v0.3.3
Prioritize differentiable rendering
Final API change for DiffDRR
DRR objects now act purely as a differentiable renderer. That means DRR objects take in rotations and translations parameters to specify (batched) camera poses, and render X-rays from those poses.
The pose parameters are not stored as nn.Parameters of the module. This introduced too many annoyances with the DRR module, such as gradients not flowing backwards past the pose parameters, or only being able to render a specific batch size.
2D-to-3D registration is still possible with DiffDRR. This is done by creating a lightweight wrapper module that stores rotations and translations as nn.Parameters, and calls a DRR object in its forward call. An example is available in the updated documentation.
Small changes:
- DRR generation can now be performed patch-wise (i.e., generate a 600^2 DRR as four 300^2 in series). This means
DiffDRRcan generate insanely large DRRs since computation memory is no longer a constraint. See an example in the updated timing notebook.
What's Changed
- Differentiable rendering and large DRRs by @eigenvivek in #139
Full Changelog: v0.3.2...v0.3.3
