fix(*:skip) Fix querying dataset size functions

[KarhouTam]

Issue

Description

I found that in many places, when flower collaborates with pytorch, it uses len(self.trainloader) and len(self.testloader) to query the sizes of clients train set and test set.
However, in pytorch, len(dataloader) will return the num of batches that the dataloader.dataset can split into, not the real size of dataloader.dataset. Means that len(dataloader) != len(dataloader.dataset), and the latter one is more reasonable in federated learning scenario, I think.

For example:

>>> from torch.utils.data import DataLoader, Subset
>>> from torchvision.datasets import MNIST
>>> mnist = MNIST(...)
>>> subset_idxs = list(range(64))
>>> mnist_subset = Subset(mnist, indices=subset_idxs)
>>> dataloader_mnist_whole = DataLoader(mnist, batch_size=32)
>>> dataloader_mnist_subset = DataLoader(mnist_subset, batch_size=32)
>>> len(dataloader_mnist_whole)
1875
>>> len(dataloader_mnist_subset)
2
>>> len(dataloader_mnist_whole.dataset)
60000
>>> len(dataloader_mnist_subset.dataset)
64

Related issues/PRs

Proposal

So I just search len(self.trainloader) and len(self.testloader) globally and replace them with len(self.trainloader.dataset) and len(self.testloader.dataset).

Checklist

• Implement proposed change
• Write tests
• Update documentation
• Make CI checks pass
• Ping maintainers on Slack (channel #contributions)

[jafermarq]

Hi @KarhouTam, thanks for opening the PR. This is just an implementation choice right (and not a bug) The weighing factor used for aggregation is just a function of the batchsize rather than the total number of examples. In most cases this will result in near identical (because some batches might be incomplete) results as if the total number of samples was instead communicated (assuming all clients use the same batch size)

We can follow this approach in newer examples. For the baselines i don't think we'll make the change.

[KarhouTam]

I think this change has no risk at all and is better to uniform the aggregation weight factor in flower.

But anyway, I respect your decision. 🫡

(This PR can be reopened anytime if you change your mind.)