EfficientVIT | Luxonis

Our new model ZOO works with DepthAI V3. Find out more in our documentation.

0 Likes

Model Details

Model Description

EfficientViT is a transformer-based model with novel multi-scale linear attention. Unlike previous state-of-the-art transformer-based models, it is built only by using lightweight and hardware-efficient operations. The model delivers better performance on edge devices.

Developed by:
Shared by:
Model type: Computer Vision
License:
Resources for more information:

Training Details

Training Data

The classification model was trained on dataset, actually on ImageNet-1K with random initialization (300 epochs + 20 warmup epochs) using supervised learning.

Testing Details

Metrics

The get more information about evaluation, please check .

Model	Resolution	ImageNet Top-1 Acc.	ImageNet Top-5 Acc.
EfficientViT-B1	224x224	79.39	94.35

Technical Specifications

Input/Output Details

Input:
- Name: input
  - Info: NCHW BGR un-normalized image
Output:
- Name: output
  - Info: The non-softmaxed values for 1000 classes.

Model Architecture

Backbone:
- Composed of an input stem and four stages, each reducing spatial resolution while increasing channel depth.
- EfficientViT modules are embedded in Stages 3 and 4 for efficient global and local feature extraction.
- Downsampling is handled by MobileNet-like blocks (MBConv).
Head:
- The head processes features from Stages 2, 3, and 4 using simple 1x1 convolutions and upsampling.
- Features are fused through addition to reduce computational overhead.
- Designed for simplicity and effectiveness, suitable for both segmentation and classification tasks.

Throughput

Model variant: efficientvit:b1-224x224

• Input shape: [1, 3, 224, 224] • Output shape: [1, 1000]

• Params (M): 9.095 • GFLOPs: 0.552

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	35.55	N/A
RVC4	FP16	229.41	3.14

* Benchmarked with , using 2 threads (and the DSP runtime in balanced mode for RVC4).

* Parameters and FLOPs are obtained from the package.

Utilization

Models converted for RVC Platforms can be used for inference on OAK devices. DepthAI pipelines are used to define the information flow linking the device, inference model, and the output parser (as defined in model head(s)). Below, we present the most crucial utilization steps for the particular model. Please consult the docs for more information.

Install DAIv3 and depthai-nodes libraries:

pip install depthai
pip install depthai-nodes

Define model:

model_description = dai.NNModelDescription(
    "luxonis/efficientvit:b1-224x224"
)

nn = pipeline.create(ParsingNeuralNetwork).build(
    <CameraNode>, model_description
)

Inspect model head(s):

ClassificationParser that outputs message (detected classes and scores).

Get parsed output(s):

while pipeline.isRuning():
    parser_output: Classifications = parser_output_queue.get()

Example

You can quickly run the model using our script. It automatically downloads the model, creates a DepthAI pipeline, runs the inference, and displays the results using our DepthAI visualizer tool. To try it out, run:

python3 main.py \
    --model luxonis/efficientvit:b1-224x224

Transformed-based classification model.
License	Apache 2.0 Commercial use
Downloads	155
Tasks	Classification
Model Types	ONNX

Model Variants

Name	Version	Available For	Created At	Deploy
		RVC2, RVC4	6 months ago