linreg_gradient_descent

linreg_gradient_descent.py

b0, b1 = 0.0, 1.0

lr = 0.001
epochs = 10000
error = []

# run 10000 times
for epoch in range(epochs):
    # initialize to 0 -> cost of epoch, Jb_0, Jb_1
    epoch_cost, cost_b0, cost_b1 = 0, 0, 0
    
    for i in range(len(x)):
        # make prediction
        y_pred = (b0 + b1*x[i])
        # append squared error
        epoch_cost += (y[i] - y_pred)**2 
        
        for j in range(len(x)):
            # partial derivative of b0 and b1 for current row
            partial_wrt_b0 = -2 * (y[j] - (b0 + b1*x[j]))
            partial_wrt_b1 = (-2 * x[j]) * (y[j] - (b0 + b1*x[j]))
            
            # increase cost of coeffs
            cost_b0 += partial_wrt_b0
            cost_b1 += partial_wrt_b1

        # calculate new coeffs
        b0 = b0 - lr * cost_b0
        b1 = b1 - lr * cost_b1
        
    # keep track of errors - for visualization purposes
    error.append(epoch_cost)