Skip to main content
Version: Next

FAQs

H2O Hydrogen Torch is an application that lets novice and expert data scientists build deep learning models for diverse problem types in computer vision, natural language, and audio. No code is required.

H2O Hydrogen Torch lets you generate good models with default hyperparameter values derived from best practices used by top Kaggle grandmasters. You can tune default hyperparameter values to obtain the best state-of-the-art deep learning models. Simple and interactive charts in H2O Hydrogen Torch let you understand the impact of selected hyperparameter values on the training process. For model deployment, you can deploy built models in the H2O Hydrogen Torch UI, external Python environments, or directly to H2O MLOps.

H2O Hydrogen Torch optimizes and simplifies training deep learning models by streamlining the training process.

The below sections provide answers to frequently asked questions. If you have additional questions, please send them to cloud-feedback@h2o.ai.

General​

How is H2O Hydrogen Torch different than any other software capable of training deep learning models?​

The following points distinguish H2O Hydrogen Torch on the market:

  • UI built with H2O Wave for non-code deep learning model training
  • Variety of text, image, and audio base problem types
  • Model training best practices from top Kaggle grandmasters
  • Search for optimal model parameters to get the best model
  • Easy flexible deployment

How often do new versions come out?​

The frequency of major new H2O Hydrogen Torch releases has historically been about every four to eight weeks.

What is the difference between H2O Hydrogen Torch, Driverless AI (DAI), and H2O-3?​

Driverless AI (DAI), H2O-3, and H2O Hydrogen Torch are designed to democratize machine learning. Therefore at first glance, they might have functionality overlaps.

  1. H2O-3 is an open-source product, while H2O Hydrogen Torch is a commercial offering part of the H2O AI Cloud.
  2. DAI and H2O Hydrogen Torch are two different machine learning backends, and there is a set of tasks where both of the backends can be applied. However, there is a set of tasks (such as computer vision, natural language processing, and audio) where deep learning models are expected to outperform other methods:
    • DAI uses classic machine learning techniques and typically selects gradient boosting trees or linear regression models.
    • H2O Hydrogen Torch is for fitting deep learning models exclusively.
    • DAI has computer vision and natural language processing support with limited functionality. H2O Hydrogen Torch is explicitly developed and maintained for deep learning models in focus, therefore providing more functionality for fitting such models and more types of problems you can solve.
    • It is advised to use DAI for machine learning problems that rely on tabular data (iid, time series, unsupervised tasks).
    • It is advised to use Hydrogen Torch for machine learning problems based on images, short texts, videos, and audio data.

Do you need a license to run H2O Hydrogen Torch?​

To run H2O Hydrogen Torch, you don't need a license. H2O Hydrogen Torch comes with the H2O AI Cloud for free as a core service.

What are some example use cases that I can achieve through H2O Hydrogen Torch?​

H2O Hydrogen Torch enables novice and expert data scientists to solve an array of use cases in various computer vision and natural language problems. To learn more, see Use cases.

How can I request a new feature for H2O Hydrogen Torch (e.g., a new loss function)?​

To request a new feature, please get in touch with cloud-feedback@h2o.ai.

How can I interpret my built model?​

H2O Hydrogen Torch displays random train visual samples after augmentations. As well, it visualizes validation samples and their prediction. For more information, see Experiment tabs.

Datasets​

What are the dataset formats for every supported problem type in H2O Hydrogen Torch?​

Before uploading your dataset to H2O Hydrogen Torch, your dataset needs to be preprocessed in a particular format depending on the problem type it aims to solve. To learn about the different formats, see Dataset formats.

What are the data types H2O Hydrogen Torch supports?​

H2O Hydrogen Torch supports image, text, and audio data types.

Note

Videos: Although H2O Hydrogen Torch doesn't natively support video data types, you can still work with video data by breaking the video into individual frames to treat them as images later. This approach allows you to leverage the available image models and techniques in H2O Hydrogen Torch. The recommended workflow involves the following steps:

  1. Use a frame-by-frame approach: Split the video into individual frames (images).
  2. Label each frame: To prepare your video data (images) for training and testing, you'll need to label each frame (image).
    • H2O.ai provides a convenient solution called H2O Label Genie, which enables you to rapidly label your datasets for various tasks, including computer vision, natural language processing, and audio. With H2O Label Genie, you can easily create labeled datasets compatible with H2O Hydrogen Torch. By leveraging zero-shot learning models, H2O Label Genie streamlines the labeling process, allowing you to focus on other aspects of your project
  3. Select an appropriate image model: Select an appropriate image model to train and test, considering factors such as the type of analysis, use case, available computing resources, and desired accuracy.
  4. Feed individual frames into the model: Feed individual frames (images) into the model to generate predictions.
  5. Reference other frames: Reference (score) the model using other frames from the same video or different videos, enabling you to evaluate the model's performance and fine-tune it as needed.
    • You can display a set of scored frames as a video output

Following this workflow, you can effectively adapt image models for video analysis tasks in H2O Hydrogen Torch, even though it doesn't directly support video data types.

Does the data used in H2O Hydrogen Torch need to be labeled?​

Yes. In particular, datasets must be labeled and formated in a particular way depending on the problem type. To learn more, see Dataset formats.

Does H2O Hydrogen Torch support unlabeled data (unsupervised learning)?​

H2O Hydrogen Torch supports a semi-supervised mode where labeled and unlabeled data are supported. For problem types that support labeled and unlabeled data, H2O Hydrogen Torch first trains a model with the provided labeled data. Immediately after, it predicts so-called pseudo labels for the provided unlabeled data. At last, H2O Hydrogen Torch retrains 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:

What number of supported image and audio extensions are available for image and audio processing in H2O Hydrogen Torch?​

H2O Hydrogen Torch supports various image and audio extensions.

Experiments​

Does H2O Hydrogen Torch support custom model architectures and custom weights?​

  • All text problem types support custom model architectures and custom weights. Custom architectures are supported through custom Hugging Face models (requires the enablement of the trust_remode_code setting). You can read more about creating such models in Hugging Face's guide.
  • Image and audio classification/regression/metric learning problem types support custom weights following the timm's guide. One can provide a backbone name in the H2O Hydrogen Torch settings in the following form: hf_hub:nateraw/resnet18-random. However, custom model architectures are not available.
  • Other problem types (image object detection, image segmentation, text-to-speech, etc.) do not support custom architectures or custom weights.

What is grid search, and how does it work?​

Grid search enables you to define several options for certain hyperparameters (grid search hyperparameters). To learn more, see Grid search.

H2O Hydrogen Torch runs only one experiment at a time while other experiments are queued? Or does the number depend on GPUs?​

H2O Hydrogen Torch can run multiple experiments simultaneously, but only on different GPUs. For example, starting one experiment on GPUs 1-2 and another on 3-4 enables the experiments to run simultaneously. In contrast, starting a new experiment on GPUs 1-2 leads H2O Hydrogen Torch to queue the experiment.

Multiple-fold experiments can automatically run several experiments simultaneously. Also, grid search automatically enables the running of several experiments simultaneously.

Note

The following setting controls the number of GPUs per experiment in multiple-fold or grid search experiments: Number of GPUs per run.

In the context of text classification, can I specify the length of the token input sequence?​

Yes, you can specify the maximum length of the token input sequence with the Max length experiment setting.

Which experiment settings should I train an audio classification or regression model with audio files longer than one minute?​

For an audio classification or regression model, the default settings in H2O Hydrogen Torch truncate the audio files to one minute. To train an audio classification or regression model with audio files longer than one minute, you need to modify the following settings as follows (when defining the hyperparameter values for the model):

  • Experience level: Master
    note

    To learn more about this setting, see Experience levels.

  • Audio parameters: Manual
  • Training chunk seconds: 1200
  • Inference chunk seconds: 1200

As the audio classification or regression model can require more GPU memory, the following setting can help:

  • Automatically adjust batch size: On
  • Hop size: A larger hop size (for example, 1024)
  • Mel frequency bins: A low number of frquency bins (for example, 64)
  • Backbone: A smaller backbone (for example, resnet34)
Feedback