We propose a novel hybrid Mamba-Transformer backbone, denoted as MambaVision, which is specifically tailored for vision applications. Our core contribution includes redesigning the Mamba formulation to enhance its capability for efficient modeling of visual features. In addition, we conduct a comprehensive ablation study on the feasibility of integrating Vision Transformers (ViT) with Mamba. Our results demonstrate that equipping the Mamba architecture with several self-attention blocks at the final layers greatly improves the modeling capacity to capture long-range spatial dependencies. Based on our findings, we introduce a family of MambaVision models with a hierarchical architecture to meet various design criteria. For Image classification on ImageNet-1K dataset, MambaVision model variants achieve a new State-of-the-Art (SOTA) performance in terms of Top-1 accuracy and image throughput. In downstream tasks such as object detection, instance segmentation and semantic segmentation on MS COCO and ADE20K datasets, MambaVision outperforms comparably-sized backbones and demonstrates more favorable performance. Code: https://github.com/NVlabs/MambaVision.

Model Performance

MambaVision-L-21K is pretrained on ImageNet-21K dataset and finetuned on ImageNet-1K.

Name	Acc@1(%)	Acc@5(%)	#Params(M)	FLOPs(G)	Resolution
MambaVision-L-21K	86.1	97.9	227.9	34.9	224x224

In addition, the MambaVision models demonstrate a strong performance by achieving a new SOTA Pareto-front in terms of Top-1 accuracy and throughput.

Model Usage

It is highly recommended to install the requirements for MambaVision by running the following:

pip install mambavision

For each model, we offer two variants for image classification and feature extraction that can be imported with 1 line of code.

Image Classification

In the following example, we demonstrate how MambaVision can be used for image classification.

Given the following image from COCO dataset val set as an input:

The following snippet can be used for image classification:

Python

from transformers import AutoModelForImageClassification
from PIL import Image
from timm.data.transforms_factory import create_transform
import requests

model = AutoModelForImageClassification.from_pretrained("nvidia/MambaVision-L-21K", trust_remote_code=True)

# eval mode for inference
model.cuda().eval()

# prepare image for the model
url = 'http://images.cocodataset.org/val2017/000000020247.jpg'
image = Image.open(requests.get(url, stream=True).raw)
input_resolution = (3, 224, 224)  # MambaVision supports any input resolutions

transform = create_transform(input_size=input_resolution,
                             is_training=False,
                             mean=model.config.mean,
                             std=model.config.std,
                             crop_mode=model.config.crop_mode,
                             crop_pct=model.config.crop_pct)

inputs = transform(image).unsqueeze(0).cuda()
# model inference
outputs = model(inputs)
logits = outputs['logits'] 
predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])

The predicted label is brown bear, bruin, Ursus arctos.

Feature Extraction

MambaVision can also be used as a generic feature extractor.

Specifically, we can extract the outputs of each stage of model (4 stages) as well as the final averaged-pool features that are flattened.

The following snippet can be used for feature extraction:

Python

from transformers import AutoModel
from PIL import Image
from timm.data.transforms_factory import create_transform
import requests

model = AutoModel.from_pretrained("nvidia/MambaVision-L-21K", trust_remote_code=True)

# eval mode for inference
model.cuda().eval()

# prepare image for the model
url = 'http://images.cocodataset.org/val2017/000000020247.jpg'
image = Image.open(requests.get(url, stream=True).raw)
input_resolution = (3, 224, 224)  # MambaVision supports any input resolutions

transform = create_transform(input_size=input_resolution,
                             is_training=False,
                             mean=model.config.mean,
                             std=model.config.std,
                             crop_mode=model.config.crop_mode,
                             crop_pct=model.config.crop_pct)
inputs = transform(image).unsqueeze(0).cuda()
# model inference
out_avg_pool, features = model(inputs)
print("Size of the averaged pool features:", out_avg_pool.size())  # torch.Size([1, 640])
print("Number of stages in extracted features:", len(features)) # 4 stages
print("Size of extracted features in stage 1:", features[0].size()) # torch.Size([1, 80, 56, 56])
print("Size of extracted features in stage 4:", features[3].size()) # torch.Size([1, 640, 7, 7])

License:

NVIDIA Source Code License-NC