Release notes¶

H2O Hydrogen Torch v1.1.0 | May 03, 2022¶

For image and audio regression and classification experiments, GradCam interpretability is now available. To learn more, see Validation interpretation insights.
H2O Hydrogen Torch now can redefine global setting values used across H2O Hydrogen Torch (e.g., AWS credentials). To learn more, see Global settings.
For an image object detection experiment, the following new object detection models are available: Faster Region-based Convolutional Neural Networks (RCNN) & Convolutional One-stage Object Detector (FCOS). To learn more, see Model type.
H2O Hydrogen Torch supports YAML config files' import and export functionality. Users can now download config settings of finished experiments, make changes, and re-upload them when starting a new experiment. To learn more, see Import config from YAML.
Through a demo mode, H2O Hydrogen Torch (HHT) can now offer the ability to view HHT simplistically. Users with demo-mode enabled can view pre-run (non-functional) datasets and experiments for all supported problem types. To learn more, see Demo mode.

A new list of demo datasets is available. To learn more, see Preprocessed datasets
New validation options for experiments
Faster download of artifacts
Nested settings when defining the settings of an experiment
Allow any number of channels for input images
Handling of missing values in labels and observations
An optional gradient checkpointing setting for NLP models
The number of GPUS setting is now a Grid search hyperparameter
Improved output formats
UX improvements and bug fixes
Label class name support for certain problem types
Best epoch functionality when training experiments
Plotting improvements
ETA estimation improvements and style changes when displaying a list of experiments

New landing page for the documentation website
New documentation for all the settings and features for audio classification and regression
New structure to view all experiment settings for all support problem types
New documentation for all new features and settings

Image Instance Segmentation: This new problem type can be seen as a refined version of semantic segmentation where we also detect each instance of a category. For example, identify each cell in an image.

Renamed output files
Removed option to specify image extension
New dropout option for text models
New hyperparameter settings for object detection
Classification data can be imported with a dense format or with one-hot-encoded labels
H2O Hydrogen Torch can now import data for text experiments in a .pq, .txt, and .zip file format
Changed custom augmentation format
Improved validation plots for text experiments
Improved output format for MLOps Pipeline

During and after model training, you can observe actual augmented images used to train in the Train Data Insights tab
During and after model training, metrics beyond the one selected for the experiment are calculated for classification and regression experiments. For classification, you can modify the threshold for the calculated metrics (e.g., confusion matrix). Confusion matrices are available for multi-class and multi-label classification
NLP problems during and after model training can view insights around random, worst, and best samples
In the new Test Predictions Insights tab, during and after a prediction, test samples and their prediction are render
Inference-specific settings are available for new predictions (e.g., test time augmentations)

Renamed Hydrogen Torch to H2O Hydrogen Torch
Added the following documentation for an image instance segmentation experiment:
- dataset format
- all available settings
- prediction files
Updated the H2O Hydrogen Torch overview video
Restructured and edited all documentation pages to improve readability
Added new documentation about all tabs around datasets, experiments, and predictions

Fixes an issue with unwarranted final activation for text regression problems
Fixes an issue with indexing in the insights tab for text classification problems

New UX rework with a new home page and side-navigation bar
New user complexity roles
New problem type: Text Span Prediction
New problem type: Text Sequence to Sequence
Internal logging rework and new live logs are available when an experiment is running
Flexible backbone selection using a combo box: a custom backbone can be entered manually
Reworking of label specification of different problem types
Migration to Wave 0.18
New option to train on full data
Neptune logging is now available
New option to calculate train metric

Grid search is now available
Boken can now be used to interact with insights generated for the following problem types: Semantic Segmentation, Object Detection, Image Classification, Metric Learning, and Image Regression.
Several UI improvements
H2O Hydrogen Torch is now able to auto-select the correct problem type where possible
Users can now enable Multi-GPU training for token classification and semantic segmentation

New problem type: Token Classification
Built models can now be used to score on new data through the H2O Hydrogen Torch UI
Automatic scoring pipelines
Initial MLOps pipeline
Experiment queuing
New demo datasets

Feedback