search

leondgarse / keras_cv_attention_models

Keras beit,caformer,CMT,CoAtNet,convnext,davit,dino,efficientdet,edgenext,efficientformer,efficientnet,eva,fasternet,fastervit,fastvit,flexivit,gcvit,ghostnet,gpvit,hornet,hiera,iformer,inceptionnext,lcnet,levit,maxvit,mobilevit,moganet,nat,nfnets,pvt,swin,tinynet,tinyvit,uniformer,volo,vanillanet,yolor,yolov7,yolov8,yolox,gpt2,llama2, alias kecam
MIT License
595 stars 95 forks source link

EfficientViT-B0/B1/B2/B3 Models #119

Closed MarkWijkhuizen closed 1 year ago

MarkWijkhuizen commented 1 year ago

The original EfficientVit paper mentions 4 model, namely EfficientVit-B0/B1/B2/B3.

These models are published on the paper's official GitHub repository.

In the latest release EfficientVit-M1/M2/M3/M4/M5 were added, which does differ from the official models.

Would it be possible to add the EfficientVit-B0/B1/B2/B3 from their official GitHub repository?

leondgarse commented 1 year ago

Uh, they got a same name. These M ones are from Github microsoft/Cream/EfficientViT, paper 2305.07027 EfficientViT: Memory Efficient Vision Transformer with Cascaded Group Attention, and I put the wrong one...

MarkWijkhuizen commented 1 year ago

What a coincidence, these models with the identical names are released within a month of each other. With a name like EfficientVit this is not exactly unsurprising... If you would be in the position to add the other EfficientVit models, that would be highly appreciated!

leondgarse commented 1 year ago

It's almost done, just uploading. Just wondering where should I put them, as the B and M series are actually totally different.

MarkWijkhuizen commented 1 year ago

Thanks for picking it up so quickly!

Where to put them is a good question.

The B and M models are from different authors with different architectures, this should be explicitly stated. From a programming point of view it would be convenient to have them under the same EfficientVit class, since their purpose is the same: small efficient vision transformers.

As the project admin you know best what to do :)

leondgarse commented 1 year ago

Ya, currently they are both placed under keras_cv_attention_models/efficientvit and acceptable through kecam.efficientvit, and also separately acceptable in kecam.efficientvit_m and kecam.efficientvit_b. Just be it.

MarkWijkhuizen commented 1 year ago

Perfect, pulled the updates locally and all working smoothly. Many thanks for adding the EfficientVit-B models!