CRNN Model Configuration

This is a simple CRNN (Convolutional Recurrent Neural Network) configuration supporting the following layer types:

• conv2d
• lstm
• fc (fully connected)
• dropout
• batchnorm
• layernorm

Example Configuration (config_crnn.yaml)

name: crnn_simple
layer_configs:
  - type: conv2d
    filters: 64
    kernel_size: [3, 3]
    strides: [1, 1]
    activation: relu

  - type: batchnorm
    momentum: 0.99
    epsilon: 1e-3

  - type: dropout
    rate: 0.2

  - type: lstm
    units: 128

  - type: layernorm
    epsilon: 1e-5

  - type: fc
    units: 64
    activation: relu

  - type: dropout
    rate: 0.3

  - type: fc
    units: 1
    activation: linear

Layer Descriptions

• conv2d: 2D convolutional layer for spatial feature extraction

• batchnorm: Batch normalization layer

• momentum: Decay rate for moving average (e.g. 0.99)

• epsilon: Small constant to avoid division by zero (e.g. 1e-3)

• dropout: Regularization layer to prevent overfitting

• rate: Fraction of input units to drop (e.g. 0.2)

• lstm: Long Short-Term Memory layer for sequential modeling

• units: Number of hidden units

• layernorm: Layer normalization across features

• epsilon: Small constant for numerical stability (e.g. 1e-5)

• fc: Fully connected (dense) layer

• units: Number of neurons

• activation: Activation function (e.g. relu, sigmoid, linear)

Only the above layers and parameters are allowed. Ensure your configuration adheres to this format for compatibility.