Squeeze-and-Excitation: Recalibrating the Features
Beginner
The problem, from first principles
After the gates are computed, the block multiplies each channel of the original feature map by its gate. The gate is the same at every spatial position, so the whole channel is scaled.
From theory to code
Implement recalibrate(x, s), scaling each channel of the feature map by its gate.
Constraints
- The gates broadcast over height and width.
Hints
Hint 1
Multiply the map by the gate vector, which NumPy broadcasts over the last axis.
Keywords
research-paperscomputer-visionattentionarchitecture
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.