Редактиране

Споделяне чрез

ResNet

  • Статия

This article describes how to use the ResNet component in Azure Machine Learning designer, to create an image classification model using the ResNet algorithm..

This classification algorithm is a supervised learning method, and requires a labeled dataset.

Note

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 a model and a 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 ResNet

Refer to this paper for more details about ResNet.

How to configure ResNet

  1. Add the ResNet component to your pipeline in the designer.

  2. For Model name, specify name of a certain ResNet structure and you can select from supported resnet: 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152', 'resnet152', 'resnext50_32x4d', 'resnext101_32x8d', 'wide_resnet50_2', 'wide_resnet101_2'.

  3. 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.

  4. For Zero init residual, specify whether to zero-initialize the last batch norm layer in each residual branch. If selected, the residual branch starts with zeros, and each residual block behaves like an identity. This can help with convergence at large batch sizes according to https://arxiv.org/abs/1706.02677.

  5. Connect the output of ResNet component, training and validation image dataset component to the Train PyTorch Model.

  6. Submit the pipeline.

Results

After pipeline run is completed, to use the model for scoring, connect the Train PyTorch Model to Score Image Model, to predict values for new input examples.

Technical notes

Component parameters

Name Range Type Default Description
Model name Any Mode resnext101_32x8d Name of a certain ResNet structure
Pretrained Any Boolean True Whether to use a model pre-trained on ImageNet
Zero init residual Any Boolean False Whether to zero-initialize the last batch norm layer in each residual branch

Output

Name Type Description
Untrained model UntrainedModelDirectory An untrained ResNet model that can be connected to Train PyTorch Model.

Next steps

See the set of components available to Azure Machine Learning.