Small Is Beautifull

Edge AI for Image Aesthetic Assesment

Created on May 17|Last edited on November 17

Comment

﻿
Edge AI High level.Hardware:Mobile device best Image selectionImage Aesthetic Quality Assessment (IAQA):Selecting Best network ConViT, CvTs and CNNs. Related Work and Networks Pipeline and ML OpsPre Processing and Meta-dataTraining and AugmentationTraining and TuningTesting CPU inference on Rasberry Pi 4 (4GB)Deployment at the Edge (Monitoring Inference)Code & Reproducibility
﻿
Edge AI﻿
 High level.There are a huge number of examples including Industrial Internet of Things(IoT), ML deployed in mobile phones, urban sensing, cartography, and research.
Edge of a network and generally GPU pre-trained (but not always)
Involves both specialist and no specialist hardware.
Within computer vision models designed for Edge AI computer vision are optimized that have the smallest number of FLOPs, small size (eg. 5mb) , low-latency (fast inference), has small power consumption. 
Hal Burche's Graph of the internet. 
Hardware:
Nvidia Jetson Nano 
﻿
Mobile device best Image selectionThis example uses transfer learning on mobile/edge-optimized CNNs and CvT's.
This example uses:
Messy real-world data.
Training small rather than large.
Uses Weights and Biases for training and edge device network deployment. 
[timm](https://rwightman.github.io/pytorch-image-models/)
Image Aesthetic Quality Assessment (IAQA):Method A Binary Classifier in Regressiors shoes 
Dataset first web scraped by N. Murray et al. ﻿in 2012 from https://www.dpchallenge.com/ 
Using transfer learning of pre-training mobile/edge-optimized devices using Ross Weighmans' Timm repository and Pytorch Framework.
Quantized and thresholded into binary class from a probability distribution;
Trained to classify good or bad images;
Comparing models.
Selecting Best network ConViT, CvTs and CNNs. Exploring reproducibility and domain-adaptation using transfer learning and getting SOTA metrics for the network(s) pre-trained on Image-Net 1k using  Ross Weighman's timm package, the docs for this can be found here.
The latest bleeding-edge release is  available to check out here, At the time of writing is needed if you want to use the latest MobileViTs∈{xxs,xs,s}Mobile ViTs\in \{xxs,xs,s\}MobileViTs∈{xxs,xs,s}﻿﻿
pip install git+https://github.com/rwightman/pytorch-image-models.git
Related Work and Networks ﻿

PaperURLModel Name
MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision TransformerhereMobileViT
Self-training with Noisy StudenthereEfficientNet NoisyStudent (B0-B7, L2)
Adversarial Examples Improve Image RecognitionhereEfficientNet AdvProp (B0-B8)
EfficientNet: Rethinking Model Scaling for Convolutional Neural NetworkshereEfficientNet (B0-B7)
EfficientNet-EdgeTPU: Creating Accelerator-Optimized Neural Networks with AutoMLhereEfficientNet-EdgeTPU (S, M, L)
EffecientNetV2: Smaller Models and Faster TraininghereEfficientNet V2
MobileNetV2: Inverted Residuals and Linear BottleneckshereMobileNet-V2
Model Rubik's Cube: Twisting Resolution, Depth and Width for TinyNetshereTinyNet
FBNet: Hardware-Aware Efficient ConvNet Design via Differentiable Neural Architecture SearchhereFBNet-C
MnasNet: Platform-Aware Neural Architecture Search for MobilehereMNASNet B1, A1 (Squeeze-Excite), and Small
Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4 HourshereSingle-Path NAS
﻿
Pipeline and ML Ops
Pre Processing and Meta-data
Training and Augmentation
Training and Tuning
Testing CPU inference on Rasberry Pi 4 (4GB)
Deployment at the Edge (Monitoring Inference)﻿
Run set3
﻿
Code & Reproducibility

typesource
colab .ipynb
repohere
datahere
﻿
﻿
﻿

Paper	URL	Model Name
MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer	here	MobileViT
Self-training with Noisy Student	here	EfficientNet NoisyStudent (B0-B7, L2)
Adversarial Examples Improve Image Recognition	here	EfficientNet AdvProp (B0-B8)
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks	here	EfficientNet (B0-B7)
EfficientNet-EdgeTPU: Creating Accelerator-Optimized Neural Networks with AutoML	here	EfficientNet-EdgeTPU (S, M, L)
EffecientNetV2: Smaller Models and Faster Training	here	EfficientNet V2
MobileNetV2: Inverted Residuals and Linear Bottlenecks	here	MobileNet-V2
Model Rubik's Cube: Twisting Resolution, Depth and Width for TinyNets	here	TinyNet
FBNet: Hardware-Aware Efficient ConvNet Design via Differentiable Neural Architecture Search	here	FBNet-C
MnasNet: Platform-Aware Neural Architecture Search for Mobile	here	MNASNet B1, A1 (Squeeze-Excite), and Small
Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4 Hours	here	Single-Path NAS

type	source
colab .ipynb
repo	here
data	here

Add a comment