Introducing locally connected 1d layers

[Riccardo231]

Starting to create locally connected 1d layers. I was able to create a prototype of this, with the same syntax you used in conv2d layer.

EDIT: I added two new files for generalized reshaping; i still get an error in nf_layer_constructors_submodule. I have to work on this. I think you have to help me here because I am stuck.

This is what i get:

/home/r_orsi/neural-fortran/src/nf/nf_layer_constructors_submodule.f90:164:50:

164 | module function reshape_generalized(output_shape) result(res)
| 1
Error: Rank mismatch in argument 'output_shape' (1/0) at (1)
make[2]: *** [CMakeFiles/neural-fortran.dir/build.make:309: CMakeFiles/neural-fortran.dir/src/nf/nf_layer_constructors_submodule.f90.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:160: CMakeFiles/neural-fortran.dir/all] Error 2
make: *** [Makefile:146: all] Error 2

I know I added a lot of stuff. You don't have to look it all. Just look at the reshape_generalized files.

You have to help me here. I added also a handling of the scalar values in input but i keep getting the same error; i don't understand what it is. If i change the name of the function, I get all well, but then in trying reshape_generalized i get other shape errors. I truly don't understand what is going on.

I wanted to change directly the reshape 3d to a generalized reshape but that function was used way too much and I didn't want to take a bigger step than what I already did.

Still don't look at the implementation of locally connected 1d. I think it is still not what is described in Keras.

[OneAdder]

Wow, looks smart!

Figuring out your issue might not be super easy, given the fact, that everything here is about the shape.
I see that you're using the usual reshape in CNN 1D for MNIST. Is it intentional?
If this is not the problem, then I can only suggest compiling the reshape generalized layer only and testing it for rank issues with ouput_shape...

[Riccardo231]

Hello, it is intentional since I still haven't come to the cnn_mnist_1d part because I was stuck in compiling my reshape_generalized function

[OneAdder]

Hm. The only suggestion I have is to add some tests for the layers. And figure out the issue this way. I usually prefer sort of test-driven approach when working with manual memory allocation.
Hopefully, the issue can get resolved this way. So far I can only see, that for some reason your changes break memory allocation for all layers. Making tests for your layers can help identify where the issue with that lies.

[OneAdder]

I tried compiling your code, but didn't manage to pinpoint the exact culprit. So, I can only suggest a way I would go about figuring it out 🙃

[Riccardo231]

I tried compiling your code, but didn't manage to pinpoint the exact culprit. So, I can only suggest a way I would go about figuring it out 🙃

I patched the bug i was talking about. Now i need to implement the reshape_generalized function into the network but i still haven't written anything

[OneAdder]

The newer version does compile, but tests fail with Fortran runtime error: Dimension 1 of array 'res' has extent 0 instead of <some_random_unallocated_value>

[Riccardo231]

Thank you, I still didn't try the tests

[milancurcic]

Sorry that I only see this now and somehow missed the email notifications (too many).

Thank you for this addition; definitely needed and in scope.

I only need to figure out how to change this PR so that it's against modern-fortran/neural-fortran main, instead of my own fork. Perhaps the easiest approach could be, when this PR is ready, to merge it into milancurcic/neural-fortran main, and then I would open a new PR to merge from milancurcic/neural-fortran main into modern-fortran/neural-fortran main.

[Riccardo231]

Going to apply these modifies today. But I still need to add a lot