Neural network from scratch in Python – ThinkInfi
Neural Network is used everywhere like speech recognition, face recognition, marketing, healthcare, etc. Artificial Neural networks mimic the behavior of human brain and try to solve any given (data-driven) problem like human. Neural Network consists of multiple layers of Perceptrons. When you fed some input data to Neural Network, this data is then passed through those multiple layers of Perceptrons to produce the desired output.
The Perceptron Algorithm | Working,…
Please enable JavaScript
The Perceptron Algorithm | Working, Learning, and Code Implementation
In this tutorial, I will explain each step to train a neural network from scratch in python with a simple example and write Neural Network from Scratch using numpy Python. After reading this tutorial you will have answers for below questions:
- What is Neural Network
- How Neural Network Works
- Steps to build a Neural Network
- How Forward propagation works
- Error Calculation Neural Network
- How Back Propagation works
- Matrix calculation of neural network in Python
Before moving into each step of neural network let me give you an overview of Neural Network Architecture.
Architecture of Neural Network
Architecture of Neural Network
A neural network consists of three layers:
- Input Layer: In this layer, input data needs to feed. Input of input layer goes to the hidden layer
- Hidden Layer: Locate between input and output layer. The input of hidden layer is output of input layer. In the real-world example, there can be multiple hidden layers. To explain neural network, I am using one hidden layer in this article
- Output Layer: Output of hidden layer goes to output layer. This layer generate predicted output of Neural Network. In the above picture and for this article I am considering two class Neural Network (Out y1, Out y2)
Neural Network Formation
Before listing down all equations of a simple neural network, let me clear you that, an artificial neural network equation consist of three things:
- Linear function
- Bias
- Activation function
Output of any layer is the combination of bias and activation function with a linear function.
For example
Input of H1 (or Output of x1) = x1w1+ x2w2 + b1
Here
x1w1+ x2w2 is the linear function
b1 is the bias (constant)
Activation function is required to calculate output of any layer.
Now let’s calculate output of H1
To calculate output of H1 you need to apply activation function to input of H1. You can use any activation function like Sigmoid, Tanh, ReLu, etc. For this tutorial, I am using sigmoid function as my activation function.
Let me show you the equation for the sigmoid function.
So after applying activation function with input of H1, we will get Output of H1
Steps to train Neural Network
There are three steps to train a Neural Network
1.
Forward Propagation
2.
Error Calculation
3.
Back Propagation
Now let’s explore each steps of neural network in detail.
In this tutorial I am denoting
·
h1
(after applying linear function and bias) as input of H1
·
h2
(after applying linear function and bias) as input of H1
·
Out h1
(after applying activation function) as output of H1
·
Out h2
(after applying activation function) as output of H2
·
y1
(after applying linear function and bias) as input of y1 layer
·
y2
(after applying linear function and bias) as input of y2 layer
·
Out y1
(after applying activation function) as output of y1 layer
·
Out y2
(after applying activation function) as output of y2 layer
·
ETotal
as total error of the Neural Network model
Also Read:
Naive Bayes algorithm in Machine Learning with Python
Forward Propagation in Neural Network
Let’s assume we want to apply Neural Network in below dataset, where output (T1 | T2) have two class of probability (for example probability to win and probability to loss)
To explain how neural network works, let’s assume we have only one row of below dataset.
X1X2T (T1/T2)0.030.09(0.01/ 0.99)0.040.1(0.99/ 0.01)0.050.11(0.01/ 0.99)0.060.12(0.99/ 0.01)
In forward propagation of Neural Network left to right directional calculation happens.
1.
First:
Input data (x1, x2) fed into input layer
2.
Second:
Hidden Layer (H1, H2) calculation
3.
Third:
Predict output by output layer (y1, y2) calculation
To calculate forward propagation in hand, let’s take some numbers for weights, bias and target value or actual output (first row output T1 | T2) along with input value (first row of our dataset).
Input ValueBiasWeights
1st
layerWeights
2nd
layerActual/ Target outputx1 = 0.03b1 = 0.39w1 = 0.11w5 = 0.44T1 = 0.01×2 = 0.09b2 = 0.42w2 = 0.27w6 = 0.48T2 = 0.99 w3 = 0.19w7 = 0.23 w4 = 0.52w8 = 0.29
Now let’s start calculation for each step of forward propagation
1. Hidden layer Calculation
In this tutorial I am using two hidden units (
h1, h2
) in hidden layer. Let’s calculate output of those hidden units.
h1 = x1w1+ x2w2 +b1
= 0.03*0.11 + 0.09*0.27 + 0.39 = 0.4176
Now,
In the similar way
h2= x1w3 + x2w4 +b1
= 0.03*0.19 + 0.09*0.52 + 0.39 = 0.4425
So,
Now that we have done with hidden layer calculation, we will move on to output layer.
2. Output layer Calculation
In the similar way of hidden layer:
y1= Outh1 * w5 + Outh2 * w6 + b2
= 0.60290881 * 0.44 + 0.60885456921 * 0.48 + 0.42= 0.97753007
So now,
In the similar way we can calculate:
y2= Outh1 * w7 + Outh2 * w8 + b2
= 0.60290881 * 0.23 + 0.60885457 * 0.29 + 0.42= 0.73523685
And,
Now that we got Out y1 and Out y2 (predicted target value), we will calculate the errornow to find out how accurately our Neural Network algorithm is predicting.
Error Calculation in Neural Network
To find out accuracy of any algorithm error calculation is essential. There are many techniques to calculate error. In this tutorial I am using (or I will calculate) Mean Square Error to find out accuracy.
For example the target value y1 is 0.01 but neural network predicted output (value) for y1(out y1) is 0.72661785, therefore it is an error.
So calculating mean square error for y1
Similarly calculating mean square error for y2
So the total error for our neural network (after one iteration) is the sum of these errors:
ETotal= E1 + E2
= 0.25677057 + 0.04931263 = 0.3060832
Back Propagation in Neural Network
Now that we know how much error our algorithm has (after one iteration). Now we need to improve accuracy (decrease error) of our algorithm. One of the standard ways to improve accuracy is updating weight values. The way to update weight values in Neural Network is called Back Propagation.
In Back Propagation our goal is to update each of the weights (w1, w2, …w8) in the network so that they cause the actual output to be closer the target output. In this way we can minimize the error for each output neuron (y1 and y2)
Now let’s see how back propagation works in Neural Network.
1.Back Propagation for Output Layer
We will apply partial differentiation to get how much changes are required to update w5.
Consider weight w5, we want to know how much changes in weight w5affect the total error of our neural network ().
Note:
is the partial derivative (or gradient) of ETotal with respect to w5
By applying chain rule we can get.
Now,
ETotal = E1+ E2
So,
So,
Now,
So,
= Out y1 (1-Out y1)
= 0.72661785 * (1 – 0.72661785)
= 0.19864435
So,
Again,
y1 = Out h1 *w5 +Out h2 * w6 +b2
So,
So finally,
Update w5with change value
1) to affect total error of our neural network model, to decrease error, we will subtract this change value (
After calculating how much changes required in weight (w) to affect total error of our neural network model, to decrease error, we will subtract this change value (
) from the current weight (w5).
=0.44 – 0.3 * 0.08582533
[Taking
η
= 0.3]
(new)w5 = 0.41425240
In this similar way we can calculate all updated weights or new weights (w6, w7, w8) of output layer.
2.Back Propagation for Hidden Layer
After calculating new values for output layer weights, we will continue backward pass to calculate new values for hidden layer weights (w1, w2, w3, w4)
Consider weight w1
w1 affect the total error of our neural network (
, we want to know how much changes in weightaffect the total error of our neural network (
)
Note
:
Output of each hidden layer neuron (Out h1, Out h2) contributes to the output of each output neurons (Out y1, Out y2) and therefore contributes to the error.
Now,
Now again,
Now,
And,
So finally,
Similarly we can calculate,
Now,
And,
So finally,
So now putting it all together,
Now coming back to the main equation,
Now,
Now,
So now finally putting it all together,
Update w1with change value
1) to affect total error of our neural network model, to decrease error, we will subtract this change value (
After calculating how much changes required in weight (w) to affect total error of our neural network model, to decrease error, we will subtract this change value (
) from the current weight (w1)
In this similar way we can calculate all new weights (w2, w3, w4) of output layer.
******* This is the end of 1st iteration of our Neural Network model *******
It is important to note that the model is not trained properly yet, as we only back-propagated through one sample (first row) from the training set. Doing all we did, all over again for all the samples (each row) will yield a complete model.
Neural Network Matrix Calculation in Python
While applying neural network you should not apply each steps (forward propagation, error calculation, back propagation) for entire dataset sample-by-sample (row by row). It will be then a time consuming process where as you need to repeat same thing for so many times (total number of row of the training dataset).
Instead of that we will calculate each steps of neural network at once (all row once) by doing matrix calculation.
Let me show you each step of neural network from scratch using numpy in Python.
Hidden Layer Matrix Calculation
Denoting
Φ
(phi) as activation function (for this example sigmoid function).
Output Layer Matrix Calculation
Error Calculation
Neural Network in Numpy Python
########################################################################## # Neural Network from Scratch using numpy ########################################################################## import numpy as np # input data x variable x_val = np.array([[0.03, 0.09], [0.04, 0.10], [0.05, 0.11], [0.06, 0.12]]) # output data y variable y_val = np.array([[0.01, 0.99], [0.99, 0.01], [0.01, 0.99], [0.99, 0.01]]) ############################################### # Initializing weights # 1st layer Weights w1 = 0.11 w2 = 0.27 w3 = 0.19 w4 = 0.52 # 2nd layer weights w5 = 0.44 w6 = 0.48 w7 = 0.23 w8 = 0.29 # Bias b1 = 0.39 b2 = 0.42 # Learning rate eta = 0.3 # setting 100 iteration to tune our neural network algorithm iteration = 100 # 1st layer weights matrix weights_h1 = np.array([[w1], [w2]]) weights_h2 = np.array([[w3], [w4]]) # 2nd layer weights matrix weights_y1 = np.array([[w5], [w6]]) weights_y2 = np.array([[w7], [w8]]) ##################### Forward Propagation ########################## # Entire hidden layer weight matrix weights_h = np.row_stack((weights_h1.T, weights_h2.T)) # Entire output layer weight matrix weights_y = np.row_stack((weights_y1.T, weights_y2.T)) # Sigmoid Activation function ==> S(x) = 1/1+e^(-x) def sigmoid(x, deriv=False): if deriv == True: return x * (1 - x) return 1 / (1 + np.exp(-x)) h = np.dot(x_val, weights_h.T) + b1 # Entire 1st layer output matrix out_h = sigmoid(h) y = np.dot(out_h, weights_y.T) + b2 # Entire 2nd layer output matrix out_y = sigmoid(y) ##################### Error Calculation ########################## # E as E total E_total = (np.square(y_val - out_y))/2 ##################### Back Propagation ########################## # 1. Update 2nd layer weights with change value 111111111111111111 # (dE_Total)/(dOut y_1 ) dE_total_dout_y = -(y_val - out_y) # (d Out y_1)/(dy_1 ) dout_y_dy = out_y * (1 - out_y) # (dy_1)/(dw_5 ) dy_dw = out_h # For each iteration for iter in range(iteration): # Foreach row of input data update 2nd layer weight matrix for row in range(len(x_val)): # row = 0 # (dE_Total)/(dw_5 ) = (dE_Total)/(dOut y_1 )*(dOut y_1)/(dy_1 )*(dy_1)/(dw_5 ) dE_Total_dw5 = dE_total_dout_y[row][0] * round(dout_y_dy[row][0], 8) * dy_dw[0][0] dE_Total_dw5 = round(dE_Total_dw5, 8) # (dE_Total)/(dw_5 ) = (dE_Total)/(dOut y_1 )*(dOut y_1)/(dy_1 )*(dy_1)/(dw_5 ) dE_Total_dw6 = dE_total_dout_y[row][0] * round(dout_y_dy[row][0], 8) * dy_dw[0][1] dE_Total_dw6 = round(dE_Total_dw6, 8) # (dE_Total)/(dw_5 ) = (dE_Total)/(dOut y_1 )*(dOut y_1)/(dy_1 )*(dy_1)/(dw_5 ) dE_Total_dw7 = dE_total_dout_y[row][0] * round(dout_y_dy[row][1], 8) * dy_dw[0][0] dE_Total_dw7 = round(dE_Total_dw7, 8) # (dE_Total)/(dw_5 ) = (dE_Total)/(dOut y_1 )*(dOut y_1)/(dy_1 )*(dy_1)/(dw_5 ) dE_Total_dw8 = dE_total_dout_y[row][0] * round(dout_y_dy[row][1], 8) * dy_dw[0][1] dE_Total_dw8 = round(dE_Total_dw8, 8) # Combine all differential weights dE_Total_dw_2nd_layer = np.array([[dE_Total_dw5, dE_Total_dw6], [dE_Total_dw7, dE_Total_dw8]]) # Updated weights for 2nd layer # (new)w_5 = w_5-η*(dE_Total)/(dw_5 ) [η is learning rate] weights_y = weights_y - (eta * dE_Total_dw_2nd_layer) weights_y # 2. Update 1st layer weights with change value 22222222222222222 # (dE_2)/(dy_2 )=(dE_2)/(d〖Out y〗_2 )*(d〖Out y〗_2)/(dy_2 ) dE_dy = -(y_val - out_y) * (out_y * (1-out_y)) # (dE_1)/(dOut h_1 )= (dE_1)/(dy_1 )*(dy_1)/(dOut h_1 ) dE_dOut_h1 = dE_dy * np.array([[w5, w7]]) # (dE_2)/(dOut h_1 )= (dE_2)/(dy_2 )*(dy_2)/(dOut h_1 ) dE_dOut_h2 = dE_dy * np.array([[w6, w8]]) # (dE_Total)/(dOut h_1 )=(dE_1)/(dOut h_1 )+(dE_2)/(dOut h_1 ) dE_Total_dOut_h1 = dE_dOut_h1[row][0] + dE_dOut_h1[row][1] # (dOut h_1)/(dh_1 )=Outh_1 (1-Outh_1) dOut_h_dh = out_h * (1-out_h) # dh1_dw1 = x dh_dw = x_val # (dE_Total)/(dw_1 )=(dE_Total)/(dOut h_1 )*(dOut h_1)/(dh_1 )*(dh_1)/(dw_1 ) dE_Total_dw1 = dE_Total_dOut_h1 * dOut_h_dh[row][0] * dh_dw[row][0] dE_Total_dw1 = round(dE_Total_dw1, 8) # (dE_Total)/(dw_1 )=(dE_Total)/(dOut h_1 )*(dOut h_1)/(dh_1 )*(dh_1)/(dw_1 ) dE_Total_dw2 = dE_Total_dOut_h1 * dOut_h_dh[row][0] * dh_dw[row][1] dE_Total_dw2 = round(dE_Total_dw2, 8) # (dE_Total)/(dw_1 )=(dE_Total)/(dOut h_1 )*(dOut h_1)/(dh_1 )*(dh_1)/(dw_1 ) dE_Total_dw3 = dE_Total_dOut_h1 * dOut_h_dh[row][1] * dh_dw[row][0] dE_Total_dw3 = round(dE_Total_dw3, 8) # (dE_Total)/(dw_1 )=(dE_Total)/(dOut h_1 )*(dOut h_1)/(dh_1 )*(dh_1)/(dw_1 ) dE_Total_dw4 = dE_Total_dOut_h1 * dOut_h_dh[row][1] * dh_dw[row][1] dE_Total_dw4 = round(dE_Total_dw4, 8) # Combine all differential weights dE_Total_dw_1st_layer = np.array([[dE_Total_dw1, dE_Total_dw2], [dE_Total_dw3, dE_Total_dw4]]) # update weights w1 weights_h = weights_h - (eta * dE_Total_dw_1st_layer) print('iteration: ' + str(iter) + ' complete')
To ease to remember everything, let me list down all equations of Neural Network
Also Read:
Naive Bayes algorithm in Machine Learning with Python
Neural Network Equations
Forward Propagation
h1= x1w1 + x2w2 +b1
h2 = x1w3 + x2w4+b1
y1 = Outh1*w5 + Outh2*w6 + b2
y2 = Outh1*w7 + Outh2*w8 + b2
Error Calculation
Back Propagation
1.Update 2nd layer weights
2.Update 1st layer weights
Conclusion
In this tutorial I have shown just one epoch (one round of forward pass to get the predicted output value and one round of backward pass to update all weights (w1,w2,…w5,..w8). After updating again you need to calculate total error by doing forward pass.
After one round of iteration you may get better accuracy than initial stage of neural network. But after repeating this process (iteration) 1,000 or 10,000 times you can achieve predicted output near to actual target value.
Derivative of sigmoid function
Explanation
If you have any question or suggestion regarding this topic see you in comment section. I will try my best to answer.
Share this