Skip to main content

Customize a Training Job

The SliceX AI Trainer allows you a high degree of customization over several model and training parameters for all tasks. Currently, we support the following :

  • Batch size: In general, when training on smaller datasets, we recommend using a lower batch size. A good rule of thumb is having a batch size not greater than 1/5th size of your training data. By default it is set to 64.
  • Number of epochs: limited to 20 at the moment, note that our models tend to converge under 10 epochs in most cases. By default it is set to 10
  • Learning rate: By default it is set to 1e-3.
  • Pretrained SliceX AI models for initialization.
  • Lowercase: Choose to make your training job case-sensitive (or not!) as your task demands.
  • Model family: Review the task specific SliceX AI™ Cloud Platform model families here.

For text classification we also support three additional options:

  • Data augmentation
  • Language of the pretrained SliceX AI model (if using for initialization): We currently provide pretrained SliceX AI models for model initialization that have been optimized for three languages-English, Spanish, and Japanese. Note that there are no language specific limitations on custom training itself. SliceX AI APIs can support multilinguality by default.
  • Tokenization: Option for multilingual data that might benefit from added tokenization eg. in non-segmented languages like Chinese and Japanese. Check out a tutorial for configuring the SliceX AI Trainer for multilingual data here.

These custom options are packaged under either preprocessing, training_config or model_config in the training REQUEST (quick reference here).

An example REQUEST body for a binary text classification training job customization is:

Binary Text Classification: Training Job Configuration Example
{
  "preprocessing": {
  "lowercase": true,
  "add_tokenization":"" 
    },
  "training_config": {
    "batch_size": 64,
    "learning_rate": 0.001,
    "num_epochs": 5,
    "language": "english",
    "data_augmentation": true,
    "use_pretrained": true
  },
  "model_config": {
    "name": "model_name",
    "type": "binary-text-classification",
    "family": "Papaya",
    "size": "mini"
  },
  "dataset_url": "DATASET_URL"
}
info

The different task types are:

  • Binary Text Classification: "binary-text-classification"
  • Multiclass Text Classification: "multiclass-text-classification"
  • Multilabel Text Classification: "multilabel-text-classification"
  • Question Answering: "question-answering"
  • Sequence Labeling: "sequence-labeling"