moonbox-ai/MobileSAM-MLX

MobileSAM v2 MLX

MLX builds of MobileSAM v2, made by Moonbox for Apple Silicon Macs. Each precision is a folder; keep a folder's files together. Only the parts listed below are converted: the rest of the pipeline comes from the original repository unless the section below says otherwise.

Files

The last column is how far each build is from the original PyTorch model on the same inputs (see the section below for what was measured).

Precision	File	Size	sha256	Checked against the original
bf16	bf16/config.json	0.3 kB	18c4a51430fe	mask IoU 0.996 mean / 0.972 worst
bf16	bf16/mlx_model.py	20.7 kB	57f29a58bfa3	mask IoU 0.996 mean / 0.972 worst
bf16	bf16/model.safetensors	19.6 MB	948290b7a2e3	mask IoU 0.996 mean / 0.972 worst
fp16	fp16/config.json	0.3 kB	d6876abfe8f9	mask IoU 0.999 mean / 0.988 worst
fp16	fp16/mlx_model.py	20.7 kB	57f29a58bfa3	mask IoU 0.999 mean / 0.988 worst
fp16	fp16/model.safetensors	19.6 MB	e8a52d0c013f	mask IoU 0.999 mean / 0.988 worst
8bit	8bit/config.json	0.4 kB	a5c89d1aea72	below threshold (experimental): mask IoU 0.989 mean / 0.899 worst
8bit	8bit/mlx_model.py	20.7 kB	57f29a58bfa3	below threshold (experimental): mask IoU 0.989 mean / 0.899 worst
8bit	8bit/model.safetensors	12.7 MB	d5daba02bdec	below threshold (experimental): mask IoU 0.989 mean / 0.899 worst
6bit	6bit/config.json	0.4 kB	6cd61ce1c00d	below threshold (experimental): mask IoU 0.989 mean / 0.914 worst
6bit	6bit/mlx_model.py	20.7 kB	57f29a58bfa3	below threshold (experimental): mask IoU 0.989 mean / 0.914 worst
6bit	6bit/model.safetensors	10.8 MB	6699f046a09d	below threshold (experimental): mask IoU 0.989 mean / 0.914 worst
4bit	4bit/config.json	0.4 kB	4f556f197ff7	below threshold (experimental): mask IoU 0.922 mean / 0.012 worst
4bit	4bit/mlx_model.py	20.7 kB	57f29a58bfa3	below threshold (experimental): mask IoU 0.922 mean / 0.012 worst
4bit	4bit/model.safetensors	9.0 MB	9beb710a859d	below threshold (experimental): mask IoU 0.922 mean / 0.012 worst

How to run

MobileSAM (TinyViT-5M image encoder + SAM's prompt encoder and mask decoder, weights/mobile_sam.pt) ported to MLX: the point / box segmentation half of MobileSAM v2. Each folder holds model.safetensors, config.json and mlx_model.py, the port (only mlx, numpy and pillow needed), with SamPredictor's interface:

import numpy as np, mlx_model
from PIL import Image
model = mlx_model.load("fp16")
rgb = np.array(Image.open("in.jpg").convert("RGB"))
emb = mlx_model.embed(model, rgb)                                     # once per image
masks, scores, low_res, logits = mlx_model.predict(model, emb, points=[[320, 240]], labels=[1], multimask=True)
masks, scores, _, _ = mlx_model.predict(model, emb, box=[100, 80, 400, 380], multimask=False)

Not included: MobileSAM v2's object-aware prompt sampler (a YOLOv8 from ultralytics, AGPL-3.0, weights on Google Drive only) and its prompt-guided decoder (shipped as a pickled PyTorch module). For "segment everything", feed boxes from any detector to predict.

How they were checked. Against the original in PyTorch fp32 on 8 photos, one point prompt (3 masks) and one box prompt each: IoU of each of the 32 masks with PyTorch's. Pass: IoU >= 0.95 on every mask.

Precision	Size	Mask IoU mean / worst	Status
fp16/	19.7 MB	0.999 / 0.988	passes
bf16/	19.7 MB	0.996 / 0.972	passes
8bit/	12.7 MB	0.989 / 0.899	below threshold, experimental
6bit/	10.9 MB	0.989 / 0.914	below threshold, experimental
4bit/	9.0 MB	0.922 / 0.012	below threshold, experimental: some prompts give a wrong mask

Use fp16/: the model is 10 M parameters, so quantizing saves 7-11 MB and costs accuracy. About 65 ms per image (encoder + two prompts) on an Apple M5.

How these were made

From MobileSAM v2 at revision f706ad9c4e: the original weights loaded in MLX and stored as bf16 or fp16, or quantized with MLX's affine quantization (8bit, 6bit, 4bit, group size 64; layers MLX can't quantize, such as convolutions and norms, stay bf16). Every build was checked against the original before upload; the sha256 column above is what the files must hash to.

License

apache-2.0. These files are a derivative of the original model and carry its license: it is copied in this repository, and it governs your use of these files. The original model card follows unchanged.

---

Original model card

Faster Segment Anything (MobileSAM) and Everything (MobileSAMv2)

:pushpin: MobileSAMv2, available at ResearchGate and arXiv, replaces the grid-search prompt sampling in SAM with object-aware prompt sampling for faster segment everything(SegEvery).

:pushpin: MobileSAM, available at ResearchGate and arXiv, replaces the heavyweight image encoder in SAM with a lightweight image encoder for faster segment anything(SegAny).

Support for ONNX model export. Feel free to test it on your devices and share your results with us.

A demo of MobileSAM running on CPU is open at hugging face demo. On our own Mac i5 CPU, it takes around 3s. On the hugging face demo, the interface and inferior CPUs make it slower but still works fine. Stayed tuned for a new version with more features! You can also run a demo of MobileSAM on your local PC.

:grapes: Media coverage and Projects that adapt from SAM to MobileSAM (Thank you all!)

• 2023/07/03: joliGEN supports MobileSAM for faster and lightweight mask refinement for image inpainting with Diffusion and GAN.

• 2023/07/03: MobileSAM-in-the-Browser shows a demo of running MobileSAM on the browser of your local PC or Mobile phone.

• 2023/07/02: Inpaint-Anything supports MobileSAM for faster and lightweight Inpaint Anything

• 2023/07/02: Personalize-SAM supports MobileSAM for faster and lightweight Personalize Segment Anything with 1 Shot

• 2023/07/01: MobileSAM-in-the-Browser makes an example implementation of MobileSAM in the browser.

• 2023/06/30: SegmentAnythingin3D supports MobileSAM to segment anything in 3D efficiently.

• 2023/06/30: MobileSAM has been featured by AK for the second time, see the link AK's MobileSAM tweet. Welcome to retweet.

• 2023/06/29: AnyLabeling supports MobileSAM for auto-labeling.

• 2023/06/29: SonarSAM supports MobileSAM for Image encoder full-finetuing.

• 2023/06/29: Stable Diffusion WebUIv supports MobileSAM.

• 2023/06/28: Grounding-SAM supports MobileSAM with Grounded-MobileSAM.

• 2023/06/27: MobileSAM has been featured by AK, see the link AK's MobileSAM tweet. Welcome to retweet.

:star: How is MobileSAM trained? MobileSAM is trained on a single GPU with 100k datasets (1% of the original images) for less than a day. The training code will be available soon.

:star: How to Adapt from SAM to MobileSAM? Since MobileSAM keeps exactly the same pipeline as the original SAM, we inherit pre-processing, post-processing, and all other interfaces from the original SAM. Therefore, by assuming everything is exactly the same except for a smaller image encoder, those who use the original SAM for their projects can adapt to MobileSAM with almost zero effort.

:star: MobileSAM performs on par with the original SAM (at least visually) and keeps exactly the same pipeline as the original SAM except for a change on the image encoder. Specifically, we replace the original heavyweight ViT-H encoder (632M) with a much smaller Tiny-ViT (5M). On a single GPU, MobileSAM runs around 12ms per image: 8ms on the image encoder and 4ms on the mask decoder.

• The comparison of ViT-based image encoder is summarzed as follows:

  Image Encoder	Original SAM	MobileSAM
  Parameters	611M	5M
  Speed	452ms	8ms

• Original SAM and MobileSAM have exactly the same prompt-guided mask decoder:

  Mask Decoder	Original SAM	MobileSAM
  Parameters	3.876M	3.876M
  Speed	4ms	4ms

• The comparison of the whole pipeline is summarized as follows:

  Whole Pipeline (Enc+Dec)	Original SAM	MobileSAM
  Parameters	615M	9.66M
  Speed	456ms	12ms

:star: Original SAM and MobileSAM with a point as the prompt.

:star: Original SAM and MobileSAM with a box as the prompt.

:muscle: Is MobileSAM faster and smaller than FastSAM? Yes! MobileSAM is around 7 times smaller and around 5 times faster than the concurrent FastSAM. The comparison of the whole pipeline is summarzed as follows:

Whole Pipeline (Enc+Dec)	FastSAM	MobileSAM
Parameters	68M	9.66M
Speed	64ms	12ms

:muscle: Does MobileSAM aign better with the original SAM than FastSAM? Yes! FastSAM is suggested to work with multiple points, thus we compare the mIoU with two prompt points (with different pixel distances) and show the resutls as follows. Higher mIoU indicates higher alignment.

mIoU	FastSAM	MobileSAM
100	0.27	0.73
200	0.33	0.71
300	0.37	0.74
400	0.41	0.73
500	0.41	0.73

Installation

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 Mobile Segment Anything:

pip install git+https://github.com/ChaoningZhang/MobileSAM.git

or clone the repository locally and install with

git clone git@github.com:ChaoningZhang/MobileSAM.git
cd MobileSAM; pip install -e .

Demo

Once installed MobileSAM, you can run demo on your local PC or check out our HuggingFace Demo.

It requires latest version of gradio.

cd app
python app.py

Getting Started

The MobileSAM can be loaded in the following ways:

from mobile_sam import sam_model_registry, SamAutomaticMaskGenerator, SamPredictor

model_type = "vit_t"
sam_checkpoint = "./weights/mobile_sam.pt"

device = "cuda" if torch.cuda.is_available() else "cpu"

mobile_sam = sam_model_registry[model_type](checkpoint=sam_checkpoint)
mobile_sam.to(device=device)
mobile_sam.eval()

predictor = SamPredictor(mobile_sam)
predictor.set_image(<your_image>)
masks, _, _ = predictor.predict(<input_prompts>)

or generate masks for an entire image:

from mobile_sam import SamAutomaticMaskGenerator

mask_generator = SamAutomaticMaskGenerator(mobile_sam)
masks = mask_generator.generate(<your_image>)

Getting Started (MobileSAMv2)

Download the model weights from the checkpoints.

After downloading the model weights, faster SegEvery with MobileSAMv2 can be simply used as follows:

cd MobileSAMv2
bash ./experiments/mobilesamv2.sh

ONNX Export

MobileSAM now supports ONNX export. Export the model with

python scripts/export_onnx_model.py --checkpoint ./weights/mobile_sam.pt --model-type vit_t --output ./mobile_sam.onnx

Also check the example notebook to follow detailed steps. We recommend to use onnx==1.12.0 and onnxruntime==1.13.1 which is tested.

BibTex of our MobileSAM

If you use MobileSAM in your research, please use the following BibTeX entry. :mega: Thank you!

@article{mobile_sam,
  title={Faster Segment Anything: Towards Lightweight SAM for Mobile Applications},
  author={Zhang, Chaoning and Han, Dongshen and Qiao, Yu and Kim, Jung Uk and Bae, Sung-Ho and Lee, Seungkyu and Hong, Choong Seon},
  journal={arXiv preprint arXiv:2306.14289},
  year={2023}
}

Acknowledgement

This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)),(No.2019-0-01287, Evolvable Deep Learning Model Generation Platform for Edge Computing)

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}
}

TinyViT (TinyViT: Fast Pretraining Distillation for Small Vision Transformers) [bib]

@InProceedings{tiny_vit,
  title={TinyViT: Fast Pretraining Distillation for Small Vision Transformers},
  author={Wu, Kan and Zhang, Jinnian and Peng, Houwen and Liu, Mengchen and Xiao, Bin and Fu, Jianlong and Yuan, Lu},
  booktitle={European conference on computer vision (ECCV)},
  year={2022}