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.
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.
Instead of creating a deep learning model from scratch, get a pretrained model, which you can apply directly or adapt to your task.
Explore MATLAB® Deep Learning Model Hub to access the latest models by category and get tips on choosing a model.
Load most models at the command line. For example:
net = darknet19;
Convert TensorFlow™, PyTorch®, and ONNX™ models to MATLAB networks by using an import function. For example:
net = importTensorFlowNetwork("EfficientNetV2L")
Apply pretrained models to image classification, computer vision, audio processing, lidar processing, and other deep learning workflows.
For image classification tasks, use a convolutional neural network such as NASNet-Mobile and EfficientNet. The majority of CNNs is trained on the ImageNet database.
Analyze images and videos using object detection (YOLO), semantic/instance segmentation (AdaptSeg/MASK R-CNN), and video classification (SlowFast).
Analyze point cloud data using classification (PointNet), object detection (PointPillars), and semantic segmentation (PointSeg).
There are many pretrained models to choose from, and each model has tradeoffs:
To get started, choose one of the faster models, such as SqueezeNet or GoogleNet. Then iterate quickly and try different preprocessing and training options.
Once you understand which settings work well, try a more accurate model, such as Inception-v3 or ResNet-50, and see if that improves your results.
When you deploy to edge devices such as Raspberry Pi® or FPGAs, choose a model with a low memory footprint, such as SqueezeNet or MobileNet-v2.