Neural Network predicting Fourier Coef for Segmentation - Help for french beginner

vision
1

Hi everybody,

I’m trying to develop a neural network in charge of predicting Fourier’s coefficients of a mask from an image. Then I use this coefficients to reconstruct a mask by convulving a gaussian over predicted points.

My implementation of these functions looks like :

Capture_transform_to_mask
Capture_transform_to_mask788×693 21.9 KB

image
image1446×665 42 KB

My neural network is created using this function :

Capture_nn

All these functions seems to work correctly

Next, I train my model using this function :

Capture_train
Capture_train1439×647 35.6 KB

and these parameters :

Capture_train2
Capture_train21389×269 11.2 KB

But the problem is, it doesn’t work :cry:
The loss doesn’t decrease, and I didn’t find what.
I’ve tried with different epoch, batch, LR and nothing solves the problem

Can someone have an idea ?

Thank you for all,
A beginner with love

2

I cannot read the code on my phone since you’ve posted it via screenshots instead of formatted code snippets.
In any case, you could try to overfit a small dataset (e.g. just 10 samples) first to make sure your model architecture is capable of doing so and no other errors are in your training code.

3

Hi ptrblck,

thanks for your answer and sorry for format’s images I’ve posted.
I’ve already tried to overfit it, and it doesn’t work.
So I tried with different implementation and formula … and it works ! I think problem in the code I posted was in relation with the device, which was not the same for all variables I used