Deep Train a Model using Clarifai_ResNet_AngularMargin – Clarifai

Need a model that can distinguish between large number of people or objects by using only a small number of sample images? Clarifai_ResNet_AngularMargin is here to help.

Clarifai_ResNet_AngularMargin consistently outperforms the state-of-the-art and has been implemented with efficient computational overhead so that predictions are made quickly.

Originally conceived asa a state-of-the-art face recognition method for obtaining highly discriminative features for face recognition, Clarifai_ResNet_AngularMargin has become a popular solution for many "visual search" use cases where accurate search results can be obtained by exaggerating the fine differences between objects.

Create your App in Portal

To begin, you will need to create an app to contain your project. You can learn more about creating deep training apps here.

Upload Your Training Data

Upload and label your training data in Portal or use our API. You can use this upload script for your convenience:

https://gitlab.infra.cf.ai/clarifai/clarifai/-/blob/master/conf/dst/upload_dataset_to_app.py

Create Model

Once your app is created, you will need to create a new model. You can do this in model mode, or by clicking the "Create a Model" button on your app home screen.

Choose Visual Embedder

ArcFace template is a "Visual Embedder" model type. Select Visual Embedder from the "Create a Custom Model" menu.

Choose "Clarifai_ResNet_AngularMargin", this gives you access to the ArcFace model template.

Choose the hyper-parameter settings that you would like to use for your model and click create model. You can find information about available hyper-parameter settings here.

Train Your Model

Finally, you will need to train your model. Just click the "Train Model" button in the top right corner of the screen.

You can check the status of your model training job in Explorer. The blue progress bar on in the lefthand sidebar will indicate training progress. Deep training can take longer than transfer learning. You can learn more about managing deep training projects here.

Evaluation

Once the model is trained you will want to check the evaluations metrics. Click "evaluate" for the model version that you would like to view. Once evaluation is complete you will be able to view the results by clicking "view". Learn more about model evaluation here.

Upload Your Training Data

Create Model

Choose Visual Embedder

Train Your Model

Evaluation

Related articles