Skip to main content
Version: v1.2.0

Release notes

v1.2.0 | Sep 2, 2022​

UI & UX

  • The buttons to start a new experiment, rename an experiment, stop an experiment, and delete an experiment, can now be located inside a Kebab menu in the experiment's row in the experiments table.
  • The buttons to edit a dataset, delete a dataset, and start a new experiment with the selected dataset, can now be located inside a Kebab menu in the dataset's row in the datasets table.

Datasets

  • Data connector: H2O Hydrogen Torch now supports Azure data lake (as a data connector). To learn more, see Data connectors.
  • Image object detection: H2O Hydrogen Torch now supports several dataset (data) formats for an image object detection experiment. Supported formats are as follows:
  • Image semantic segmentation: H2O Hydrogen Torch now supports several dataset (data) formats for an image semantic segmentation experiment. Supported formats are as follows:
  • Image instance segmentation: H2O Hydrogen Torch now supports several dataset (data) formats for an image instance segmentation experiment. Supported formats are as follows:
  • New additional checks/validations have been introduced when importing a dataset. To learn more, see Import a dataset.
  • You can now extend a dataset with new data (e.g., to increase your dataset size). To learn more, see Extend a dataset with new data.
  • You can now merge imported datasets into one. To learn more, see Merge datasets.

Experiments

  • H2O Hydrogen Torch now offers several grid search modes. To learn more, see Grid search.
  • You can now enter custom values for any grid search hyperparameter values.
  • Now, for image classification experiments, in the following three tabs, H2O Hydrogen Torch for each class separately highlights with Gradient-weighted Class Activation Mapping (Grad-CAM) the areas of an image the model considered the most when generating a prediction for the image:
  • NLP interpretability: Now, for a regression and classification text experiment, on the Validation interpretation insights tab, H2O Hydrogen Torch highlights with Integrated Gradients the words the model considered the most when generating a prediction for the text. To learn more, see Validation interpretation insights.
  • Semi-supervised learning: H2O Hydrogen Torch now supports a semi-supervised mode where labeled, and unlabeled data are supported. For problem types that support labeled and unlabeled data, H2O Hydrogen Torch will first train a model with the provided labeled data. Immediately after, it will predict so-called pseudo labels for the provided unlabeled data. At last, H2O Hydrogen Torch will retrain the model while utilizing the original and generated pseudo labels. As labeling can be expensive, unlabeled data is beneficial and can improve the model quality. As follows are problem types that support a semi-supervised mode:
  • You can now start a new experiment using the pre-trained weights from a completed experiment.
  • You can now specify multiple text columns for the following problem types:
  • For audio experiments, H2O Hydrogen Torch now allows you to define several audio settings or have H2O Hydrogen Torch calculate the Sample rate and Training chunk seconds using the training samples and assign default values to several audio settings.
  • You can now define the learning rate to apply to certain model layers.
  • H2O Hydrogen Torch now caches masks for image segmentation experiments, making the experiment much faster.
  • H2O Hydrogen Torch increased the speed of the metric calculations for image object detection experiments.
  • H2O Hydrogen Torch now supports dynamic sequence padding for text experiments, making the experiments faster.
  • H2O Hydrogen Torch now supports mixed precision inference.
  • H2O Hydrogen Torch now has new additional state-of-the-art techniques for model training.
  • You can now compare experiments regardless of their containment inside a project. You can compare experiments in the experiments table. To learn more, see Compare experiments.

Projects

  • You can now organize experiments into projects. Projects enable you to organize experiments into groups related to a specific business problem or use case. In addition, projects in H2O Hydrogen Torch enable you to view and manage in one place experiments stemming from the same grid search experiment. To learn more, see Create a project.
  • H2O Hydrogen Torch will now organize experiments generated from a selected grid search mode into one project. To learn more, see Grid search.
  • You can now compare experiments in a project to understand their similarities and differences. For example, comparing experiments in one project can help you visually understand similarities in the experiments learning rate, training batch loss, validation batch loss, and validation metric. To learn more, see Compare a project's experiments.

Predictions

  • Image instance segmentation: For an image instance segmentation, the validation and test .pkl files have been restructured. To learn more, see Image instance segmentation.
  • The H2O Hydrogen Torch Python scoring pipeline for v1.2.0 only supports Ubuntu 16.04+ with Python 3.8. To learn more, see Deploy model.
  • You can now name your predictions.

Documentation

  • You can now access a new set of tutorials where tutorials are divided into two major areas: experiments and predictions. To learn more, see Tutorials.
  • All new features and settings for v1.2.0 have been documented.
  • All documentation files before v1.2.0 have been reviewed and rewritten and, in some cases, restructured (e.g., Dataset formats).
Feedback