Oscar Knagg comments

Results 5 comments of


                                            Oscar Knagg

Could you please elaborate the following line? why not use the y coming from batch?

Yes thats right. The y coming from the batch is the underlying class from the whole dataset (>1000 classes for Omniglot, 80 for miniImageNet). We don't want want to use...

Best practice for evaluation code

I only used the callback when evaluating so I don't have any other evaluation code. This function is what the evaluate callback uses. https://github.com/oscarknagg/few-shot/blob/672de83a853cc2d5e9fe304dc100b4a735c10c15/few_shot/eval.py#L9 You can call it on its...

Implementation of 2nd order gradient in maml

I use the hook method for 1st order MAML i.e. I use the sum of the gradients on the fast weights to update the slow weights. https://github.com/oscarknagg/few-shot/blob/672de83a853cc2d5e9fe304dc100b4a735c10c15/few_shot/maml.py#L100 However I do...

Handle new iTunes login

I can confirm that this PR solves my issue with the new auth type. Thanks @doguh

How big can you make a fully fused MLP while retaining performance benefits?

For reference, here are the specs of an A100 vs a 3090: - A100: 256 KB register per SM, 164 KB shared memory per SM, 108 SMs ([source](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/nvidia-ampere-architecture-whitepaper.pdf)) - 3090:...