We did not plan the model from scratch, but created the model through transfer learning from a previously trained model.

We use Mobilenetv2.

retrain conditions are 
Training steps: 4000
Learning rate: 0.01
Testing percentage: 10% of Dataset
Validation percentage: 10% of Dataset
Train batch size: 100
Test batch size: entire test set
Validation batch size: 100

As a result of retraining with MobilenetV2 with 2800 datasets, the third photo showed the highest jump shot probability with 95%.