Enable empty split

[rayrayraykk]

No description provided.

[joneswong]

In my opinion, when there is no train split, users should be responsible for failure of accessing train split. So this behavior is ok:

FederatedScope/federatedscope/contrib/trainer/torch_example.py

Line 41 in e2ce0ce

for x, y in self.data['train']:

As for val and test splits, what is our intended behavior?

[joneswong]

In my opinion, when there is no train split, users should be responsible for failure of accessing train split. So this behavior is ok:

FederatedScope/federatedscope/contrib/trainer/torch_example.py

Line 41 in e2ce0ce

for x, y in self.data['train']:

As for val and test splits, what is our intended behavior?

e.g., is this still ok?

FederatedScope/federatedscope/contrib/trainer/torch_example.py

Line 68 in e2ce0ce

for x, y in self.data[target_data_split_name]:

[rayrayraykk]

In my opinion, when there is no train split, users should be responsible for failure of accessing train split. So this behavior is ok:

FederatedScope/federatedscope/contrib/trainer/torch_example.py

Line 41 in e2ce0ce

for x, y in self.data['train']:

As for val and test splits, what is our intended behavior?

e.g., is this still ok?

FederatedScope/federatedscope/contrib/trainer/torch_example.py

Line 68 in e2ce0ce

for x, y in self.data[target_data_split_name]:

Now, val and test splits follow the same principle in GeneralTorchTrainer. But users are free to set the rules for no val and test splits in user-defined trainer, e.g.

if target_data_split_name not in self.data:
    do nothing
    return {f'{target_data_split_name}_loss': 0,
                f'{target_data_split_name}_total': 0,}

[joneswong]

LGTM

[joneswong]

approved. It seems that we'd better make our proj more "typed".