So far, we have seen how to prepare the training set; now we have a challenge ahead. We have to train 234,545 images. Although the testing phase would be less exhaustive with only 500 images, it's better to train each CNN using batch-mode using DL4j's MultipleEpochsIterator. Here's a list of important hyperparameters and their details:
- Layers: As we have already observed with our simple 5 layers MNIST, we received outstanding classification accuracy, which is very promising. Here I will try to construct a similar network having.
- The number of samples: If you're training all the images, it will take too long. If you train using the CPU, not the GPU, it will take days. When I tried with 50,000 ...