DANet

A neural network designed to enhance tabular data processing by grouping correlated features and reducing computational complexity.

Functions

class AcceleratedCreator(object)

Creates accelerated versions of neural networks by extracting and compressing layers.

Parameters:

  • input_dim (int) - Input dimension.

  • base_out_dim (int) - Base output dimension.

  • k (int) - Number of branches.

Methods:

  • __call__(self, network) - Accelerate a network by modifying its layers.

  • extract_module(self, basicblock, base_input_dim, fix_input_dim) - Extract and compress a module.

class Extractor(object)

Extracts parameters from abstract layers and computes compressed weights.

Parameters:

  • k (int) - Number of branches.

Methods:

  • get_parameter(self, abs_layer) - Extract parameters from an abstract layer.

  • compute_weights(self, a, b, eps, mu, var, sw, pw, pb, base_input_dim, base_output_dim, k) - Compute compressed weights.

  • __call__(self, abslayer, input_dim, base_out_dim) - Extract and compress a layer.

class CompressAbstractLayer(nn.Module)

Compressed abstract layer with attention and feature weights.

Parameters:

  • att_w (Tensor) - Attention weights.

  • f_w (Tensor) - Feature weights.

  • att_b (Tensor) - Attention bias.

  • f_b (Tensor) - Feature bias.

Input:

  • x (Tensor) - Input tensor.

Output:

  • Tensor - Compressed output with attention mechanism.

def get_parameter(abs_layer)

Extract parameters from an abstract layer.

Parameters:

  • abs_layer - Abstract layer to extract parameters from.

Returns:

  • tuple - Extracted parameters (alpha, beta, eps, mu, var, sparse_weight, process_weight, process_bias).

def compute_weights(a, b, eps, mu, var, sw, pw, pb, base_input_dim, base_output_dim, k)

Compute compressed weights from extracted parameters.

Parameters:

  • a (Tensor) - Alpha parameter.

  • b (Tensor) - Beta parameter.

  • eps (float) - Epsilon value.

  • mu (Tensor) - Mean parameter.

  • var (Tensor) - Variance parameter.

  • sw (Tensor) - Sparse weights.

  • pw (Tensor) - Process weights.

  • pb (Tensor) - Process bias.

  • base_input_dim (int) - Base input dimension.

  • base_output_dim (int) - Base output dimension.

  • k (int) - Number of branches.

Returns:

  • tuple - Computed weights (W_att, W_fc, B_att, B_fc).

Referencses:

Chen, J., Liao, K., Wan, Y., Chen, D. Z., & Wu, J. (2022). DANets: Deep Abstract Networks for Tabular Data Classification and Regression. arXiv:2112.02962 [cs.LG]. https://arxiv.org/abs/2112.02962