View an experiment's tabs

Overview

An experiment contains the following tabs during and after model training:

• Charts
• Summary
• Metrics
• Train data insights
• Validation prediction insights
• Validation interpretation insights
• Logs
• Config
• MLOps

Instructions

To access the tabs of a completed or running experiment, consider the following instructions:

1. Open the experiments list

   In the H2O Hydrogen Torch navigation menu, click View experiments.

2. Open the experiment

   In the Experiments table, click the name of the experiment whose tabs you want to view.

Tabs

Charts

The Charts tab of an experiment displays four graphs that highlight performance metrics during the experiment. The graphs are as follows:

• Learning rate
• Train batch loss
• Validation batch loss
• Validation {Metric}

Learning rate

The Learning rate graph displays the current learning rate throughout the training process visualized after each training batch.

Train batch loss

The Train batch loss graph displays the current training loss throughout the training process visualized after each training batch.

Validation batch loss

The Validation batch loss graph displays the current validation loss throughout the training process. Loss is calculated as the average of all validation batches. By default, it is calculated and visualized after each training epoch, but you can adjust this frequency with the Evaluation epochs setting when creating an experiment.

Validation {Metric}

The Validation {Metric} graph displays the current validation metric throughout the training process. In this case, the {Metric} refers to the selected metric for the problem type (for example, MAE for an image regression experiment). {Metric} is calculated on full validation set predictions. By default, it is calculated and visualized after each training epoch, but you can adjust this frequency with the Evaluation epochs setting when creating an experiment.

Summary

On the Summary tab, H2O Hydrogen Torch displays the following information about the experiment.

Name	Description
Parent Experiment	The name and fold of the experiment that this experiment was created from. This row appears only when the experiment was created from another experiment.
Name	Name of the experiment.
Dataset	Dataset used for the experiment.
Problem type	Problem type of the experiment.
Fold	The validation fold value.
Seed	The random seed value that you selected for model training.
Gpu list	List of GPUs that H2O Hydrogen Torch used during model training.
Loss	The loss function that model training directly optimizes.
Metric	The metric selected to evaluate model performance.
Train metric	The evaluation metric computed on the training data. Available only when the Calculate train metric setting is turned on in the experiment settings.
Val metric	The evaluation metric computed on the validation data.
Test metric	The evaluation metric computed on the test data. Available only when the experiment has test data.
Batch inference speed	The average wall-clock time per batch across the experiment's validation and test passes, including prediction postprocessing.

Metrics

On the Metrics tab, H2O Hydrogen Torch displays the following experiment metrics:

• An array of scorer values (metrics) applicable to the problem type of the experiment (such as Accuracy, MAE, MSE, or MAPE), where the scorer values refer to the validation dataset
  • If you turned on the Calculate train metric setting for the experiment, H2O Hydrogen Torch also provides scorer values for the training dataset
  • If you provided a test dataset, H2O Hydrogen Torch also provides scorer values for the test dataset
• Best thresholds and their corresponding metric values are shown for threshold-dependent metrics
• For classification experiments, H2O Hydrogen Torch displays confusion matrices where an interactive threshold slider is available in the case of binary or multi-label classification. Moving the slider changes the metric values for threshold-dependent metrics along with the confusion matrices

note

Precision, recall, F1, F05, and F2 are micro-average. To learn more, see Metrics.

Train data insights

H2O Hydrogen Torch displays random visual samples of the training data on the Train data insights tab after the experiment's augmentations.

Note

• For text models, you can view text samples tokenized and in their original form. For example, observing a tokenized text allows you to see the beginning ([CLS]) and end ([SEP]) of a sample text + its label
  • The [CLS] token representation depends on the tokenizer model backbone. For example, it may look like <s> for some models
  • The [SEP] token may not be present for some models
• The Train data insights tab is available for the following problem types:
  • Image regression
  • 3D image regression
  • Image classification
  • 3D image classification
  • Image object detection
  • Image semantic segmentation
  • 3D image semantic segmentation
  • Image instance segmentation
  • Image metric learning
  • Image and text classification
  • Text regression
  • Text classification
  • Text sequence to sequence
  • Text metric learning
  • Audio regression
  • Audio classification

Validation prediction insights

On the Validation Prediction Insights tab, H2O Hydrogen Torch visualizes the validation samples of the experiment while providing the loss value of each sample. In particular, insights are divided as follows:

• Random validation samples
  • In the Random validation samples tab, H2O Hydrogen Torch visualizes random validation samples of an experiment while providing the loss value of each sample.
• Best validation samples
  • In the Best validation samples tab, H2O Hydrogen Torch visualizes the best validation samples of an experiment while providing the loss value of each sample. Samples in this tab are regarded as the model's smallest significant errors.
• Worst validation samples
  • In the Worst validation samples tab, H2O Hydrogen Torch visualizes the worst validation samples of an experiment while providing the loss value of each sample. Samples in this tab are regarded as the model's most significant errors. Samples in this tab can help identify issues the model might have.

Note

Evaluating the best and the worst performance differs for each problem type and might vary from the selected loss and the selected metric to assess the model.

Validation interpretation insights

The Validation interpretation insights tab is only available for image, text, and audio classification and regression experiments.

• Image
  • For an image regression and classification experiment, on the Validation interpretation insights tab, H2O Hydrogen Torch 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.
• Text
  • 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.
    • Highlighted green text: The greener the highlight, the higher the positive contribution to the prediction the text had
    • Highlighted red text: The redder the highlight, the higher the negative contribution to the prediction the text had
• Audio
  • For an audio regression and classification experiment, on the Validation Interpretation Insights tab, H2O Hydrogen Torch highlights with Gradient-weighted Class Activation Mapping (Grad-CAM) the areas of a spectrogram the model considered the most when generating a prediction for the image.
  Note

  H2O Hydrogen Torch only presents the first five seconds of an audio (spectrogram) if the audio is longer than five seconds.

Insights are divided as follows:

• Random validation samples
  • In the Random validation samples tab, H2O Hydrogen Torch visualizes random samples.
• Best validation samples
  • In the Best validation samples tab, H2O Hydrogen Torch visualizes the best validation samples. Samples in this tab are regarded as the model's smallest significant errors.
• Worst validation samples
  • In the Worst validation samples tab, H2O Hydrogen Torch visualizes the worst validation samples. Samples in this tab are regarded as the model's most significant errors. Samples in this tab can help identify issues the model might have.

Note

For image classification experiments, in the Random validation samples, Best validation samples, and Worst validation samples 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.

Logs

On the Logs tab, you can view the experiment's logs.

Config

On the Config tab, you can view the experiment's configurations (setting values) used during training.

MLOps

On the MLOps tab, you can upload the built model (experiment) to H2O MLOps. Deployments are created and managed in the H2O MLOps UI. The MLOps tab displays the following cards:

• Upload model to H2O MLOps
• Uploaded models

Note

The MLOps tab is available for a training experiment with a Finished or Uploaded status when H2O MLOps is turned on for your H2O Hydrogen Torch instance. The tab is not available for demo experiments or for predictions.

Upload model to H2O MLOps

In the Upload model to H2O MLOps card, you can select an H2O MLOps workspace and upload the built model to it.

• To learn how to upload a built model to H2O MLOps, see Deploy a model to H2O MLOps through the H2O Hydrogen Torch UI.

Uploaded models

In the Uploaded models card, H2O Hydrogen Torch tracks the model artifacts that it uploaded to H2O MLOps for the model (experiment). For each upload, the card lists the registered model name, the H2O MLOps workspace name and UID, the model UID, and the date and time of the upload. The model UID identifies the registered model in H2O MLOps, so repeated uploads of the same experiment to the same workspace share a model UID and appear as separate rows, one per version.

Note

Deployments of the uploaded models are created, configured, and monitored in the H2O MLOps UI. To learn more, see the H2O MLOps documentation.

---

Feedback

• Submit and view feedback for this page
• Send feedback about H2O Hydrogen Torch to cloud-feedback@h2o.ai