Tiny EfficentNet

[snarb]

Is it possible to create tiny version of Efficent Net, using the configs or maybe the code without huge modification? Am am interesting in something like 200k parameters.

[rwightman]

@snarb

Looking at the bottom of
https://github.com/huggingface/pytorch-image-models/blob/main/results/results-imagenet.csv

And especially top of https://github.com/huggingface/pytorch-image-models/blob/main/results/benchmark-infer-amp-nchw-pt113-cu117-rtx3090.csv

...will give you some of the current smallest models in timm, a lot of those are closely related to EfficientNet in architecture but have design considerations for mobile use (mobilenetv2, mobilenetv3, mnasnet, tinynet, lcnet, ghostnet)

You could modify any of those to be even smaller, many have a depth and/or width scale factor that you can use to downscale the original layout, but 200k parameters isn't suitable for ImageNet (models here tend to be focused on at least ImageNet or larger scale pretrainining).