Peter O'Connor
Peter O'Connor
`-init_image` doesn't seem to work, or at least not how I expect it to. The second image here is 100 iterations into optimization after being called with the first image...
Thanks guys for releasing this great code. I noticed you actually need to run it with the ResNet101 checkpoint. I added instructions here.
NPTyping is great, but could be more concise. The main value of NDArray to me is to have a very concise way of documenting shapes. I would propose - Just...
We've already made a rough built-in hyperparameter optimizer in master which uses scipy's gaussian-process search. We should clean this up a bit and at least add random/grid search options.
It would be good to be able to see the entire list of currently-running experiments. This involves something we should probably have anyway, a "heartbeat" for experiments that indicates that...
Accessing all experiments in the file system can be slow. Maybe we should follow Sacred's lead and use mongodb for storing results.
Performance/Memory wise, it makes sense to just append the results of generator-experiments to file. Right now it just overwrites the results, so you need to build a growing data structure...
When you remove or change the name of an experiment variant, the corresponding records can be "disappeared". It would be good to at least have the option to list these...