In this article we explain how to restore a checkpoint previously trained on eIQ Portal. This feature is also explained in the eIQ Portal documentation. Loading a previously trained checkpoint will accelerate your development and testing of ModelPack. From the trained checkpoint you can convert the model, validate the model on a target device and export the model for deployment onto your target device.
Restoring a ModelPack Checkpoint
- To restore a ModelPack checkpoint it is needed to have a previously trained project as well as the project file. For this example, we will use the ModelPack Pre-Trained Checkpoints from the Playing Cards v7 eIQ Portal Project. Download a checkpoint from the downloads section for use in the next steps.
- Once you download the .eiqp file and located it into a folder, it is needed to create an empty folder with the same name of the project but without the .eiqp extension. For example, if the project is named playing_cards_v7.eiqp, the folder should be created in the same directory the project is but with the following name: playing_cards_v7.
- Inside playing_cards_v7 folder it is needed to extract the checkpoint folder. The directory should look like this:
- Open eIQ Portal and load the project playing_cards_v7.eiqp
- Click the SELECT MODEL button
- Click the RESTORE MODEL button
- You will now see all the checkpoints trained for that dataset. If this is the first time you load this project file, only the pretrained checkpoint name will be shown. Notice the name on the screen (step 6) is the same the checkpoint folder has. Select modelpack_playing_cards_v7 checkpoint
- After clicking the checkpoint, it is possible to select each of the epochs stored within the checkpoint. For this case, we only saved the epoch 57. Select that epoch and the graph will be shown in the screen
- At this point, we are ready to validate the model on the target device, export the converted model, and finally deploy the model, all without the need to re-train the model from scratch.