​
Custom configuration

Custom configuration is used to train models with advanced parameters in the Custom model training user interface.

​
Custom configuration scenarios

Custom configuration is typically used in the following scenarios:

​
Custom configuration parameters

This section contains the most commonly used and important configuration parameters. You can use these parameters in the:

​
Model type

To set the dataset and training defaults, enter the appropriate value in the dataset_config and trainer_config fields:
  • mlp_general. This is used for the general recurrent neural networks (RNN) model type.
  • mlp_ecommerce. This is used for an ecommerce RNN model type.
For example, to set General RNN model type, use: 
{
  "dataset_config": "mlp_general",
  "trainer_config": "mlp_general"
}

​
Model parameters

The parameters nested in trainer_config let you set training and model encoder parameters. The most important parameters are:
  • "trainer_config/text_processor_config": "word_en". Determines which type of tokenization and embedding is used as the base for the recurrent neural network (RNN) model. For example, word or byte-pair encoding (BPE). For information about values, see Text processor.
    Important The syntax for the text_processor_config parameter name must use a forward slash / and not a period . because it sets a group of parameters using a group name. Other nested parameters use a period ..
  • "trainer_config.encoder_config.rnn_names_list": ["gru"]. Determines which bi-directional recurrent neural network (RNN) layers are used. Options include gru and lstm.
  • "trainer_config.encoder_config.rnn_units_list": [128]. The number of units for each recurrent neural network (RNN) layer.
    Because this is a bi-directional RNN, the encoder’s vector size is two times larger than the number of units in the last layer. For example, if one layer is 128 units, the output vector size is 256.
You must specify the same number of units for trainer_config.encoder_config.rnn_units_list and its similarly-named trainer_config.encoder_config.rnn_names_list RNN layer. For example, rnn_units_list needs to be the same size as rnn_names_list.

​
Advanced custom configuration parameters

This section describes the most common advanced custom configuration parameters you can alter. Modifying the values does not typically provide a significant boost in quality. However, setting values incorrectly may cause serious quality degradation.

​
Advanced model parameters

  • "trainer_config.trn_batch_size": null. The batch size to be used for a single model training update. By default, an appropriate batch size is automatically determined based on the dataset size. If the field is set to null, the batch size is also automatically determined based on the dataset size.
  • "trainer_config.num_epochs": 64. The number of epochs the training data must complete. An epoch is a full cycle where training data passes through the designated algorithms. During one epoch, the model processes all the training data examples (queries and index documents) at least one time.
  • "trainer_config.monitor_patience": 8. The number of epochs the training passes before it stops if there is no validation metric improvement during the epochs. The best model state based on the monitor validation metric is used as the final model.
    • For the general RNN, the mrr@3 metric is monitored and the monitor_patience default value is 8.
    • For the ecommerce RNN, the ndcg@5 metric is monitored and the monitor_patience default value is 16.
  • "trainer_config.encoder_config.emb_spdp": 0.3. This field provides a regularization effect, which is the process to simplify result answers. The regularization is applied between the token embeddings layer and the first recurrent neural network (RNN) layer.
  • "trainer_config.encoder_config.emb_trainable". This field determines if fine-tuning of the token embeddings is enabled. Examples of token embedding are word or byte pair encoding (BPE) token vectors. If set, it can improve the quality of the model if the query contains less natural language that negatively impacts training. Because the embeddings layer is the largest layer in the network, the process to improve the model requires enough training data to prevent overfitting.
    The default values are:
    • "trainer_config.encoder_config.emb_trainable": false. For mlp_general models.
    • "trainer_config.encoder_config.emb_trainable": true. For mlp_ecommerce models.

​
Custom configuration examples

To create a custom configuration, set a dataset_config and trainer_config. To minimize dimished quality in the training model, only change field parameters that deviate from the default.
For detailed information about dataset_config for the index and query files, see use case training data.

​
Configuration