Model snapshots
Snapshots allow you to create a copy of your model such that if you modify the model, the snapshot is still completely unaffected.
Taking a snapshot
To take a snapshot of a model, go to the model page and click "Take snapshot" on the "Snapshots" panel of the "Configure" tab. This will create a copy of the code in the model as well as the active model state.
Note: You are billed for the total amount of storage used. When snapshotting a model, the duplicated code and state increases this amount. See pricing for details.
In addition to storing the code and state of the model, the snapshot also keeps track of the following settings:
- Filesystem size
- Input/output types
- Max durations
- Package versions
- Default launchers
These settings will be applied for a new model that is created using the snapshot. Even if you change these settings for the model, the snapshot will be unaffected.
Evaluating a snapshot
One common use case is to perform evaluations using the snapshot. You can be sure that your snapshot evaluations keep outputting results as expected even if you were to modify your model. Then, when you ready to use your new changes, create a new snapshot and modify your applications to use this new snapshot for evaluations.
Creating a model based on a snapshot
There is also another common use case for snapshots - you can use a snapshot to create a new model. This can be used to create many models that share the same underlying code. For example, let's say that you want to create one image classifier that is trained on a dataset of vehicles, and one image classifier that is trained on a dataset of dogs. Because both are image classifiers, we should be able to share the same code but train them on different datasets. To do so, first create a generic image classifier, and create a snapshot of this model. Then, create two models that uses the same underlying code by selecting the type "Based on existing" in the create model dialog. Select the image classifier model and snapshot to use as the base. The new models will not contain any code of their own. Instead, they will be linked to the code stored in the snapshot. You can now train and evaluate the two models separately, while only having to write code for a single generic model.
