I trained my first Conv2D-MaxPooling based binary image classification model and observed different regularization methods impact on accuracy/loss - val-acc/val-loss graphics. Trained for thumbs-up / thumbs-down image classification. Top predict accuracy performance with 200x480x640x3 dataset : about %90 But this model lacks consistency because most of it is based on randomness and shuffles.
- Dropout
- Data Augmentation
- Weight Regularization
I couldn't include my dataset because all photos included my face. (shy person intensifies)