Supported Layers
VGSLify supports a range of layers that can be specified using the VGSL format. Each layer type has its own configuration format, allowing you to define models concisely and flexibly. This section provides an overview of the supported layers and their VGSL specifications.
Layer Specifications
Input Layer
VGSL Spec: <batch_size>,<height>,<width>[,<depth>,<channels>]
Description: Defines the input shape for the model, where the first value is the batch size (set to None for variable), followed by the height, width, and optionally the depth and channels.
Example: None,28,28,1
Defines an input layer with variable batch size, height and width of 28, and 1 channel (e.g., for grayscale images).
Conv2D Layer
VGSL Spec: C(s|t|r|l|m),<x>,<y>,[<s_x>,<s_y>,]<d>
Description: Defines a 2D convolutional layer with a kernel size of <x> by <y>, optional strides <s_x>,<s_y>, and <d> filters. Activation functions are specified as follows:
s: Sigmoid
t: Tanh
r: ReLU
l: Linear
m: Softmax
Example: Cr3,3,32
Adds a convolutional layer with ReLU activation, a 3x3 kernel, default strides (1,1), and 32 filters.
Pooling2D Layer
VGSL Spec: <p>(<x>,<y>,<s_x>,<s_y>)
Mp for max-pooling, Ap for average pooling.
Description: Specifies a pooling operation, which reduces the spatial dimensions by applying a window of <x> by <y> and strides of <s_x>,<s_y>.
Example: Mp2,2,2,2
Defines a max-pooling layer with a pool size of 2x2 and strides of 2x2.
Dense (Fully Connected) Layer
VGSL Spec: F(s|t|r|l|m)<d>
Description: Defines a fully connected (dense) layer with <d> units. The non-linearity can be:
s: Sigmoid
t: Tanh
r: ReLU
l: Linear
m: Softmax
Example: Fr64
Adds a dense layer with 64 units and ReLU activation.
RNN Layer (LSTM/GRU/Bidirectional)
VGSL Spec: L(f|r)[s]<n>[,D<rate>,Rd<rate>] for LSTM/GRU, B(g|l)<n>[,D<rate>,Rd<rate>] for Bidirectional RNN
Description: Specifies an RNN layer with n units. The optional dropout D and recurrent dropout Rd rates can be included.
L: LSTM
G: GRU
B: Bidirectional
f: Forward direction, r: Reverse direction, g: GRU, l: LSTM
Example: Lf64,D50,Rd25
Defines an LSTM layer with 64 units, 50% dropout, and 25% recurrent dropout.
Dropout Layer
VGSL Spec: D<rate>
Description: Specifies a dropout layer, where <rate> is the dropout percentage (0–100).
Example: D50
Adds a dropout layer with a 50% dropout rate.
Output Layer
VGSL Spec: O(2|1|0)(l|s)<n>
Description: Defines the output layer. The first value (2, 1, or 0) specifies whether the output is 2D, 1D, or scalar, followed by the activation type (l: linear, s: softmax), and the number of output units (n).
Example: O1s10
Defines a softmax output layer with 10 classes for a 1D sequence.
Reshape Layer
VGSL Spec: Rc2, Rc3, or R<x>,<y>,<z>
Description: The Reshape layer reshapes the output tensor from the previous layer. It has two primary functions:
Rc2: Collapses the spatial dimensions (height, width, and channels) into a 2D tensor. This is typically used when transitioning to a fully connected (dense) layer.
Example: Reshaping from (batch_size, height, width, channels) to (batch_size, height * width * channels).
Rc3: Collapses the spatial dimensions into a 3D tensor suitable for RNN layers. This creates a 3D tensor in the form of (batch_size, time_steps, features).
Example: Reshaping from (batch_size, height, width, channels) to (batch_size, height * width, channels) for input to LSTM or GRU layers.
R<x>,<y>,<z>: Directly reshapes to the specified target shape.
Example:
Rc2 collapses the output from (None, 8, 8, 64) to (None, 4096) for a fully connected layer.
Rc3 collapses the output from (None, 8, 8, 64) to (None, 64, 64) for input to an RNN layer.
R64,64,3 reshapes the output to (None, 64, 64, 3).
More Examples
Explore additional examples and advanced configurations in the tutorials.