Why does regularization in pytorch and scratch code does not match and what is the formula used for regularization in pytorch?

Question

I have been trying to do L2 regularization on a binary classification model in PyTorch but when I match the results of PyTorch and scratch code it doesn't match, Pytorch code:

class LogisticRegression(nn.Module):
  def __init__(self,n_input_features):
    super(LogisticRegression,self).__init__()
    self.linear=nn.Linear(4,1)
    self.linear.weight.data.fill_(0.0)
    self.linear.bias.data.fill_(0.0)

  def forward(self,x):
    y_predicted=torch.sigmoid(self.linear(x))
    return y_predicted

model=LogisticRegression(4)

criterion=nn.BCELoss()
optimizer=torch.optim.SGD(model.parameters(),lr=0.05,weight_decay=0.1)
dataset=Data()
train_data=DataLoader(dataset=dataset,batch_size=1096,shuffle=False)

num_epochs=1000
for epoch in range(num_epochs):
  for x,y in train_data:
    y_pred=model(x)
    loss=criterion(y_pred,y)
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

Scratch Code:

def sigmoid(z):
    s = 1/(1+ np.exp(-z))
    return s  

def yinfer(X, beta):
  return sigmoid(beta[0] + np.dot(X,beta[1:]))

def cost(X, Y, beta, lam):
    sum = 0
    sum1 = 0
    n = len(beta)
    m = len(Y)
    for i in range(m): 
        sum = sum + Y[i]*(np.log( yinfer(X[i],beta)))+ (1 -Y[i])*np.log(1-yinfer(X[i],beta))
    for i in range(0, n): 
        sum1 = sum1 + beta[i]**2
        
    return  (-sum + (lam/2) * sum1)/(1.0*m)

def pred(X,beta):
  if ( yinfer(X, beta) > 0.5):
    ypred = 1
  else :
    ypred = 0
  return ypred

beta = np.zeros(5)
iterations = 1000
arr_cost = np.zeros((iterations,4))
print(beta)
n = len(Y_train)
for i in range(iterations):
    Y_prediction_train=np.zeros(len(Y_train))
    Y_prediction_test=np.zeros(len(Y_test)) 

    for l in range(len(Y_train)):
        Y_prediction_train[l]=pred(X[l,:],beta)
    
    for l in range(len(Y_test)):
        Y_prediction_test[l]=pred(X_test[l,:],beta)
    
    train_acc = format(100 - np.mean(np.abs(Y_prediction_train - Y_train)) * 100)
    test_acc = 100 - np.mean(np.abs(Y_prediction_test - Y_test)) * 100   
    arr_cost[i,:] = [i,cost(X,Y_train,beta,lam),train_acc,test_acc]
    temp_beta = np.zeros(len(beta))

    ''' main code from below '''

    for j in range(n): 
        temp_beta[0] = temp_beta[0] + yinfer(X[j,:], beta) - Y_train[j]
        temp_beta[1:] = temp_beta[1:] + (yinfer(X[j,:], beta) - Y_train[j])*X[j,:]
    
    for k in range(0, len(beta)):
        temp_beta[k] = temp_beta[k] +  lam * beta[k]  #regularization here
    
    temp_beta= temp_beta / (1.0*n)
    
    beta = beta - alpha*temp_beta

graph of the losses

graph of training accuracy

graph of testing accuracy

Can someone please tell me why this is happening? L2 value=0.1

Answer 1

Great question. I dug a lot through PyTorch documentation and found the answer. The answer is very tricky. Basically there are two ways to calculate regulalarization. (For summery jump to the last section).

The PyTorch uses the first type (in which regularization factor is not divided by batch size).

Here's a sample code which demonstrates that:

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import torch.optim as optim
 
class model(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(1, 1)
        self.linear.weight.data.fill_(1.0)
        self.linear.bias.data.fill_(1.0)

    def forward(self, x):
        return self.linear(x)


model     = model()
optimizer = optim.SGD(model.parameters(), lr=0.1, weight_decay=1.0)

input     = torch.tensor([[2], [4]], dtype=torch.float32)
target    = torch.tensor([[7], [11]], dtype=torch.float32)

optimizer.zero_grad()
pred      = model(input)
loss      = F.mse_loss(pred, target)

print(f'input: {input[0].data, input[1].data}')
print(f'prediction: {pred[0].data, pred[1].data}')
print(f'target: {target[0].data, target[1].data}')

print(f'\nMSEloss: {loss.item()}\n')

loss.backward()

print('Before updation:')
print('--------------------------------------------------------------------------')
print(f'weight [data, gradient]: {model.linear.weight.data, model.linear.weight.grad}')
print(f'bias [data, gradient]: {model.linear.bias.data, model.linear.bias.grad}')
print('--------------------------------------------------------------------------')
 
optimizer.step()

print('After updation:')
print('--------------------------------------------------------------------------')
print(f'weight [data]: {model.linear.weight.data}')
print(f'bias [data]: {model.linear.bias.data}')
print('--------------------------------------------------------------------------')

which outputs:

input: (tensor([2.]), tensor([4.]))
prediction: (tensor([3.]), tensor([5.]))
target: (tensor([7.]), tensor([11.]))

MSEloss: 26.0

Before updation:
--------------------------------------------------------------------------
weight [data, gradient]: (tensor([[1.]]), tensor([[-32.]]))
bias [data, gradient]: (tensor([1.]), tensor([-10.]))
--------------------------------------------------------------------------
After updation:
--------------------------------------------------------------------------
weight [data]: tensor([[4.1000]])
bias [data]: tensor([1.9000])
--------------------------------------------------------------------------

Here m = batch size = 2, lr = alpha = 0.1, lambda = weight_decay = 1.

Now consider tensor weight which has value = 1 and grad = -32

case1(type1 regularization):

 weight = weight - lr(grad + weight_decay.weight)
 weight = 1 - 0.1(-32 + 1(1))
 weight = 4.1

case2(type2 regularization):

 weight = weight - lr(grad + (weight_decay/batch size).weight)
 weight = 1 - 0.1(-32 + (1/2)(1))
 weight = 4.15

From the output we can see that updated weight = 4.1000. That concludes PyTorch uses type1 regularization.

So finally In your code you are following type2 regularization. So just change some last lines to this:

# for k in range(0, len(beta)):
#    temp_beta[k] = temp_beta[k] +  lam * beta[k]  #regularization here

temp_beta= temp_beta / (1.0*n)

beta = beta - alpha*(temp_beta + lam * beta)

And also PyTorch loss functions doesn't include regularization term(implemented inside optimizers) so also remove regularization terms inside your custom cost function.

In summary:

1. Pytorch use this Regularization function:

   

2. Regularization is implemented inside Optimizers (weight_decay parameter).

3. PyTorch Loss functions doesn't include Regularization term.

4. Bias is also regularized if Regularization is used.

5. To use Regularization try:

   torch.nn.optim.optimiser_name(model.parameters(), lr, weight_decay=lambda).