Apply Custom Augmentation to MNIST and create Data Loader

vision
1

I have three types of custom augmentations to be performed on the MNIST(written three different functions for the same). How do I do create a data loader comprising of augmented data? The method I’m currently using throws an error for augmentation functions that accept a parameter.

# the augmentation function

def pad_rotate_project(entry, theta, show = False, requires_theta = False):
    if (type(entry) != torch.Tensor):
        entry = torch.tensor(entry.T)
    copy = entry.clone()
    if len(entry.shape) == 2:
        copy = copy.unsqueeze(dim = 0)
    copy = F.pad(copy, pad=(1,1,1,1), value=0)
    rot = transforms.RandomRotation(degrees=(theta, theta))
    rotated = rot(transforms.ToPILImage()(copy))
    plt.imshow(rotated)
    tensor = transforms.ToTensor()(rotated)
    sum_axis = torch.sum(tensor, dim = 2)
    # If this is true, then we append the rotation angle applied to the output 
    if requires_theta:
        to_add =  np.full(shape=copy.size()[0], fill_value=theta)
        rot_tens = torch.tensor(to_add, dtype=torch.float).unsqueeze(dim = 1)
        sum_axis = torch.cat([sum_axis, rot_tens], dim = 1)
    return sum_axis

# class to create custom dataset. I plan to create 3 custom datasets and then concatenate them

class pad_rotate_project_data(data.Dataset):  
 
    def __init__(self, X_data,  theta = 100, transform2=None):

        self.X_data = X_data
        self.theta = theta
        self.transform2 = transform2


    def __getitem__(self, index):
      
        img = self.X_data[index][0]
        label = self.X_data[index][1]
        
        #perform augmentation
        if self.transform2:
            img = self.transform2(img, **{'theta' : self.theta})
            
        
        return img,label

    def __len__(self):
        return len(self.X_data)


# part where i apply transformation
ds2 = convolve_noise_data(mnist_train, transform2 = pad_rotate_project)

train_loader2 = data.DataLoader(ds2, batch_size,
                                    #sampler = RandomSampler(train_ds),
                                    num_workers = THREADS,
                                    pin_memory= USE_CUDA )

Below is the error I face: ---------------------------------------------------------------------------

TypeError                                 Traceback (most recent call last)
<ipython-input-58-2e42ff52d82a> in <module>()
----> 1 ds2 = convolve_noise_data(mnist_train, transform2 = pad_rotate_project)
      2 
      3 train_loader2 = data.DataLoader(ds2, batch_size,
      4                                     #sampler = RandomSampler(train_ds),
      5                                     num_workers = THREADS,

TypeError: __init__() got an unexpected keyword argument 'transform2'
2

Your code runs fine if I change convolve_noise_data to pad_rotate_project_data, since the firmer one is undefined.

PS: I’ve formatted your code so that I could easily copy and debug it. You can add code snippets using three backticks ``` :wink:

1 Like
4

Thanks for your response. Can you please explain why this works: transform2 = pad_rotate_project ?
Especially since pad_rotate_project is a function with arguments, how come just mentioning the name of the function is enough to ensure the function is applied to each image in the dataset?

5

Inside your Dataset you are assigning this functions to self.transform2.
This will basically assign the reference to this function as a class attribute.
In __getitem__ you are then calling self.transform2 with the necessary arguments.