Hey there awesome People!
I am wondering what is the best way to put meta data as extra information into my CNN. I thought of normalizing it and concatenating the tensor of meta data with the output-tensor of my last Conv2D Layer so the Fully connected layers can take advantage of this meta-data. Is this common practice in PyTorch?
Thanks a lot!
What kind of metadata do you have?
For the case that you deal with numerical metadata:
You could surely do this.
Another approach would be to repeat your Metadata until it has the same dimensions as your images and start feeding it into your Conv layers. This would have the benefit, that you do not have a mixture of abstract/learned features and “handcrafted” features/Metadata. Your network would learn to build abstract features out of your metadata as well.
However, your approach could work too.
In my opinion it is something like a philosophical decision.
It’s numerical data, output of another Net.