How to feed the output of a neural network into another neural network?

1

Hello,
this question might sound strange but I couldn’t find information about my use case.
My goal is to combine two different neural networks. The first one is running, it gets an image and classifies it. Now I want to feed this output into another network in order to classify a subcategory of this image.
Concrete this means that the first net can classify a buch of flowers and trees, e.g Rose or Palm Tree. The next net should now classify any kind of tree as “tree” and any kind of flower as "flower.

I have got the labels and the output-tensors but the problem is, that these tensors are no images, so they have only 3 dimensions, since they are missing size parameters.
Now I don’t know if there is a more generic net structure to provide just the training on tensors without the image dimensions because I think that convolutional networks are not the right approach.

I wrote a data loader to provide inputs and lables in the following structure:

24
241252×692 55.2 KB

Those are seven output tensors, that should be represented in two different classes.

In this case, all of the input tensors are flowers, so the labels are structured as flowers:

33

I defined flowers as 0 and trees as 1.
If I try to feed this input into a pre-trained model, it is missing a dimension, my own networks are also failing and I wonder where to start.

Thank you very much in advance.

2

If I understand the use case correctly, you would like to get the general classes based on sub-categories from another model.
If that’s the case, wouldn’t a simple mapping just work, i.e. check for the predicted class and apply a mapping to get the more general class (tree, flower)?

I can imagine the opposite use case might work with a staged model, but I’m not sure if a simple if-condition couldn’t work “perfectly” here.

Let me know, if I misunderstood your use case.

3

Thank you for your answer.

Yes that is exactly what I want but I would like to implement it as a neural network. At this stadium it would not make much sense, since it will be a prototype for future use. Later I want to do multiple classifications on different levels, so i thought of a simple task to get started with the problem. And that is why I try to generalize subclasses.

Is there any way to solve it at this way?