Problems
Loading...
Step 5 of 6
VGG Feature Extractor
Medium

Implement a configuration-driven VGG feature extractor in NHWC layout. Each integer entry applies the next supplied same-padded convolution and bias followed by ReLU. Each M entry applies non-overlapping 2 by 2 max pooling. Return the final features as a float64 NumPy array.

Loading visualization...

Examples

Input: x = [[[[1],[2]],[[3],[4]]]], config = [1,"M"], kernels = [[[[[2]]]]], biases = [[-1]]

Output: [[[[7]]]]

Explanation: The integer entry applies a convolution and ReLU before M pools the resulting spatial grid.

Input: x = [[[[1,-1]]]], config = [2,1], kernels = [[[[[1,0],[0.5,1]]]],[[[[0.5],[2]]]]], biases = [[0,0.1],[-0.2]]

Output: [[[[0.05]]]]

Input: x = [[[[1],[-1]],[[2],[0]]]], config = [1,"M"], kernels = [[[[[1]]]]], biases = [[0.5]]

Output: [[[[2.5]]]]

Hint 1

Keep a separate index for kernels because M consumes no weights.

Hint 2

Reuse the same convolution operation for every integer entry.

Hint 3

Apply max pooling immediately when M is encountered.

Requirements

  • Use NumPy.
  • Consume one kernel and bias for each integer entry.
  • Apply same-padded convolution and ReLU for integer entries.
  • Apply 2 by 2 max pooling for M entries.
  • Return a float64 NumPy array.

Constraints

  • x uses NHWC layout and dtype float64.
  • config contains positive integers and M strings.
  • kernels and biases contain one compatible pair per integer entry.
  • Spatial dimensions are even whenever pooling occurs.
  • Every supplied numeric array has dtype float64.
Prev5 / 6Next