...

/

Constructing an Optimized Baseline Convolutional Network

Constructing an Optimized Baseline Convolutional Network

Learn how to build a baseline convolutional network with a strategic swap of pooling and activation.

Baseline network

Construct a simple sequential baseline model with layer structure as shown below. Note to swap activation and pooling layers.

Press + to interact

Convolution layer

  • Conv1D vs. Conv2D vs. Conv3D: A Conv‘x’D is chosen based on the number of spatial axes in the input. Use Conv1D, Conv2D, and Conv3D for inputs with 1, 2, and 3 spatial axes, respectively. Follow the table below for details.

Axes and Channels in Grid-Like Inputs to Convolutional Networks


Time Series

Image

Video

Axis-1 (Spatial dim1)

Time

Height

Height

Axis-2 (Spatial dim2)

-

Width

Width

Axis-3 (Spatial dim3)

-

-

Time

Channels

Features (one in univariate time series)

Colors

Colors

Conv’x’d

Conv1D

Conv2D

Conv3D

Input Shape

(samples, time, features)

(samples, height, width, colors)

(samples, height, width, time, colors)

kernel_size3

An integer, t, specifying the time window

An integer tuple (h, w) specifying the height and width window

An integer tuple,

(h, w, t), specifying the height, width, and time window

Kernel shape

(t, features)

(h, w, colors)

(h, w, t, colors)

  • Kernel: The kernel_size argument is a tuple in which each element represents the size of the kernel along a spatial axis. The size can be taken as:

    spatial axis size2\frac{\sqrt{\text{spatial axis size}}}{2} ...