This article describes how to use the DenseNet component in Azure Machine Learning designer, to create an image classification model using the Densenet algorithm.
This classification algorithm is a supervised learning method, and requires a labeled image directory.
This component does not support labeled dataset generated from Data Labeling in the studio, but only support labeled image directory generated from Convert to Image Directory component.
You can train the model by providing the model and the labeled image directory as inputs to Train Pytorch Model. The trained model can then be used to predict values for the new input examples using Score Image Model.
More about DenseNet
For more information on DenseNet, see the research paper, Densely Connected Convolutional Networks.
How to configure DenseNet
Add the DenseNet component to your pipeline in the designer.
For Model name, specify name of a certain DenseNet structure and you can select from supported DenseNet: 'densenet121', 'densenet161', 'densenet169', 'densenet201'.
For Pretrained, specify whether to use a model pre-trained on ImageNet. If selected, you can fine-tune model based on selected pre-trained model; if deselected, you can train from scratch.
For Memory efficient, specify whether to use checkpointing, which is much more memory-efficient but slower. For more information, see the research paper, Memory-Efficient Implementation of DenseNets.
Connect the output of DenseNet component, training, and validation image dataset component to the Train Pytorch Model.
Submit the pipeline.
|Model name||Any||Mode||densenet201||Name of a certain DenseNet structure|
|Pretrained||Any||Boolean||True||Whether to use a model pre-trained on ImageNet|
|Memory efficient||Any||Boolean||False||Whether to use checkpointing, which is much more memory efficient but slower|
|Untrained model||UntrainedModelDirectory||An untrained DenseNet model that can be connected to Train Pytorch Model.|
See the set of components available to Azure Machine Learning.