ml-mdm
ml-mdm copied to clipboard
ResNet MLX
Key Architectural Changes
-
Unsqueeze Process
Adjusted tensor dimension expansion logic to ensure proper shape compatibility between temporal embeddings and spatial feature maps. -
Temporal Embedding Broadcast
Modified broadcasting logic to align temporal embeddings with spatial dimensions without introducing shape mismatches.
Testing & Validation
-
Test Strategy
Implemented straightforward test cases to verify:- Layer behavior consistency
- Numerical stability across resolutions
-
Resolved Issues
- Fixed GroupNorm portability issues related to:
- Parameter initialization
- Axis/dimension handling during inference
- Validated edge cases for temporal embedding integration
- Fixed GroupNorm portability issues related to: