EfficientViT: Multi-Scale Linear Attention for High-Resolution Dense Prediction (paper, poster)

News

If you are interested in getting updates, please join our mailing list here.

• [2023/09/18] EfficientViT for Segment Anything Model (SAM) is released. EfficientViT SAM runs at 1009 images/s on A100 GPU, compared to ViT-H (12 images/s), mobileSAM (297 images/s), and nanoSAM (744 image/s, but much lower mIoU)
• [2023/09/12] EfficientViT is highlighted by MIT home page and MIT News.
• [2023/07/18] EfficientViT is accepted by ICCV 2023.

Demo

EfficientViT-L0 for Segment Anything (1009 image/s on A100 GPU)

EfficientViT-L1 for Semantic Segmentation (45.9ms on Nvidia Jetson AGX Orin, 82.716 mIoU on Cityscapes)

About EfficientViT Models

EfficientViT is a new family of ViT models for efficient high-resolution dense prediction vision tasks. The core building block of EfficientViT is a lightweight, multi-scale linear attention module that achieves global receptive field and multi-scale learning with only hardware-efficient operations, making EfficientViT TensorRT-friendly and suitable for GPU deployment.

Third-Party Implementation/Integration

• timm: link
• X-AnyLabeling: link

Getting Started

Installation

conda create -n efficientvit python=3.10
conda activate efficientvit
conda install -c conda-forge mpi4py openmpi
pip install -r requirements.txt

Dataset

Our code expects the ImageNet dataset directory to follow the following structure:

imagenet
├── train
├── val

Our code expects the Cityscapes dataset directory to follow the following structure:

cityscapes
├── gtFine
|   ├── train
|   ├── val
├── leftImg8bit
|   ├── train
|   ├── val

Our code expects the ADE20K dataset directory to follow the following structure:

ade20k
├── annotations
|   ├── training
|   ├── validation
├── images
|   ├── training
|   ├── validation

Pretrained Models

Latency/Throughput is measured on NVIDIA Jetson Nano, NVIDIA Jetson AGX Orin, and NVIDIA A100 GPU with TensorRT, fp16. Data transfer time is included.

Segment Anything

In this version, the EfficientViT segment anything models are trained using the image embedding extracted by SAM ViT-H as the target. The prompt encoder and mask decoder are the same as SAM ViT-H.

ImageNet

All EfficientViT classification models are trained on ImageNet-1K with random initialization (300 epochs + 20 warmup epochs) using supervised learning.

Cityscapes

ADE20K

Usage

# segment anything
from efficientvit.sam_model_zoo import create_sam_model

efficientvit_sam = create_sam_model(
  name="l2", weight_url="assets/checkpoints/sam/l2.pt",
)
efficientvit_sam = efficientvit_sam.cuda().eval()

from efficientvit.models.efficientvit.sam import EfficientViTSamPredictor

efficientvit_sam_predictor = EfficientViTSamPredictor(efficientvit_sam)

from efficientvit.models.efficientvit.sam import EfficientViTSamAutomaticMaskGenerator

efficientvit_mask_generator = EfficientViTSamAutomaticMaskGenerator(efficientvit_sam)

# classification
from efficientvit.cls_model_zoo import create_cls_model

model = create_cls_model(
  name="l3", weight_url="assets/checkpoints/cls/l3-r384.pt"
)

# semantic segmentation
from efficientvit.seg_model_zoo import create_seg_model

model = create_seg_model(
  name="l2", dataset="cityscapes", weight_url="assets/checkpoints/seg/cityscapes/l2.pt"
)

model = create_seg_model(
  name="l2", dataset="ade20k", weight_url="assets/checkpoints/seg/ade20k/l2.pt"
)

Evaluation

Please run eval_sam_coco.py, eval_cls_model.py or eval_seg_model.py to evaluate our models.

Examples: segment anything, classification, segmentation

Visualization

Please run demo_sam_model.py to visualize our segment anything models.

Example:

# segment everything
python demo_sam_model.py --model l1 --mode all

# prompt with points
python demo_sam_model.py --model l1 --mode point

# prompt with box
python demo_sam_model.py --model l1 --mode box --box "[150,70,630,400]"

Please run eval_seg_model.py to visualize the outputs of our semantic segmentation models.

Example:

python eval_seg_model.py --dataset cityscapes --crop_size 1024 --model b3 --save_path demo/cityscapes/b3/

Export TFLite

To generate TFLite files, please refer to tflite_export.py. It requires the TinyNN package.

pip install git+https://github.com/alibaba/TinyNeuralNetwork.git

Example:

python tflite_export.py --export_path model.tflite --task seg --dataset ade20k --model b3 --resolution 512 512

Export ONNX

To generate ONNX files, please refer to onnx_export.py.

To export ONNX files for EfficientViT SAM models, please refer to the scripts shared by CVHub.

Training

Please see TRAINING.md for detailed training instructions.

Contact

Han Cai: [email protected]

TODO

• ImageNet Pretrained models
• Segmentation Pretrained models
• ImageNet training code
• EfficientViT L series, designed for cloud
• EfficientViT for segment anything
• EfficientViT for super-resolution
• Segmentation training code

Citation

If EfficientViT is useful or relevant to your research, please kindly recognize our contributions by citing our paper:

@article{cai2022efficientvit,
  title={Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition},
  author={Cai, Han and Gan, Chuang and Han, Song},
  journal={arXiv preprint arXiv:2205.14756},
  year={2022}
}