Deep Neural net with forward and back propagation from scratch - Python
This article aims to implement a deep neural network from scratch. We will implement a deep neural network containing two input layers, a hidden layer with four units and one output layer. The implementation will go from scratch and the following steps will be implemented.
Algorithm:
1. Loading and visualizing the input data
2. Deciding the shapes of the Weight and bias matrix
3. Initializing matrix, function to be used
4. Implementing the forward propagation method
5. Implementing the cost calculation
6. Backpropagation and optimizing
7. Prediction and visualisation of the output
The Architecture of the Model:
The architecture of the model has been defined by the figure below where the hidden layer uses the Hyperbolic Tangent as the activation function. In contrast, the output layer, being the classification problem uses the sigmoid function.
Weights and bias:
The weights and the bias used for the layers have to be declared initially. The weights will be declared randomly in order to avoid the same output of all units, while the bias will be initialized to zero. The calculation will be done from the scratch itself and according to the rules given below where W1, W2 and b1, b2 are the weights and bias of first and second layer respectively. Here 'a' stands for the activation function of a particular layer.
Cost Function:
The cost function of the above model will pertain to the cost function used with logistic regression. Hence, in this tutorial we will be using this cost function:
Code: Loading and Visualizing the data
import matplotlib.pyplot as plt
import numpy as np
import sklearn
def sigmoid(x):
s = 1/(1+np.exp(-x))
return s
def load_planar_dataset():
np.random.seed(1)
m = 400 # number of examples
N = int(m/2) # number of points per class
D = 2 # dimensionality
X = np.zeros((m,D)) # data matrix where each row is a single example
Y = np.zeros((m,1), dtype='uint8') # labels vector (0 for red, 1 for blue)
a = 4 # maximum ray of the flower
for j in range(2):
ix = range(N*j,N*(j+1))
t = np.linspace(j*3.12,(j+1)*3.12,N) + np.random.randn(N)*0.2 # theta
r = a*np.sin(4*t) + np.random.randn(N)*0.2 # radius
X[ix] = np.c_[r*np.sin(t), r*np.cos(t)]
Y[ix] = j
X = X.T
Y = Y.T
return X, Y
X, Y = load_planar_dataset()
# Visualize the data:
plt.scatter(X[0, :], X[1, :], c = Y, s = 40, cmap = plt.cm.Spectral);
Code: Initializing the Weight and bias matrix
Here, the number of hidden units are four, so, the W1 weight matrix will be of shape (4, number of features) and bias matrix will be of shape (4, 1) which after broadcasting will add up to the weight matrix according to the formula mentioned above. Same procedure can be applied to the W2.
# X --> input dataset of shape (input size, number of examples)
# Y --> labels of shape (output size, number of examples)
W1 = np.random.randn(4, X.shape[0]) * 0.01
b1 = np.zeros(shape =(4, 1))
W2 = np.random.randn(Y.shape[0], 4) * 0.01
b2 = np.zeros(shape =(Y.shape[0], 1))
Code: Forward Propagation :
Now, we will perform the forward propagation using the W1, W2 and the bias b1, b2. In this step the corresponding outputs are calculated in the function defined as forward_propagation.
def forward_propagation(X, W1, W2, b1, b2):
Z1 = np.dot(W1, X) + b1
A1 = np.tanh(Z1)
Z2 = np.dot(W2, A1) + b2
A2 = sigmoid(Z2)
# here the cache is the data of previous iteration
# This will be used for backpropagation
cache = {"Z1": Z1,
"A1": A1,
"Z2": Z2,
"A2": A2}
return A2, cache
Code: Defining the cost function :
# Here Y is actual output
def compute_cost(A2, Y):
m = Y.shape[1]
# implementing the above formula
cost = -(1/m)*np.sum(Y*np.log(A2) + (1-Y)*np.log(1-A2))
# Squeezing to avoid unnecessary dimensions
cost = np.squeeze(cost)
return cost
Code: Back-propagating function:
This is a crucial step as it involves a lot of linear algebra for implementation of backpropagation of the deep neural networks. The Formulas for finding the derivatives can be derived with some mathematical concept of linear algebra, which we are not going to derive here. Just keep in mind that dZ, dW, db are the derivatives of the Cost function w.r.t Weighted sum, Weights and Bias of the layers.
def back_propagation(W1, b1, W2, b2, cache, learning_rate):
# Retrieve also A1 and A2 from dictionary "cache"
A1 = cache['A1']
A2 = cache['A2']
m = Y.shape[1]
# Backward propagation: calculate dW1, db1, dW2, db2.
dZ2 = A2 - Y
dW2 = (1 / m) * np.dot(dZ2, A1.T)
db2 = (1 / m) * np.sum(dZ2, axis = 1, keepdims = True)
dZ1 = np.multiply(np.dot(W2.T, dZ2), 1 - np.power(A1, 2))
dW1 = (1 / m) * np.dot(dZ1, X.T)
db1 = (1 / m) * np.sum(dZ1, axis = 1, keepdims = True)
# Updating the parameters according to algorithm
W1 = W1 - learning_rate * dW1
b1 = b1 - learning_rate * db1
W2 = W2 - learning_rate * dW2
b2 = b2 - learning_rate * db2
return W1, W2, b1, b2
Code: Training the custom model
Now we will train the model using the functions defined above, the epochs can be put as per the convenience and power of the processing unit.
# Please note that the weights and bias are global
# Here iterations is epochs
iterations = 10000
learning_rate = 0.01
for i in range(0, iterations):
# Forward propagation. Inputs: "X, parameters". return: "A2, cache".
A2, cache = forward_propagation(X, W1, W2, b1, b2)
# Cost function. Inputs: "A2, Y". Outputs: "cost".
cost = compute_cost(A2, Y)
# Backpropagation. Inputs: "parameters, cache, X, Y". Outputs: "grads".
W1, W2, b1, b2 = back_propagation(W1, b1, W2, b2, cache, learning_rate)
# Print the cost every 1000 iterations
if(i%(iterations/10) == 0):
print("cost after ", i, "iteration is : ", cost)
Output with learnt parameters
After training the model, take the weights and predict the outcomes using the forward_propagation function above then use the values to plot the figure of output. You will have similar output.
Conclusion:
Deep Learning is a world in which the thrones are captured by the ones who get to the basics, so, try to develop the basics so strong that afterwards, you may be the developer of a new architecture of models which may revolutionalize the community.
