vision-cair / cs326-few-shot-classification Goto Github PK

I noticed that in the outer loop, the outer_loop_loss.backward() and optim.step() cannot update the MAMLModel.params, which causes each episode cannot update the model.params.

Finally, the meta part cannot change the model.params and has no effect.

I’m wondering whether this is deliberately designed? If yes, it seems that we need to add some code except the “TODO” to update the initial params using the gradients of fast weights?

I understand that we for save of computing power measure accuracy and loss via moving model, but it may be the case that current implementation that just returns last loss and last acc in the last batch a bit incorrect.

I suggest replace it:
``` sum(losses)/len(losses), sum(accs)/len(accs)``

It's still approximate computation of loss and accuracy because we evaluate acc and loss in different points.

line 34 in maml.py:
def __call__(self, x: Tensor) -> Tensor:
Should this function be changed to
def __call__(self, x: Tensor, params: Tensor = None) -> Tensor: