I'm currently trying to create a CNN but am having issues figuring out what parameters to be using. Choosing a filter size, number of filters, and number of convolutional layers just isn't clicking with me. All the resources I have found or been given just say to 'look at your dataset' and determine them, which just isn't helpful.
I have a dataset with around 40000 instances and 500 classes. I'm using keras on top of Tensorflow-gpu. Currently I'm using 6 1D convolutional layers with tanh activation and using relu and softmax for my 2 dense layers. I do max pooling after every 2 conv layers. I have two dropouts (between my dense layers and after) with set to .4 each. With my current filter sizes and numbers, I'm getting around 35% accuracy but am expecting close to 80% based on current research.
I'm not necessarily looking for someone to tell me exactly what numbers or configurations I should be plugging into my model, but just some guidance on really how to even begin determining such values. I've really just guess and checked thus far. Additionally, I'm unsure about how the values should relate to each other: should I be doubling the number of filters each layer? How often should I repeat a filter size? etc.
I've looked at other questions on stack overflow and all the ones with answers seem to tell people to try simpler models or to base their work off of keras examples, which is not something that would work for my case.
