search

FluxML / fluxml.github.io

Flux Website
https://fluxml.ai
MIT License
20 stars 45 forks source link

MNIST experiment accuracy #28

Closed mcognetta closed 3 years ago

mcognetta commented 5 years ago

The accuracy of the conv classifier for MNIST on the website is rather poor. This is because the model was trained for only one epoch (if the code from model-zoo) was actually used.

See: https://github.com/FluxML/model-zoo/pull/72.

I have weights for a model in bson format that I can make a PR for if necessary.

MikeInnes commented 5 years ago

Worth talking to @Roboneet. I'm not sure what the specs of the model are at this point; for a while it seemed to be working much better.

Roboneet commented 5 years ago

The current model has an accuracy of approx 96% on the dataset (iirc). Also, the first layer of the model used a size of (3, 3) (instead of (2, 2) as in model-zoo), because of an issue with the javascript layers (This may not come up in newer versions of tf-js). It'd be nice to have a better model :)

mcognetta commented 5 years ago

@Roboneet Thanks for the info. I figured it used the weights that would result from the linked example on model zoo. Have you tried it out on the website recently? I don't think it is as accurate as advertised.

Roboneet commented 5 years ago

@mcognetta, you can try the model using this (use Flux.jl#5d8b63d commit) on julia 0.6

ToucheSir commented 3 years ago

Since the experiments have long since become unmaintained and moved into the attic, I'm going to close this up. We can always re-open if/when the experiments are rejuvenated.