TABLE 3.
Parameters for CNNs.
Layer | Parameters (2D-CNN) | Parameters (3D-CNN) |
Conv2D (for 2D-CNN) or Conv3D (for 3D-CNN) | Filters = 64 | Filters = 64 |
Kernel_size = (3,3) | Kernel_size = (3,3,3) | |
Padding = “same” | Padding = “same” | |
Activation = “relu” | Activation = “relu” | |
Input_shape = (5,5,7) | Input_shape = (5,5,5,7) | |
Conv2D (for 2D-CNN) or Conv3D (for 3D-CNN) | Filters = 128 | Filters = 128 |
Kernel_size = (2,2) | Kernel_size = (2,2,3) | |
Padding = “same” | Padding = “same” | |
Activation = “relu” | Activation = “relu” | |
Conv2D (for 2D-CNN) or Conv3D (for 3D-CNN) | Filters = 256 | Filters = 256 |
Kernel_size = (1,1) | Kernel_size = (1,1,5) | |
Padding = “valid” | Padding = “valid” | |
Activation = “relu” | Activation = “relu” | |
Reshape | Target_shape = (5,5,256) | |
Flatten | NA | |
Dense | Units = 256 | |
Activation = “relu” | ||
Dense | Units = 5,400 | |
Reshape | Target_shape = (5,5,256) | |
Conv2D | Filters = 128 | |
Kernel_size = (2,2) | ||
Padding = “same” | ||
Activation = “relu” | ||
Conv2D | Filters = 64 | |
Kernel_size = (3,3) | ||
Padding = “same” | ||
Activation = “relu” | ||
Conv2D | Filters = 1 | |
Kernel_size = (1,1) | ||
Padding = “same” | ||
Activation = “linear” |
Each Keras layer refers to a building block of the neural network, including convolution layers (Conv2D and Conv3D), reshaping layers (Flatten and Reshape), and fully connected layers (Dense). Besides the last layer, all layers used a rectified linear unit (“relu”) activation function that directly outputs the input, if positive, or zero otherwise.