RFC-0039-incremental-pca.md

RFC-0039-incremental-pca.md

@@ -0,0 +1,91 @@
+# [Implementation of Incremental PCA]
+
+**Authors:**
+* @sirluk
+
+
+## **Summary**
+Implement a class for Incremental PCA in vanilla PyTorch with native GPU support.
+
+
+## **Motivation**
+- Incremental PCA is important where data arrives in streams or is too large to fit in memory.
+- PyTorch currently lacks a built-in implementation of incremental PCA, and the best available implementations, like `sklearn`, do not support have GPU support.
+- Requested in issue tracker: [issue40770](https://github.com/pytorch/pytorch/issues/40770)
+
+
+## **Proposed Implementation**
+The proposed implementation would be similar to the one in `sklearn` ([sklearn.decomposition.IncrementalPCA](https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.IncrementalPCA.html)) but with the following differences:
+- torch.tensor objects instead of numpy arrays
+- GPU support
+- Option to use truncated SVD via torch.svd_lowrank
+
+Draft implementation can be found [here](https://github.com/sirluk/pytorch_incremental_pca). Potentially it could make sense to inherit from torch.nn.Module and define internal state as buffers.
+
+
+## **Metrics **
+- Ability to handle datasets larger than available RAM
+- GPU vs. CPU performance comparison
+- Accuracy comparison with sklearn's IncrementalPCA
+
+## **Drawbacks**
+- Implementation could be relatively self-contained, no additional dependencies required. Can reuse existing linear algebra functions in pytorch. Draft implementation already created.
+- Tests required to ensure numerical stability accross various data types and sizes.
+
+
+## **Alternatives**
+What other designs have been considered?
+- Use scikit-learn's Incremental PCA by moving data to CPU. Very slow.
+What is the impact of not doing this?
+- Users might resort to less efficient workarounds or third-party libraries
+- Potential migration of users to other tools for specific PCA-heavy workflows
+
+## **Prior Art**
+- Implemented in sklearn: [sklearn.decomposition.IncrementalPCA](https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.IncrementalPCA.html)
+- Large parts of the sklearn implementation can be reused
+
+
+## **How we teach this**
+- "IncrementalPCA" or "IPCA" for the main class name. Clearly indicates the incremental nature of the algorithm
+- Use terms like "partial_fit" for incremental updates.
+- Emphasize use cases: streaming data, large datasets, online learning
+- Demonstrate GPU speedup with benchmarks
+- Provide code examples demonstrating integration with PyTorch datasets and dataloaders
+
+## **Unresolved questions**
+- Where in the pytorch codebase should this live?
+- What should the API look like?
+- Should this be implemented as a torch.nn.Module or should it be a standalone class?
+- Should the device be implicit (inferred from the input tensor) or explicit?
+- Should this work with multiple GPUs out of the box?
+Out of scope for this RFC:
+- Multi-GPU support
+
+## Resolution
+We decided to do it. X% of the engineering team actively approved of this change.
+
+### Level of Support
+Choose one of the following:
+* 1: Overwhelming positive feedback.
+* 2: Positive feedback.
+* 3: Majority Acceptance, with conflicting Feedback.
+* 4: Acceptance, with Little Feedback.
+* 5: Unclear Resolution.
+* 6: RFC Rejected.
+* 7: RFC Rejected, with Conflicting Feedback.
+
+
+#### Additional Context
+Some people were in favor of it, but some people didn’t want it for project X.
+
+
+### Next Steps
+Will implement it. 
+
+
+#### Tracking issue
+<github issue URL>
+
+
+#### Exceptions
+Not implementing on project X now. Will revisit the decision in 1 year.