How to add noise to image in denoising autoencoders

1

I want to implement denoising autoencoder in pytorch.
when I am in the training loop:

for inputs,labels in trainloader:
        noised_inputs=add_noise(inputs)
        opt.zero_grad()
        outputs = net(noised_inputs)
        loss = loss_fn(outputs,inputs)
        loss.backward()
        opt.step()

Please tell me how to implement this add_noise function which can add gaussian noise to the entire batch input of images

2

You can use the torch.randn_like() function to create a noisy tensor of the same size of input. Then add it.
In your case ,

def add_noise(inputs):
     noise = torch.randn_like(inputs)
     return inputs + noise
3

It worked!!! also we can multiply it with factor like 0.2 to reduce the noise

def add_noise(inputs):
     noise = torch.randn_like(inputs)*0.2
     return inputs + noise

Original Image:
image

Noised Image:
image