0.1% Data Makes Segment Anything Slim
Zigeng Chen, Gongfan Fang, Xinyin Ma, Xinchao Wang
Learning and Vision Lab, National University of Singapore
Paper: [Arxiv]
- 🚀 March 22, 2024: Awesome-Efficient-Segment-Anything is now available. Find more efficient SAMs here.
- 🚀 January 10, 2024: Run SlimSAM in your browser with 🤗 Transformers.js (demo).
- 🚀 January 9, 2024: Quickly loading using huggingface 🤗 🤗 🤗 .
- 🚀 January 7, 2024: Release models using uniform local pruning for easier state dict loading.
- 🚀 December 19, 2023: Release the Colab example for SlimSAM.
- 🚀 December 11, 2023: Release the training code, inference code and pre-trained models for SlimSAM.
from PIL import Image
from transformers import SamModel, SamProcessor
model = SamModel.from_pretrained("Zigeng/SlimSAM-uniform-50").to("cuda")
processor = SamProcessor.from_pretrained("Zigeng/SlimSAM-uniform-50")
img_url = "https://huggingface.co/ybelkada/segment-anything/resolve/main/assets/car.png"
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert("RGB")
input_points = [[[450, 600]]] # 2D localization of a window
inputs = processor(raw_image, input_points=input_points, return_tensors="pt").to("cuda")
outputs = model(**inputs)
masks = processor.image_processor.post_process_masks(outputs.pred_masks.cpu(), inputs["original_sizes"].cpu(), inputs["reshaped_input_sizes"].cpu())
scores = outputs.iou_scores
SlimSAM is a novel data-efficient SAM compression method that achieves superior performance with extremely less training data. The essence of SlimSAM is encapsulated in the alternate slimming framework which effectively enhances knowledge inheritance under severely limited training data availability and exceptional pruning ratio. Diverging from prior techniques, our framework progressively compresses the model by alternately pruning and distilling distinct, decoupled sub-structures. Disturbed Taylor pruning is also proposed to address the misalignment between the pruning objective and training target, thereby boosting the post-distillation after pruning.
SlimSAM yields significant performance improvements while demanding over 10 times less training data than any other existing compression methods. Even when compared to the original SAM, SlimSAM achieves approaching performance while reducing parameter counts to merely 1.4% (9.1M), MACs to 0.8% (23G), and requiring only 0.1% (10k) of the SAM training data.
Qualitative comparison of results obtained using point prompts, box prompts, and segment everything prompts are shown.
We conducted a comprehensive comparison encompassing performance, efficiency, and training costs with other SAM compression methods and structural pruning methods.
The code requires python>=3.8, as well as pytorch>=1.7 and torchvision>=0.8. Please follow the instructions here to install both PyTorch and TorchVision dependencies. Installing both PyTorch and TorchVision with CUDA support is strongly recommended.
Install with
pip install -e .
The following optional dependencies are necessary for mask post-processing, saving masks in COCO format.
pip install opencv-python pycocotools matplotlib
We use the original SA-1B dataset in our code. See here for an overview of the datastet. The dataset can be downloaded here.
The download dataset should be saved as:
<train_data_root>/
sa_xxxxxxx.jpg
sa_xxxxxxx.json
......
<val_data_root>/
sa_xxxxxxx.jpg
sa_xxxxxxx.json
......
To decode a mask in COCO RLE format into binary:
from pycocotools import mask as mask_utils
mask = mask_utils.decode(annotation["segmentation"])See here for more instructions to manipulate masks stored in RLE format.
The base model of our method is available. To enhance collaboration with our dependency dectection algorithm, we have split the original image encoder's qkv layer into three distinct linear layers: q, k, and v.
Click the links below to download the checkpoints of orginal SAM-B.
SAM-B: SAM-B model.
The check points of our SlimSAM are avalable. We release two versions, which are SlimSAM-50 (pruning ratio = 50%) and SlimSAM-77 (pruning ratio = 77%).
Click the links below to download the checkpoints for the corresponding pruning ratio.
SlimSAM-50: SlimSAM-50 model.SlimSAM-77: SlimSAM-77 model.
Above models can be instantiated by running
import torch
SlimSAM_model = torch.load(<model_path>)
SlimSAM_model.image_encoder = SlimSAM_model.image_encoder.module
def forward(self, x):
x = self.patch_embed(x)
if self.pos_embed is not None:
x = x + self.pos_embed
for blk in self.blocks:
x,qkv_emb,mid_emb,x_emb = blk(x)
x = self.neck(x.permute(0, 3, 1, 2))
return x
import types
funcType = types.MethodType
SlimSAM_model.image_encoder.forward = funcType(forward, SlimSAM_model.image_encoder)
SlimSAM_model.to(device)
SlimSAM_model.eval()SlimSAM-50-uniform: SlimSAM-50 model.SlimSAM-77-uniform: SlimSAM-77 model.
Above models can be instantiated by running
import torch
from segment_anything import sam_model_registry
model_type = 'vit_p50'
checkpoint = 'checkpoints/SlimSAM-50-uniform.pth'
SlimSAM_model = sam_model_registry[model_type](checkpoint=checkpoint)
SlimSAM_model.to(device)
SlimSAM_model.eval()First download SlimSAM-50 model or SlimSAM-77 model for inference
We provide detailed instructions in 'inference.py' on how to use a range of prompts, including 'point' and 'box' and 'everything', for inference purposes.
CUDA_VISIBLE_DEVICES=0 python inference.py
First download a SAM-B model into 'checkpoints/' as the base model.
The model after step1 is saved as 'checkpoints/vit_b_slim_step1_.pth'
CUDA_VISIBLE_DEVICES=0 python prune_distill_step1.py --traindata_path <train_data_root> --valdata_path <val_data_root> --prune_ratio <pruning ratio> --epochs <training epochs>
The model after step2 is saved as 'checkpoints/vit_b_slim_step2_.pth'
CUDA_VISIBLE_DEVICES=0 python prune_distill_step2.py --traindata_path <train_data_root> --valdata_path <val_data_root> --prune_ratio <pruning ratio> --epochs <training epochs> --model_path 'checkpoints/vit_b_slim_step1_.pth'
You can adjust the training settings to meet your specific requirements. While our method demonstrates impressive performance with just 10,000 training data, incorporating additional training data will further enhance the model's effectiveness
If you use SlimSAM in your research, please use the following BibTeX entry. Thank you!
@article{chen20230,
title={0.1\% Data Makes Segment Anything Slim},
author={Chen, Zigeng and Fang, Gongfan and Ma, Xinyin and Wang, Xinchao},
journal={arXiv preprint arXiv:2312.05284},
year={2023}
}SAM (Segment Anything) [bib]
@article{kirillov2023segany,
title={Segment Anything},
author={Kirillov, Alexander and Mintun, Eric and Ravi, Nikhila and Mao, Hanzi and Rolland, Chloe and Gustafson, Laura and Xiao, Tete and Whitehead, Spencer and Berg, Alexander C. and Lo, Wan-Yen and Doll{\'a}r, Piotr and Girshick, Ross},
journal={arXiv:2304.02643},
year={2023}
}Torch Pruning (DepGraph: Towards Any Structural Pruning) [bib]
@inproceedings{fang2023depgraph,
title={Depgraph: Towards any structural pruning},
author={Fang, Gongfan and Ma, Xinyin and Song, Mingli and Mi, Michael Bi and Wang, Xinchao},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={16091--16101},
year={2023}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent