Skip to content

Model Experiments Dashboard

A model experiment is a container of training and validation sessions in the experiment. The following figure shown is the "Model Experiments" page. This page will contain all the experiments that were started by the user.

Project Experiments
Project Experiments

The following figure breaks down the elements of an "Experiment" card.

Experiment Attributes
Experiment Attributes

An experiment will contain child training and validation sessions. The training sessions and validation sessions will be described in more detail in the sections below.

Training Sessions

Training sessions take in datasets and synthesize from them new AI models for object recognition (Vision) or object perception (Fusion).

From the "Model Experiments" page, we can click on the "Training Sessions" button with the icon Trainers Button to see the training sessions in the experiment. The figure below shows the layout of the training session cards under the "Training Sessions" page.

Training Sessions
Training Sessions

The following figure describes the attributes of any given training session.

Training Session Attributes
Training Session Attributes

For more details regarding deploying training sessions, please see Training Vision Models and Training Fusion Models for training Fusion models.

Training Session Details

Click on any training session card to open its details page. The details page provides a deep view into the session through the following tabs:

Training Session Details
Training Session Details
Tab Description
Session Summary of the session configuration: session ID, name, description, created by, status, dataset, annotation set, train/val groups, start date, and duration. Also shows the training parameters in YAML format and the status of each trainer process stage.
Logs Live and historical console output from the training process.
Charts Training metrics plotted over epochs, such as loss curves and mAP.
Artifacts Downloadable model artifacts produced by the session (ONNX, TFLite, etc.). Click the download arrow next to any artifact to save it locally.
Runs Individual run attempts within the session, useful when a session was restarted.

The action bar at the top of the details page provides the following operations:

  • Refresh — Reload the session state and latest log output.
  • Validate — Create a new validation session for this training session.
  • Compare — Compare the training charts of this session against others.
  • Stop — Terminate a running training session.
  • Recycle Bin — Move the session to the recycle bin.

Cloning a Training Session

Cloning creates a new training session from an existing one. To clone a session, click "clone training session" button on the training card. A dialog will appear pre-filled with the source session's configuration. Adjust the name and description as needed, then confirm.

The clone runs the source session's trainer with its configuration. To reuse the trained weights of an Ultralytics session instead, start a new Ultralytics session with its Weights set to that session:

  • With Enable Training on, the new session fine-tunes those weights, for example on a different dataset.
  • With Enable Training off, the new session exports those weights without retraining, for example at another Input Resolution or Deployment.

See Weights and Enable Training for the full behavior matrix.

Validation Sessions

The validation sessions will assess the performance of the models in the training sessions. A training session can have any number of validation sessions. From the "Model Experiments" page, we can click on the "Validation Sessions" button with the icon Validation Button to see the validation sessions in the experiment. The figure below shows the layout of the validation session cards under the "Validation Sessions" page.

Validation Sessions
Validation Sessions

The following figure describes the attributes of any given validation session.

Validation Session Attributes
Validation Session Attributes

To compare the validation charts of each session, select the validation sessions to compare by using the checkbox on the top left of each session card and then click on "Compare" at the top right corner of the "Validate Sessions" page. This will show the charts of each validation session side-by-side for a quick comparison.

Compare Validation Sessions
Compare Validation Sessions

Now the charts for each selected session are displayed side-by-side. All the charts for a single training session will be shown in one column. A new column indicates another session.

Comparing Validation Sessions
Comparing Validation Sessions

There are two types of validation: cloud validation and on-target profiling. Cloud validations are set by default in EdgeFirst Studio which triggers an EC2 instance that downloads the dataset and the inference model for validation. This type of validation is best used if you do not have an embedded platform to deploy the model. On-target profiling are hosted on embedded platforms where the dataset and the inference model will be downloaded. This type of validation is best used if you have an embedded platform available to verify how the model would perform on the platform and determine the model's inference time when deployed.

For more details regarding deploying validation sessions, please see Validating Vision Models and Validating Fusion Models.

Validation Session Details

Click on any validation session card to open its details page. The details page provides the following tabs:

Validation Session Details
Validation Session Details
Tab Description
Session Summary of the session: session ID, status, name, model used, linked training session, dataset, start date, and duration. Also shows the validation parameters in YAML format.
Logs Console output from the validation run.
Charts Validation metric charts such as precision-recall curves.
Metrics Tabular accuracy metrics (mAP, AP per class, etc.) from the completed validation run.
Runs Individual run attempts for this session.
Profile Loads the trace files measuring the performance of the model on target

The action bar at the top of the details page provides the following operations:

  • Refresh — Reload the session state and latest output.
  • Stop — Terminate a running validation session.
  • Recycle Bin — Move the session to the recycle bin.

Running a model

EdgeFirst Studio allows you to run trained models directly in the web browser. After a model has been trained, open the model details page by clicking the training session card. Only ONNX models are currently supported for live inference in the browser. On the Artifacts tab, click the Run Model button (the play icon) next to the ONNX artifact.

ONNX models only

Live browser-based model running only supports ONNX format. To validate or profile models in other formats (TFLite, Neutron, Hailo, TensorRT, etc.) on real hardware, use the EdgeFirst Profiler.

Training Session Details (Run Model button)
Training Session Details (Run Model button)

Please wait while the model loads. This may take up to 60 seconds, depending on the model size. Once the model is loaded, select a mode: Live to run the model on images from a camera stream, or Upload to choose a file from local storage.

You should now see the live inference feed on your browser running the trained model.

Live Inference
Live Inference

The model will run automatically, and the results will be displayed on the screen. The output (segmentation or detection) depends on the type of model being used.

Image Inference
Image Inference

For more information, please see the full tutorial on model deployment.

Next Steps

Now that you are familiar with the layout of the Model Experiments Dashboard, proceed to the next section for learning more about the context of the Tasks Dashboard.