Model Card for beit_base_patch16_224_pruned_65

This is a prunned version of the timm/beit_base_patch16_224.in22k_ft_in22k_in1k model in a toch-dag format.

This model has rougly 65% of the original model FLOPs with minimal metrics drop.

Model	KMAPPs*	M Parameters	Accuracy (224x224)
timm/beit_base_patch16_224.in22k_ft_in22k_in1 (baseline)	673.2	86.5	85.23%
beit_base_patch16_224_pruned_65 (ours)	438 (65%)	56.7 (66%)	84.53% (↓ 0.7%)

*KMAPPs thousands of FLOPs per input pixel

KMAPPs(model) = FLOPs(model) / (H * W * 1000), where (H, W) is the input resolution.

The accuracy was calculated on the ImageNet-1k validation dataset. For details about image pre-processing, please refer to the original repository.

Model Details

Model Description

• Developed by: TCL Research Europe
• Model type: Classification / feature backbone
• License: Apache 2.0
• Finetuned from model: timm/beit_base_patch16_224.in22k_ft_in22k_in1k

Model Sources

• Repository: timm/beit_base_patch16_224.in22k_ft_in22k_in1k

How to Get Started with the Model

To load the model, You have to install torch-dag library, which can be done using pip by

pip install torch-dag

then, clone this repository

# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/TCLResearchEurope/beit_base_patch16_224_pruned_65

and now You are ready to load the model:

import torch_dag
import torch

model = torch_dag.io.load_dag_from_path('./beit_base_patch16_224_pruned_65')

model.eval()
out = model(torch.ones(1, 3, 224, 224))
print(out.shape)