Where is profiler?

[bizzyvinci]

In the tutorial on analytics india, Profiler was used.

It was imported with from lightautoml.utils.profiler import Profiler, but there's no reference in the official documentation and importing it raises ModuleNotFoundError: No module named 'lightautoml.utils.profiler'.

Is Profiler gone or moved to another module or rebranded or ...?

[alexmryzhkov]

Hello @bizzyvinci,

That was our fault not to post release notes for the new versions of LightAutoML here on github. I'm currently working on it and add them to my answer. As you can see from them - in the latest version we remove Profiler because it takes too much computational resources to work and blocks some really important functionality like multiGPU NLP model training etc.

Release notes for 0.2.11:

• Added variable importances calculation functions for TabularAutoML and TabularUtilizedAutoML - example is in Tutorial_2.
• Added new functionality to return predictions from all algorithms before the blender layer for Train (OOF preds) and Prediction (Test preds) stages. To do that you can use return_all_predictions and weighted_blender_max_nonzero_coef params in general_params config section, example is here in cell 9
• Bug-fixes for custom metrics and TabularUtilizedAutoML custom config list setting (for example to make it work like model multistart) etc.

---

Release notes for 0.2.12:

• Added classififcation NLP models interpretation based on LIME.
• Better support for HuggingFace models for embeddings training and extraction.
• Bug-fixes for language argument in several embeddings extraction algos and tokenizers

---

Release notes for 0.2.13:

• Added regression NLP models interpretation based on LIME.

---

Release notes for 0.2.14-0.2.15:

• Bug-fixes based on user reviews in the core functionality and report generation.

---

Release notes for 0.2.16: General part:

• Removing profiler functionality (log_calls library usage) makes available to use DataParallel during Bert models training (to turn it on use multigpu param in nn_params)

NLP part:

• Migration to gensim >=4.
• Added functionaly to force big NLP models calculation without GPU (Random LSTM, Pooled Bert, etc).
• Changed logic of default parameters selection process: removing NN from pipeline if no GPU available, language and Bert model propagates to AutoNLP and NN via text_params, changed logic for embeddings size check in AutoNLP
• Added several normalization techniques for embeddings in AutoNLP module

Interpretation part:

• Minor changes in LIME part - now we can show AutoML prediction and the predictions scale colorbar.
• L2X local interpretation algorithm is now available for NLP models. It finds the most informative tokens for the target variable based on the AutoML model predictions by mutual information optimization. For more info please check this tutorial in the repo or the same in the nbviewer

[bizzyvinci]

Ok. Thanks.

https://github.com/sberbank-ai-lab/LightAutoML/releases is not up to date. Where are releases cataloged or updated?