apple / ml-cvnets

CVNets: A library for training computer vision networks
https://apple.github.io/ml-cvnets
Other
1.76k stars 225 forks source link

add apple's new model: FastViT #73

Open rotem154154 opened 1 year ago

rotem154154 commented 1 year ago

Apple just released very exiting model named FastViT. is there any plan on relese it? https://arxiv.org/abs/2303.14189

jameslahm commented 1 year ago

Hi, I try to reproduce the FastViT. Here is the coreml model of my reproduced FastViT-T12. However, I find its mobile latency is much bigger than the latency data reported in Table 5 in the paper. I use iPhone 12 and Xcode coreml performance tool to measure the latency. According to Table 9 in the mobileone paper and the Table 5 in the FastViT paper, FastViT-T12 can obtain the similar mobile latency as MobileOne-S2, \ie, about 1.2ms.

image image

However, the actual latency I measured is MobileOne-S2: 1.13ms and FastViT-T12: 1.73ms.

image image

I wonder if there is something wrong with my reproduction. Could you give me some guide or would you mind sharing the code? Thanks a lot! Any advice is appreciated.

vizaiapp commented 1 year ago

@jameslahm can you link to repo?

shituo123456 commented 1 year ago

@jameslahm Can I see your source code?

Wikidepia commented 1 year ago

It seems to be available now: https://github.com/apple/ml-cvnets/blob/c668ad13d67b3e8829e810d4b6e16ad7f77fd4aa/cvnets/models/classification/fastvit.py. But there is no pretrained model yet.