I am training the Densenet121 Model on an image dataset. I divided the dataset into 80% for training and 20% for testing. Then I further divided the training data into 85% for training and 15% for validation. Now, after every epoch, I find the accuracy of the model. I save the model after the epoch in which it gives the highest accuracy on validation dataset. Then I evaluate this model on the completely unseen test dataset. Is this the right approach?
For example, I train the model for 10 epochs. After each epoch, the model gives following validation accuracies:
Epoch 1: 0.92
Epoch 2: 0.925
Epoch 3: 0.940
Epoch 4: 0.943
Epoch 5: 0.96
Epoch 6: 0.95
Epoch 7: 0.955
Epoch 8: 0.958
Epoch 8: 0.95
Epoch 9: 0.95
Epoch 10: 0.94
Since the accuracy is highest after 5th epoch, I select the model after the 5th epoch, and then test it on the unseen test dataset.
I am suspecting that because I am selectively choosing the model with highest validation accuracy, I might be introducing bias in my model.
