Frequently Asked Questions

• Can I work with other modalities besides tables?

  • Besides tables, we have successfully used Lale for text, images, and time-series. In fact, Lale even works for multi-modal data, using the & combinator to specify different preprocessing paths per modality.

• Can I work with other tasks besides classification?

  • Besides classification, we mostly use Lale for regression. That said, you can define your own scoring metrics for evaluation, and pass them to automation tools to guide their search. The following notebook includes an example disparate_impact_scorer: https://github.com/IBM/lale/blob/master/examples/demo_aif360.ipynb

• I get an error when I instantiate an operator imported from Lale. What’s wrong?

  • Lale raises errors on invalid hyperparameter values or combinations. This ensures that the operators are used correctly. So don’t be surprised if you get any errors when you initialize Lale operators with some hyperparameter values. Chances are that those hyperpameters or combinations of hyperparameters are invalid. If not, please contact us.

• The algorithm I want to use is not present in Lale. Can I still use it?

  • Some of the features of Lale can be used if the algorithm implementation follows the scikit-learn conventions of fit/predict or fit/transform. You can turn any such operator into a Lale operator using lale_op = lale.operators.make_operator(non_lale_op). If you want to get full Lale support for your own operator, we have a separate guide for how to do that: https://github.com/IBM/lale/blob/master/examples/docs_new_operators.ipynb

• Can I use Lale for deep learning?

  • There are multiple facets to this question. The Lale library already includes a few DL operators such as a BERT transformer, a ResNet classifier, and an MLP classifier. Lale can perform joint algorithm selection and hyperparameter optimization over pipelines involving these operators. Furthermore, users can wrap additional DL operators as described. On the other hand, Lale does not currently support full-fledged neural architecture search (NAS). It can only perform architecture search when that is exposed through hyperparameters, such as hidden_layer_sizes for MLP.

• How does the search space generation work?

  • Lale includes a search space generator that takes in a planned pipeline and the schemas of the operators in that pipeline, and returns a search space for your auto-ML tool of choice. Our arXiv paper describes how that works in detail: https://arxiv.org/abs/1906.03957

• Does Lale optimize for computational performance?

  • While Lale focuses mostly on types and automation, we have also done a little bit of work on computational performance. However, it has not been a major focus. If you encounter pain-points, please reach out to us.

• What is the relationship between Lale and IBM products?

  • Lale is free and open-source and does not depend on any commercial products. It is available under the Apache license and only requires the other open-source packages listed in setup.py. Lale is used by IBM’s AutoAI SDK. The AutoAI SDK provides API access to various services on IBM cloud, including advanced pipeline search optimizers, cloud-hosted notebooks in Watson Studio, storage for datasets in Cloud Object Storage, storage for historical pipelines, deploying trained pipelines as a scoring service in Watson Machine Learning, etc. Lale does not require the AutoAI SDK to run, you can use it without the AutoAI SDK.