Adaptive Max Pool 2D

Adapctive_Max_Pool2d functions the same as Max_Pool2d but rather than defining the kernel, stride, and padding, the desired otuput dimensions are defined and the rest is calculated automatically.

When defining the output dimensions, it's important to understand what the resulting stride and kernel will be. Kernel and Stride, per axis, are calculated as follows:

\[ Stride = \left\lfloor \frac{input\ size} {output\ size} \right\rfloor \]

\[ Kernel = input\ size - (output\ size - 1) \times Stride \]

Bases: Module

Scales down an image by selecting the maximum pixel given a desired output shape

Parameters:

Name	Type	Description	Default
height	int	height of the output image	required
width	int	width of the output image	required

Shapes

Input:
    (N, channels, in_height, in_width)

Output
    (N, channels, out_height, out_width)

Implementation

import neuraltoolkit as ntk
import numpy as np

sample_count = 100
channels = 3
height = 28
width = 28

x = ntk.Tensor(
    np.random.rand(sample_count, channels, height, width)
)

# Stride = (2, 2)
# Kernel = (4, 4)

layer = ntk.Max_Pool2d(
    height=13
    width=13
)

y = layer(x)