pytorch-image-models
pytorch-image-models copied to clipboard
huggingface
[FEATURE] Support shufflenet
tkx for your wonderful work, I found the timm is not support shufflenet v1 and v2 yet, there is any plan to support it? and here is the official repository and related papers.
@flywheel1412 it's never been a priority since they didn't seem particularly stronger or better than other mobile models here already (mobilenet-v2/v3, hardcorenas, tinynet, lcnet, ghostnet, etc). Could likely be supported via efficientnet/mobilenetv3 backbones with additional block type(s) and the extra SE in head... GhostNet is also a very similar model layout that'd be a template to follow... open to well tested full impl with features_only, etc working. But hasn't been a priority for me to tackle.
tkx for your reply, where cloud i found the benchmark table of timm's models? which cloud be a better guide for everyone to choose models
@flywheel1412 the last batch of results are all in the csv files here: https://github.com/huggingface/pytorch-image-models/tree/main/results
benchmark-* are the per architecture inference and train throughput numbers, and the results-* are the per-weight instance accuracy numbers for imagenet and other test sets. You need to join them on the architecture (portion before the .) to get a per result score to plot a pareto curve like I've done in the past https://twitter.com/wightmanr/status/1463684184912711689
Latest GhostNet-V2, RepGhostNet, MobileOne models that I just added are not in the results yet... probably another week or two as it takes a while to run these days.