uvm-neurobotics-lab / anml Goto Github PK

Hi, there was a bug on the original code of OML where they did not create the computation graph for the inner loop updates. Looking the ANML code it seems to have the same issue, specifically here it is currently

grad = torch.autograd.grad(loss, fast_weights, allow_unused=False)

but to correctly backpropagate through the inner optimization it should be

grad = torch.autograd.grad(loss, fast_weights, allow_unused=False, create_graph=True)

I was wondering which version was used for the results on the paper, OML's author said fixing this bug improved performance and reduced training time.
Thanks

p.d.: congratz on the work, is really cool

I was wondering why are the batch norm stats computed as a running mean/stddev over every sample rather than over the whole batch at once.

Can ANML training regime be used for image segmentation task? Or is there any other method for meta learning as such ANML for image segmentation task?

Hello :)
First of all, thanks for your hard work
It have give me new insight of considering CLP in meta learning

By the way, can you explain the hyper-parameters used for training OML in your code.
I am trying to reproduce OML in your repository, but facing hard time finding the right parameter that matches the results of your paper. It would be really helpful if you can give the exact script like the one you've written for training ANML model.

Thanks again 😀

My understanding is that you use a single fully connected layer on top of the neuro-modulated representations.

1. Does this output layer has 963 nodes during meta-training, since you are performing 963-class classification during meta-training ?
2. If yes, how do you use the meta-learned output layer to learn/perform 600-class classification at meta-testing time ? Or do you randomly initialize a new output layer with 600 nodes?