Skip to main content
Version: v1.4.0

Release notes

v1.4.0 | July 25, 2024​

Overview​

At the core of the new version, H2O Hydrogen Torch improves and expands on its currently supported problem types. These new improvements and expansions further expand the mission to offer a platform that enables you, with no coding experience, to train state-of-the-art deep neural networks on diverse problem types. The major points of this release are as follows:

  • H2O Hydrogen Torch now supports the following problem types:
    • Graph node classification
    • Graph node regression
    • Image + text classification (taking a combined input of image & text data)
  • For the following supported problem types, H2O Hydrogen Torch can now export a model to an open neural network exchange (ONNX) format:
    • Text regression and classification
    • Image classification, regression, and object detection
    • Audio classification and regression
  • The H2O Hydrogen Torch code base and scoring clients have been upgraded from Python 3.8 to Python 3.10

To learn more about the new release, observe the below subsections.

UI & UX​

  • New: H2O Hydrogen Torch now allows you to specify a Hugging Face API token to access private model repositories. To learn more, see Hugging Face API token.
  • New: Before starting a grid search experiment, H2O Hydrogen Torch now displays the number of experiments it will run.
    • Why?: It allows you to view the actual number of experiments to be conducted and adjust the number as needed.
  • New: Now, before starting an experiment, H2O Hydrogen Torch checks whether all the settings are configured correctly and throws a warning message in case of some discrepancies.
    • Why?: It allows you to detect any experiment misconfigurations very early.
  • Improvement: Several new UX improvements are available throughout the application while improving the user experience.

Datasets​

  • New: H2O Hydrogen Torch now supports importing data from Google Cloud Storage. To learn more, see Google Cloud Storage bucket name.
    • Why?: The Google Cloud Storage connector allows you to import data from the new source.
  • New: H2O Hydrogen Torch now allows you to check all the files during a dataset import. To learn more, see Validate sample files.
    • Why?: It enables you to identify and repair any broken or missing files in the dataset before modeling.
  • New: Now, you can download datasets from H2O Hydrogen Torch.
    • Why?: It allows you to download, explore, and edit datasets locally.

Experiments​

  • New: Now, H2O Hydrogen Torch supports the following problem types:
  • New: H2O Hydrogen Torch now supports LoRA (Low-Rank Adaptation) for text problem types. To learn more, see LoRA.
    • Why?: It allows to do parameter-efficient finetuning of large pretrained models.
  • New: H2O Hydrogen Torch now supports modeling for multi-channel audio inputs for audio problem types. To learn more, see Audio channels.
    • Why?: Previously, H2O Hydrogen torch averaged together multiple channels; now, H2O Hydrogen Torch processes multi-channel audios as is.
  • New: H2O Hydrogen Torch now displays the batch inference speed in the Summary tab and when comparing multiple experiments.
    • Why?: It allows you to choose the best model based on quality and a trade-off between quality and inference speed.

Predictions​

Documentation​

  • New: All new features and settings for v1.4.0 have been documented.
Feedback