VisionActivationsStore error when provided no eval dataset

[lowlorenz]

# vit_prisma.sae.training.activations_store.py
class VisionActivationsStore:
    """
    Class for streaming tokens and generating and storing activations
    while training SAEs.
    """

    def __init__(
        self,
        cfg: Any,
        model: HookedViT,
        dataset,
        create_dataloader: bool = True,
        eval_dataset = None,
        num_workers=0,
    ):
        self.cfg = cfg
        self.model = model
        self.model.to(cfg.device)
        self.dataset = dataset
        self.image_dataloader = torch.utils.data.DataLoader(self.dataset, shuffle=True, num_workers=num_workers, batch_size=self.cfg.store_batch_size, collate_fn=collate_fn, drop_last=True)
        self.image_dataloader_eval = torch.utils.data.DataLoader(eval_dataset, shuffle=True, num_workers=num_workers, batch_size=self.cfg.store_batch_size, collate_fn=collate_fn_eval, drop_last=True)

        self.image_dataloader_iter = self.get_batch_tokens_internal()
        self.image_dataloader_eval_iter = self.get_val_batch_tokens_internal()
....

This issue serves as documentation and will be addressed later today.
Ensuring that eval_dataset is properly checked will prevent runtime errors and improve the robustness of the VisionActivationsStore class.