Versions

Choosing a model to analyze and examine in Tensorleap and managing uploaded models

The Versions view is where you track all the saved versions and models of your project. From here, you can select which model to examine in the Network Tab, add a model run (i.e. a specific evaluation) for analysis in the dashboard, delete, export, and upload new models via the Platform.

Versions Layout Overview

Click on the top left to open the Versions view. Once open, you can choose to fix the view by clicking .

The version view pinned using the top right pinning

The Versions view lists all the models you have in your project. In the Top of the version control you have the Project Name, and below it, the name of the currently selected model that is viewed within the Network Tab. The selected model is also highlighted in a different background color than the other models.

Many evaluation of the same model could be made with different code integration script & versions. Each of these would result in a model run, that could be viewed within the version control by expanding a Model version. hovering over a Run would provide extra details: when did it started, and what script version did it use for the evaluation.

Models that were not yet evaluated would not show any available runs

An Example of an expanded Model version with its Runs with the mouse hovering the oldest Run.

Switching & modifying Model versions

To review and modify different models that exists in the platform - simply hover the mouse over the required model, and click "Open Commit"

Switching Models in the Network view

Switching Reviewed models in the Network Tab has no effect on the Dashboard analysis

Each model in the platform is associated with a different code integration version. Switching between models allow you to review and change the current code integration that is associated with each model.

Deleting and renaming Models and Model Runs

To delete a model or a model run: hover over it, and click the trash can icon.

To rename a model or a model run: click over the name of the requested version and enter a new name.

Adding a model run to dashboard analysis

Once an evaluate is completed, multiple Model Runs could be added to the dashboard panel for analysis purposes. To do so, click the "Add To Dashboard" Icon that is to right of each Model Run.

If the "Add To Dashboard" Icon is not selectable next to a Model Run it indicates that the evaluate didn't start yet or it was finished unsuccessfully, and there's no data to analyze.

Saving a Model Version

In order to keep changes to the model mapping's or to the selected code integration, the model version needs to be saved. Saving a model can be done either in-place (overwriting current model configurations) or to a new model.

Saving a model version after changes
  • To save a current model we need to click on the top disk icon. This would ensure any future evaluation of this model would use the current configuration.

  • To save to a new model we can click the "disk with pencil" icon to the bottom. This would require the user to provide a new name for the experiment and then "Save it as a new version".

Export & Upload a model from Tensorleap

Exporting a model

To export a trained model out of Tensorleap:

  1. With the Versions view open, search for the model

  2. Hover your mouse on the model, then click on the right to open the Export Model window.

  3. On the Export Model window, select the format in which the model will be saved.

  4. Click to start the export process.

  5. The job appears on the list to the right with status set to Pending. A notification message also appears briefly on your screen.

6. Once Tensorleap completes compiling the file, status is set to Finished.

If status is still set to Pending after some time, you may need to refresh the page to see the change in status to Finished.

7. Click to save the file to your computer.

Downloading an exported model

The available export formats are Json (Tensorflow 2), H5 (Tensorflow 2), ONNX , SavedModel (Tensorflow)

Importing a model via the Platform

Uploading a model to the platform is done via the top most cloud icon in the version panel

To import a model from the platform, click the top most cloud icon in the version panel. This would allow you to provide a name for the Model, select its type, select it from a location on your disk and upload it to the Tensorleap system.

The model upload panel

For the ONNX model platform the option to "Transform Inputs" is available but is default off. The purpose of this flag is to switch the model between a channel-first and a channel-last architecture. Note that if you enable this flag the model would expect a different order of the input tensor than the one tested in your local integration test.

The model upload panel allows to indicate which dataset should be connected with the model after upload. Providing this context would allows automatic mapping to be applied in case it already exists.

Last updated

Was this helpful?