Backbone of Computer Vision

CNNs are a type of neural network made for images. They work by scanning small parts of an image, not the whole thing at once.

In this page I will be going through the basic intuition of CNNs, the forward pass, and how convolution + pooling work together.

— Alexi

Convolution → RELU → Pooling

Source: https://medium.com/hackernoon/visualizing-parts-of-convolutional-neural-networks-using-keras-and-cats-5cc01b214e59

Types of Convolution Layer:

1D Convolution Layer:

2D Convolution Layer:

Code and Explaination for Conv2D:

nn.Conv2d(
				in_channels=1, 
				out_channels=32, 
				kernel_size=3, 
				stride=1, 
				padding=1
				)

• Parameters in code:
  • in_channels: Number of input channels (1 for grayscale, 3 for RGB)
  • out_channels: Number of filters/feature maps to create
  • kernel_size: Size of the filter (3×3 in our example)
  • stride: How many pixels to move the filter each time (1 = move 1 pixel)
  • padding: Adding extra pixels around the border (1 = add 1 pixel border)