View a narrated version of the tutorial:
The steps for fine-tuning a net overlap with those involved in training a net. Therefore, follow the steps as indicated in Tutorial-3, but only up to Step 12.
Click on the Train View on the main screen of
Expresso. Then, click on the button Train New/Finetune Model.
From the list of net configurations in the listbox at the bottom left of the GUI titled Configuration Data, select the net
CIFAR10_TRAINED. Having done so, click on Next button at the bottom right of the GUI.
Upload a file containing solver parameters by clicking on Upload/Change button on bottom right corner of screen and navigating to the path
$EXPRESSO_ROOT/tutorials/data/mean.npy. The uploaded parameters can be viewed in current screen. Having done so, click on Next button at the bottom right of the GUI.
Select the training data,
cifar10_train, from the Training blob list and enter batch size as
50. NOTE-1:The batch size refers to the number of samples processed together in one iteration of fine-tuning.
NOTE-2: The training and validation data we use for fine-tuning is the same as the one used to train the original net. But this need not be the case in general. . Check Perform validation checkbox. This will show Validation blob list. Select
cifar10_test from the Validation blob list. The batch size refers to the number of samples processed together in one iteration of fine-tuning.
Enter the name by which you want to save configuration --
CIFAR10_FINETUNED. After entering the configuration name, click Finish. On the prompt that pops up, click on Yes to start fine-tuning.
Click on notification button in top right corner of screen, The progress bar shows the fine-tuning progress. Click on details button to view loss function vs iteration plot, which is updated as the fine-tuning proceeds. Completion of fine-tuning is indicated by a notification near the cup icon at top right corner of screen.
Click on Net View. The newly created, fine-tuned net can be observed in the Net Configuration List.
You can view the physical location where the fine-tuned net’s model is stored by right-clicking on net named
CIFAR10_FINETUNED and selecting the option Properties.