Visualizing Convolutional Networks: The Receptive Field
Intermediate
The problem, from first principles
Zeiler and Fergus (2014) visualized which input pixels drive each unit. The receptive field, the region of the input a unit depends on, depends on every layer's kernel and on the strides before it.
From theory to code
Implement receptive_field(kernels, strides), returning the receptive field of the last layer's output.
Constraints
- Strides multiply the spacing of later units.
Hints
Hint 1
Start with a field of 1 and jump of 1. For each layer add (k - 1) * jump, then multiply the jump by the stride.
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.