Kevin Musgrave

Maybe a straightforward conversion: https://ax.dev/tutorials/raytune_pytorch_cnn.html. Will look into more later.

I'm not sure why that is happening. However, I think your workaround might be equivalent, according to this issue: https://github.com/facebook/Ax/issues/133

Where does the error come from? Is it coming from your model or a loss function?

So the bug originates here? https://github.com/KevinMusgrave/powerful-benchmarker/blob/master/src/powerful_benchmarker/architectures/misc_models.py#L76

Yeah I think it only affects memory usage.

> 1. Using the train/test split already created, and have class disjoint train/val sets Could you give an example? I'm not 100% sure what you mean. Currently, this repo will...

Re: SOP, I'm not sure if there will be an easy way to do that with this library. To be honest, the code is a bit of a mess right...

I'll take a look at the code later this week to see if there's a straightforward to implement your idea.

For now, here's how you can fix this. 1. Create this custom reducer in run.py. (The code below assumes you're using pytorch-metric-learning v0.9.88) ```python from pytorch_metric_learning.reducers import ThresholdReducer class CustomAvgNonZeroReducer(ThresholdReducer):...

This is fixed in v1.7.3. The only caveat is that `len(enqueue_idx)` must be the same in all processes. In other words, the same number of embeddings have to be enqueued...