New model Fourier Reparametrization#6
Conversation
completed Fourier Reparametrization module for SIREN and ReLU
…ep models/nonlins/layers separate. Added tests. Removed unused imports.
|
Hi @ImbriacoR , |
|
Hi @kushalvyas, |
daniel360kim
left a comment
daniel360kim
left a comment
There was a problem hiding this comment.
Hi @ImbriacoR,
Overall, this looks like a strong implementation of the FR-INR paper with solid test coverage.
I left a few suggestions around the InitType and NonlinClass abstractions to improve clarity and robustness. See inline comments for details.
- Removed "linear" folder and moved FourierLayer into fourier_reparametrized.py - Simplified model creation. - Added additional test for nonlinearity error.
|
Hi @daniel360kim, I have gone over your review and applied all changes. Please let me know if any other changes are necessary. |
3 participants
Overview
This PR adds the model described in the paper "Improved Implicit Neural Representation with Fourier Reparameterized Training" by Shi et al. The current implementation allows construction of models using either the ReLU or SIREN nonlinearities.
Changes
Added a new Linear layer class for Fourier-based initialization.
Added tests and example notebook.