Assignment 4

For this lab, we recommend that students read the material in the references section of this page. This assignment aims to show the use of a SIREN in practice. In previous assignments, we worked with ReLU and LeakyReLU activation functions. In this assignment, we will work with an activation function in sinusoidal form. Click here to open a simplified example of SIREN and follow the suggestions subsection instructions.

Suggestions:

1. Create a copy of the notebook.
2. Run the cells; we recommend running them one by one to get familiar with the process.
3. The authors in the paper call attention to the adjustment of $\Omega_0$ initialization. Into the notebook are suggestions for specific sections of the paper; we recommend the students read these parts and try to adjust the $\Omega_0$ value. Afterward, the students can compare their results with the provided example and show us.
4. Do a network initialization, but remove the $\Omega_0$ in the hidden layers. Then, write about the network behavior.
5. To verify the importance of normalization, we suggest removing the init_weights function in the code and initializing the network to an arbitrary value. Then, write about the network behavior.
6. Write a review saying how we could merge gradient-domain images using SIRENs. Use the Poisson Image Editing material available in the References section.

Important

The deadline for submitting the assignment is 07/01/2023.

References

Mestrado: Processamento de Imagens - 2022

Papers

SIREN, click here.

Poisson Image Editing, click here.