This topic details procedures to create a new custom model and deploy existing custom models. For information about custom models, see Custom model training user interface.

​
Create a new model

To create a new model, navigate to the megamenu and click Models > Custom Models.
  1. In the Custom Models screen, click + Create New Model.
  2. In the Train New Model screen, enter a user-defined value in the Name field, select a Region where the training of the model occurs, and click Continue.
    These fields cannot be changed after the model is created. If you need a different name or a model in a different region, you must create another model with the new information.
  3. Click one of the following values for the Use Case:
    • General. This use case uses general data schema with mrr@3 as the primary retrieval metric. This option is typically for informational sites.
    • ecommerce. This use case uses ecommerce catalog and signals data schema with ndcg@5 as the primary retrieval metric. This option is typically used for online sales sites.
    • Classification. This use case uses uses classification data schema for texts labeling with F1 scores as the primary evaluation metric. This option is used to compute similarity scores between the incoming text and the labels.
If your Google Cloud Storage (GCS) bucket is publicly accessible, the service account key is optional and you can click Continue without entering a key. If you enter a key, it is validated when you click Continue. If the key is not valid, the model cannot be configured.
  1. In the Query Training Data field, enter the location of signals in the training data in Google Cloud Storage (GCS). For configuration information, see use case training data query file.
  2. In the Index Training Data field, enter the location of the index training data in Google Cloud Storage (GCS). For configuration information, see use case training data index file.
  3. In the Upload Service Account Key section, the information about how to generate your service account key. To upload a JSON blob that contains the default information from GCS, click Click here to browse and select the JSON file. To view an example, click Sample Service Account Key.
    If your Google Cloud Storage (GCS) bucket is publicly accessible, the service account key is optional and you can click Continue without entering a key. If you enter a key, it is validated when you click Continue. If the key is not valid, the model cannot be configured.
  4. Click Continue. You can select values in the fields on this screen, or you can click Switch to Manual Entry to manually enter a configuration.
  5. To select values on this screen, click one of the following values for the Model Type:
    • General RNN. This model type is the RNN encoder with frozen pre-trained token embeddings. The general model is typically used by knowledge management sites that contain information, such as news or documentation. For more information about this model, see custom embedding models.
    • ecommerce RNN. This model type is the RNN encoder with tuneable token embeddings. The ecommerce model is typically used by sites that provide products for purchase. For more information about this model, see custom embedding models.
    • Transformer RNN. This model is the RNN built on top of a frozen pre-trained Transformer (BERT) encoder. This model is typically used for ecommerce use cases due to slower model encoding speed and less flexibility for domain specific language. For more information about this model, see Transformer RNN models.
  6. Select values in the following fields:
    • Text Processor Config. The type of tokenization and embedding to be used as the base for the selected recurrent neural network (RNN) model type. For example, word or byte-pair encoding (BPE). For information about values, see Text processor.
    • RNN Layers. In Layer 1 of the RNN Layers field, select the bi-directional recurrent neural network (RNN) layers to use (gru or lstm), and then select the unit size of the layer. Options are gru and lstm. Layer 2 is optional. If needed, select the RNN and unit.
  7. Click Save & Run to save and execute training for the model. To exit without saving and running the model, click Cancel.

​
Manual entry

If you click Switch to Manual Entry, the Custom Config field displays.
  1. In the Custom Config field, enter the configuration parameters for your custom model.
    For detailed information about configuration options for a model, see Custom configuration.
  2. Click Save & Run to execute the model training. To exit without saving and running the model, click Cancel. To return to creating a model without entering custom configuration, click Switch to Guided Entry.

​
Create a new deployment

To deploy a custom model, navigate to the megamenu and click Models > Custom Models.
  1. In the Custom Models screen, click the model you want to deploy, and then click Deployment Details > + New Deployment.
  2. Complete the following fields:
    • Enter a unique value to describe the deployment in the Name field.
    • Click and drag the scale to set the number of Minimum Replicas for the model. The maximum value is 12.
    • Click and drag the scale to set the number of Maximum Replicas for the model. The maximum number is 12.
    • Select the Region where the model is to be deployed.
      These fields cannot be changed after the model is deployed. If you need different values, you must create another deployment with the new information.
  3. Click Save & Run to deploy the model. To exit without saving the information and deploying the model, click Cancel. The deployment displays on the Deployment Details screen for the model.