Deep Learning Models

Pretrained Deep Learning Models

Take advantage of model architectures developed by the deep learning research community. Popular models offer a robust architecture and skip the need to start from scratch.

Tips in Selecting a Model

There are many pretrained models to choose from, and each model will have tradeoffs:

Size: How much memory does the model need
The final location of the model will determine how much the network size needs to be considered.
When choosing to deploy to a low memory system, choose a model specifically designed for this task.
See Models for Edge Deployment
Accuracy: How well does the model perform prior to retraining
Typically, a model that performs well for the ImageNet dataset indicates a model that has learned informative features, and could perform well in new, similar tasks as well.
Explore Higher Accuracy Models
Prediction Speed: How fast can the model predict on new images
While prediction speed can vary based on many factors such as hardware and batch size, speed will also vary based on architecture of the chosen model, and the size of model.
Compare prediction speeds with Simple Models to Get Started.

Explore tradeoffs between models in the following sections.

To import any model into MATLAB, the structure is:

>> net = networkname

i.e.

>> net = alexnet
 >> net = resnet50

If the model is not already downloaded, a link will be provided to download the model in MATLAB.

Simple Models to Get Started

You can iterate on these models quickly and try out different settings such as data preprocessing steps and training options. Once you have a feel for which settings work well, try a more accurate network to see if it improves your results.

Explore examples:

Interactively Modify a Deep Learning Network for Transfer Learning

Train Deep Learning Network to Classify New Images

Higher Accuracy Models

Explore models that are highly effective for image-based workflows, such as image classification, object detection, and semantic segmentation.

For Object detection workflows:

DarkNet-19, DarkNet-53, and ResNet-50 are often used as the foundation for object detection problems and YOLO workflows. See examples on object detection using Yolov2 and Yolov3.

For Semantic Segmentation workflows:

Using any of the predefined network architectures provides a convenient starting point for training semantic segmentation networks. These are commonly used layer architectures for semantic segmentation problems: