Kyle Daruwalla
Kyle Daruwalla
There was a plan to open up the `pretrain` argument to allow specifying the name of an artifact to load. The intention being to have multiple artifacts per model spec....
> This regression also seems to be happening on VGG for some reason. With or without loading the weights?
Reorganizing sounds like a good plan. We can have deprecations for anything in `Layers` that moves out and is unexported at the top level. If it is exported then I...
That makes it even simpler then
This is exactly the opposite approach of what we set out to do with the Metalhead re-write in 0.6. Copying and pasting a bottleneck block over and over is good...
We made an effort to separate the building blocks from the models. But maybe we can reorganize to have the building block files (mbconv.jl, etc.) live right next to the...
It's not all on you. I will make time this week to do these cleanup steps so that additional contributors (like Carlo) can help, and we can release 0.8 soon...
One thing we want to do but have not done yet is to use the named syntax for `Chain` and `Parallel`. So we should have ```julia Chain(resblock0 = Parallel(..., residual...
This also looks good vs. my code. Only main difference I see is the use of data augmentation. It would be nice to see if that makes a meaningful difference.
> I think the Inception errors are because that family of models use an image size of 299x299, and I don't think they support alternate image sizes. AlexNet uses 224x224...