Custom Dataset class on Chapter_1

[LucianoBatista]

Hi, I was getting the followed erro when I executing this code:

from torch.utils.data import Dataset
from sklearn.datasets import fetch_openml

X, y = fetch_openml("mnist_784", version=1, return_X_y=True)

class SimpleDataset(Dataset):
    def __init__(self, X, y):
        super(SimpleDataset, self).__init__()
        self.X = X
        self.y = y
    
    def __getitem__(self, index):
        inputs = torch.tensor(self.X[index, :], dtype=torch.float32)
        targets = torch.tensor(int(self.y[index]), dtype=torch.int64)
        return inputs, targets

    def __len__(self):
        return self.X.shape[0]

dataset = SimpleDataset(X, y)
example, label = dataset[0]

InvalidIndexError: (tensor(0), slice(None, None, None))

The same was fixed when I change the code of the fetch_openml to:

X, y = fetch_openml("mnist_784", version=1, return_X_y=True, as_frame=False)

The problem was that whithout the as_frame, scikit will import the data as a DataFrame, not as numpy anymore.

[EdwardRaff]

Do you know if this was a change in scikit learn recently?

[LucianoBatista]

I don't know, currently, I'm using the newer (1.0.2) during the study of your the book:
https://pypi.org/project/scikit-learn/

[EdwardRaff]

Changed in version 0.24: The default value of as_frame changed from False to 'auto' in 0.24.

Yup, they changed it. I'll fix this soon - thank you for the catch!

[ChowZH]

Changing the getitem function by adding .loc after the dataframe seems to solve the issue. Still working my way past this point but atleast I now got the same output as the book.

def getitem(self, index):
inputs = torch.tensor(self.X.loc[index,:], dtype=torch.float32)
targets = torch.tensor(int(self.y.loc[index]), dtype=torch.int64)
return inputs, targets

[chuymtz]

I encountered this today and made a comment in the livebook from manning. I think this solution is better than what i suggested there.

If i change self.X = X.to_numpy() does that casuse problems down the road about being able to benefit from pytorch? Great book im really enjoying it.

[murphyk]

I used dataset = SimpleDataset(X.to_numpy(), y.to_numpy()) and it solved the problem.
This could go into the __init__ function as @chuymtz suggested, but
if we want to make a dataset from a torch tensor, we would have to use
dataset = SimpleDataset(X.numpy(), y.numpy()) instead (not to_numpy - argh).

XX = torch.tensor(X.to_numpy())
yy = torch.tensor(y.to_numpy(dtype=int), dtype=torch.int64)
dataset2 = SimpleDataset(XX.numpy(), yy.numpy())

[yebangyu]

a little tedious for me about murphyk's solution since it converts X to tensor and then back to numpy again.

'X, y = fetch_openml("mnist_784", version=1, return_X_y=True, as_frame=False)'

this solution is very good. thanks, LucianoBatista

[yebangyu]

another solution seems good, just fyi

X, y = fetch_openml("mnist_784", version=1, return_X_y=True)  # no need to change here
class SimpleDataset(Dataset):
    def __init__(self, X, y):
        super(SimpleDataset, self).__init__()
        self.X = X.values  # get the numpy data via values
        self.y = y.values # get the numpy data via values

dataset = SimpleDataset(X, y)
example, label = dataset[0]

[benjwolff]

@LucianoBatista: Thanks for your solution! I ran into the same problem. I created a pull request.