Conv2D¶

Conv2d is the standard layer for image processing within a neural network

This layer performs a convolution by sliding a filter (The kernel) across the image. A new, scaled down image is output, calculated by the weighted sum of the pixels under the kernel at each position as well as adding a bias term.

The out_channels parameter determins the number of learned features the model will extract from the image.

Output Dimensions¶

The dimensions of the output image is determined by the values for kernel_size, stride, and padding The formula for each dimension is the following

\[ Output\ Size = \left \lfloor \frac{input\ size + 2\times padding - kernel\ size}{stride} \right \rfloor + 1 \]

Bases: Module

Standard convolutional layer

Applies a kernel to an image output a scaled down image of feature channels

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Number of channels in the input image	required
`out_channels`	`int`	Number of channels in the output image	required
`kernel_size`	`int or tuple`	Shape of kernel (h, w). If integer then defaults to (a, a)	required
`stride`	`int or tuple`	Number of pixels to skip in each direction (h, w). If integer then defaults to (a, a)	required
`padding`	`int or tuple`	number of zeros to apply to the input image vertically and horizontally (h, w). If integer then defaults to (a, a)	required

Shapes¶

Input:
    (N, in_channels, in_height, in_width)

Output:
    (N, out_channels, out_height, out_width)

Example

layer = ntk.Conv2d(3, 32, 2, 1, 0)

y = layer(x)

implementation¶

import neuraltoolkit as ntk
import numpy as np

image_count = 100
in_channels = 3
height = 28
width = 28

x = ntk.Tensor(
    np.random.rand(image_count, in_channels, height, width)
)

layer = ntk.Conv2d(
    in_channels=in_channels,
    out_channels=32,
    kernel_size=2,
    stride=1,
    padding=0
)

y = layer(x)