While debugging regression sample of deeplearning4j I've noticed that it doesn't have normalization of data inputs and outputs. So first of all question, why it doesn't have normalization? And second question, is there somewhere in network architecture normalization mechanism?
As prof of non-normalized input is the following screenshot which was taken right before execution of line
return new ListDataSetIterator(listDs,batchSize);
We actually do normalization. We just don't do it for you automatically. It's right in our examples:
All of our image classifiation examples do this. It's also documented on our website: https://deeplearning4j.org/image-data-pipeline.html We even have videos of this.
Edit: You can also normalize the labels if you want using the same DataNormalization api calling fitLabels(true) before you put data in to the neural network.
If you don't mind could you give me feedback as to how you couldn't find this so we can improve the website? I'm not sure what was missing here.