NXP Backend: Quantize selected activation functions with Conv/Matrix multiplication
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 5k
- Forks
- 1.2k
- Avg merge
- 2d 10h
- Merged PRs (30d)
- 581
Description
🚀 The feature, motivation and pitch
Currently the Neutron Quantizer, quantizes the activation functions like Relu separatelly. What is OK for case, when the activation function is isolated. But in majority of the cases the activation function follows the Convolution or matrix multiplication. And Neutron NPU computes the Conv/MatrixMul + Activation in one pass.
Therefore the proper quantization scheme for this case shall be:
DeQuantize -> Conv -> Activation -> Quantize.
Internal ID: EIEX-455
Alternatives
No response
Additional context
No response
RFC (Optional)
No response
cc @digantdesai @JakeStevens
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
No file, test, or entry point is identified. Locate the Neutron Quantizer implementation and the existing Conv or matrix-multiplication activation quantization path first; completion should implement the stated DeQuantize → Conv/MatrixMul → Activation → Quantize scheme and validate it for the selected activation functions.
Written by the indexing model from the issue text.
Assessment
- Domain
- embedded-iot, machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100