Conv2d

Conv2D is a two-dimensional convolutional layer widely used in convolutional neural networks to process grid-like data such as images. It applies a set of learnable filters to an input tensor to produce a stack of output feature maps, one per filter. Each filter has a width and height (the kernel) and depth equal to the number of input channels. As the filter slides over the input with a defined stride, a dot product is computed between the kernel weights and the corresponding input region, producing a single value per position. Repeating this across all spatial positions yields the output feature maps, which are often passed through a nonlinearity.

Key parameters include in_channels, out_channels, kernel_size, stride, padding, dilation, groups, and bias. The number of output channels equals the number of filters. Padding controls the spatial dimensions; common choices are padding that results in smaller outputs (valid or no padding) and padding that preserves dimensions (same). Dilation spaces kernel elements, expanding receptive field without increasing parameters. Groups enables grouped convolutions, including depthwise convolutions when groups equals in_channels, which separate the filtering per input channel before combining.

Output size depends on input height and width, kernel size, padding, stride, and dilation. In general, each spatial dimension is reduced or preserved according to these parameters, while the depth equals out_channels.

In practice, Conv2D is implemented in major frameworks (for example, PyTorch uses Conv2d with in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias; TensorFlow/Keras uses Conv2D with filters, kernel_size, strides, padding, dilation_rate, groups, use_bias). It serves as a core building block for feature extraction in CNNs, typically followed by activation functions and often stacked to learn hierarchical representations.