Train ModelPack on Playing Cards V7 dataset

Train ModelPack on Playing Cards V7 dataset

1. Open eIQ Portal and load the project as before, or go to the Dataset Curator view by selecting it in the WORKSPACES menu.

2. After the dataset is loaded, the GUI will show you the Data Set Curator view. The bottom of the left panel shows the number of images contained in this dataset.  This example contains around 1328 training samples and 148 validation samples. The Data Set Curator can be used to add images, Augment images, select images for training or test.  Refer to the eIQ Portal User Guide for details on using the Data Set Curator. 

3. To start training proceed to the Model Selection page, by clicking on the SELECT MODEL button.  Then select DEEPVIEW MODELPACK ADD-ON. Click CONTINUE WITH FREE TRIAL to open the Trainer setting screen.  Configure the trainer it according to your requirements. To reproduce results shown in this article, apply the following settings:

   • Change input resolution to 416,416,3 (this resolution will provide you higher accuracy while running at > 50 FPS on the NPU)
   • Leave Initialization weights as coco (these weights comes from COCO dataset)
   • Decrease the Learning Rate to be 0.001
   • Enable learning rate schedulers and set it to Linear, set Epochs to 10 and Decay Rate to 0.9
     
   • Set Batch Size to any number your GPU/CPU is able to work with (we recommend 10)
   • Set the Epochs to Train to 150
   • Set Augmentation Setting to Default Augmentation - Oriented (handle some rotations, color transformation and flips). Since cards could come in any orientation, you could select either Default Augmentation (handles rotations with 180 degrees) or Default Augmentation - Oriented.
   • Under Model Parameters, the default values can be used
     • BatchNormalization: trainable
     • Freeze backbone: false
     • Allowed boxes: 150
     • Optimizer: Adam 
   • Enable all in the Evaluation Settings section.  Validation, training and visualization will allow visual progress of the training step by step.
     • Enable Training Metrics
     • Enable Evaluation
     • Epochs Per Evaluation - 1
     • Enable VisualizationsAt this time training can be started by clicking the START TRAINING button.  

3. Initially, eIQ Portal will take few minutes to configure the environment, initialize model and pre-cache dataset. After that, ModelPack will start training, and you should be able to see some progress on the different panels ( metrics, loss and images). Notice that images are only draw at the beginning or each epoch.

   • Training and Validation Accuracy (Mean Average Precision) Notice this training session was stopped at epoch 47 because the model started to produce very good results. In your case, you can leave it to finish the training or just stop it at 25 epochs or even before 25 epochs.
   • Training and Validation Loss
   • Predicted boxes with Images