MobileViT is a light-weight, low latency convolutional neural network that combines MobileNetV2-style layers with a new block that replaces local processing in convolutions with global processing using transformers. As with ViT (Vision Transformer), the image data is converted into flattened patches before it is processed by the transformer layers. Afterwards, the patches are "unflattened" back into feature maps. This allows the MobileViT-block to be placed anywhere inside a CNN. MobileViT does not require any positional embeddings.

The model in this repo adds a DeepLabV3 head to the MobileViT backbone for semantic segmentation.

Intended uses & limitations

You can use the raw model for semantic segmentation. See the model hub to look for fine-tuned versions on a task that interests you.

How to use

Here is how to use this model:

python

from transformers import MobileViTFeatureExtractor, MobileViTForSemanticSegmentation
from PIL import Image
import requests

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)

feature_extractor = MobileViTFeatureExtractor.from_pretrained("apple/deeplabv3-mobilevit-x-small")
model = MobileViTForSemanticSegmentation.from_pretrained("apple/deeplabv3-mobilevit-x-small")

inputs = feature_extractor(images=image, return_tensors="pt")

outputs = model(**inputs)
logits = outputs.logits
predicted_mask = logits.argmax(1).squeeze(0)

Currently, both the feature extractor and model support PyTorch.

Training data

The MobileViT + DeepLabV3 model was pretrained on ImageNet-1k, a dataset consisting of 1 million images and 1,000 classes, and then fine-tuned on the PASCAL VOC2012 dataset.

Training procedure

Preprocessing

At inference time, images are center-cropped at 512x512. Pixels are normalized to the range [0, 1]. Images are expected to be in BGR pixel order, not RGB.

Pretraining

The MobileViT networks are trained from scratch for 300 epochs on ImageNet-1k on 8 NVIDIA GPUs with an effective batch size of 1024 and learning rate warmup for 3k steps, followed by cosine annealing. Also used were label smoothing cross-entropy loss and L2 weight decay. Training resolution varies from 160x160 to 320x320, using multi-scale sampling.

To obtain the DeepLabV3 model, MobileViT was fine-tuned on the PASCAL VOC dataset using 4 NVIDIA A100 GPUs.

Evaluation results

Model	PASCAL VOC mIOU	# params	URL
MobileViT-XXS	73.6	1.9 M	https://huggingface.co/apple/deeplabv3-mobilevit-xx-small
MobileViT-XS	77.1	2.9 M	https://huggingface.co/apple/deeplabv3-mobilevit-x-small
MobileViT-S	79.1	6.4 M	https://huggingface.co/apple/deeplabv3-mobilevit-small

BibTeX entry and citation info

bibtex

@inproceedings{vision-transformer,
title = {MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer},
author = {Sachin Mehta and Mohammad Rastegari},
year = {2022},
URL = {https://arxiv.org/abs/2110.02178}
}