Enumerate/iter for multiple dataloaders

1

I am trying to divide MNIST into separate subsets where each subset is a different class.

train_dataset = datasets.MNIST('/some/dir/', train=True, download=True, transform=self.dataset_transform)

for i in range(10):
    idx = train_dataset.targets==i
    train_dataset.targets[idx]                
    train_subsets.append(torch.utils.data.Subset(train_dataset, idx))
    train_loaders.append(torch.utils.data.DataLoader(train_subsets[i], batch_size=self.batch_size, shuffle=True))
    train_iters.append(iter(train_loaders[i]))

Beginning of my training loop:

for label_id in range(10):
    data, target = train_iters[label_id].next()

When training, I get a ValueError:

only one element tensors can be converted to Python scalars

What is going on here?

2

Hello,

In my shallow opinion, there might be something wrong with your snippet.

Here, idx is tensor with values 0 and 1. So if you pass it as indices argument to Subset, the subset will only contain the first and the second target (means position 0 and position 1).
And train_dataset.targets[idx] does not make sense to me.

For this error, you could try to print train_iters[lable_id] 's shape and debug it.

1 Like
3

Finally, I was able to make this work with:

for i in range(10):
    idx = train_dataset.targets==i
    idx_np = idx.numpy()
    idx_num = idx_np * range(len(idx))
    idx_no_zeros = list(filter(lambda a: a != 0, idx_num))
    train_subsets.append(torch.utils.data.Subset(train_dataset, idx_no_zeros))
    train_loaders.append(torch.utils.data.DataLoader(train_subsets[i], batch_size=self.batch_size, shuffle=True))