Visualizing Convolutional Networks: Unpooling With Switches
Advanced
The problem, from first principles
Unpooling with switches puts each activation back at the position where its maximum came from, and leaves zeros elsewhere. The visualizer then runs the network backwards to project a feature onto image pixels.
From theory to code
Implement unpool_with_switches(pooled, switches, size), the inverse of max-pooling on the recorded maxima.
Constraints
- Zeros fill all non-switch positions.
Hints
Hint 1
Allocate zeros of length len(pooled) * size, then write each pooled value at window start plus its switch offset.
Keywords
research-paperscomputer-visioncnninterpretability
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.