Why do my val_loss fluctuate and have enormous values while val_categorical_accuracy are more or less constant throughout all epochs?

Question

I adapted the AlexNet architecture, preprocessed the images, augmented the images, used LeakyReLU activation function, utilized dropouts, and had tried adjusting a learning rate. However, these trials aren't improving my model's val_loss and val_categorical_accuracy. What should i do? Embedded are my model's fitting and history visualized =

my model.compile()

my model's fitting

Training and validation loss graph

Answer 1

This might occurs due to several reasons, the most common reasons,

---

• The number of examples of the validation set are not big enough

• A validation step at each epoch might be small

• Validation set has examples not in the training set

• Training set batch size not quite good compared with validation set size

• Model biased against (high bias) some classes at training set and those classes more likely to be in the validation set.

• ... etc

---

You could check n - randomly chosen mislabeled examples from the validation set, usually is the best thing to do so.

Answer 2

Turns out i tweaked my model too much. I tried using purely AlexNet with ReLU function and it did the job, albeit too slow (175 s/epoch). I also didn't know that to have a layer (both Conv and FC layers) use an activation function, we must specify the function within the layer's parameter (before, i had used the .add after said layer and turns out it didn't register as the activation function of said layer therefore the layer didn't learn anything throughout the fitting process). I really need to learn more about building CNNs i guess.