ 2,3 = 1,3)
    l1 = xb @ weights_1 + bias_1
    # non-linear
    l2 = tf.math.maximum(l1, tf.Variable([0.]))
    # linear (1,3 @ 3,1 = 1,1)
    l3 = l2 @ weights_2 + bias_2
    return l3

Loss function measures how close the predictions are to the real values.

def loss_func(preds, yb):
    # Mean Squared Error (MSE)
    return tf.math.reduce_mean((preds-yb)**2)

Learning rate reduces gradient making sure parameters are not changed too much in each step.

lr = tf.constant([10E-4])

Training contains three simple steps:

  1. Make prediction
  2. Calculate how good the prediction was compared to the real value (When calculating loss it automatically calculates gradient so we don’t need to think about it)
  3. Update parameters by subtracting gradient times learning rate

The code continues taking steps until the loss is less than or equal to 0.1. Finally it plots the loss change.

losses = []

while(len(losses) == 0 or losses[-1] > 0.1):
    with tf.GradientTape() as tape:
    # 1. predict
        preds = simple_neural_network(x)
        # 2. loss
        loss = loss_func(preds, y)
    dW1, db1, dW2, db2 = tape.gradient(loss, 
                                       [weights_1, bias_1, 
                                        weights_2, bias_2])

    # 3. update parameters
    weights_1.assign_sub(dW1 * lr)
    bias_1.assign_sub(db1 * lr)

    weights_2.assign_sub(dW2 * lr)
    bias_2.assign_sub(db2 * lr)


plt.plot(list(range(len(losses))), losses)
plt.ylabel('loss (MSE)')

Loss plot

It changes a lot how many steps it takes to get to loss under 0.1.

Source Code on Github

Posted on by:


markdown guide

Thanks a lot for the post! Nice showcase of GradientTape for automatic differentiation! 😀