Load and Retrain a Pretrained NN with a MATLAB GUI

Discrete Values vs. GUI

This task was about loading the pretrained NN with MATLAB® App (GUI) by clicking the desired path in the images. This method is the same method used in JupyterLab, but alternatively, since I already have steering angles with their corresponding frames, we can train a model to predict a steering angle based on new frames.

Pretrained Neural Network

A pretrained neural network is a model that’s already been trained on a large dataset for a general task (e.g., image classification on ImageNet). You can fine-tune it on your smaller, task-specific dataset—leveraging learned features to save time and improve performance.

Types of Pretrained NN

Tabelle 12: MATLAB Pretrained NN

Name	Examples	Description
Image Classification	AlexNet VGG-16/VGG-19 ResNet-50/101 SqueezeNet GoogLeNet MobileNet-v2 EfficientNet-b0	Pretrained NN for image classification tasks.
Object Detection	YOLO v2 SSD (Single Shot Multibox Detector) a Faster R-CNN	Pretrained models for detecting and localizing objects.
Semantic Segmentation	FCN-8s U-Net SegNet DeepLab-v3+	Pretrained NN to label each pixel in an image.
Super-Resolution	SRCNN ESPCN EDSR pretrained models	Pretrained NN for upscaling low-resolution images.
Image Denoising	DnCNN	Pretrained NN for removing noise from images.
Image-to-Image Translation	pix2pix CycleGAN	Pretrained NN for tasks like style transfer and domain adaptation.
Instance Segmentation	MATLAB provides Mask R-CNN (e.g., ResNet-50-FPN backbone) pretrained on COCO Annotator	NN for detecting and segmenting individual object instances.

All the above models can be downloaded from here [1].

How to Load, Retrain, and Fine-Tune a Pretrained NN

MATLAB has an integrated toolbox to run and test deep learning. ResNet18 is a pre-trained network we will be using to train our AI and then fine tune it.

1. Load: Loading a pretrained neural network means importing its saved architecture and learned weights into MATLAB’s workspace. This gives you a ready-to-use model without training from scratch [[2]][[3]].
2. Retrain: Retraining involves further training the entire loaded network on your own dataset. All layers’ weights are updated, often starting from the pretrained values [[4]][[5]].
3. Finetune: Fine-tuning freezes most of the network’s layers and only updates a subset (typically higher-level layers) on new data. You use a lower learning rate so the pretrained features aren’t drastically altered [[6]][[7]].

Tabelle 13: MATLAB Pretrained NN

Name	Application	Description
Load	Fig. 23: Load NN	Loading a fine-tuned pre-trained NN or starting with a new pre-trained NN (ResNet-18).
Retrain	Fig. 24: Retrain an NN	Inside the while hasFrame(v) loop, every time frameBatch reaches MiniBatchSize, trainedNet is called.
Fine-tune	Fig. 25: Fine-tune	The pre-trained ResNet-18 is fine-tuned by using a low InitialLearnRate (1e-4) in trainingOptions and training on your new steering batches.