Just like torch.nn.Linear, except each input gets multiplied by different weights depending its (integer) type

Returns the featuretools summary statistics about the text (num words and num_chars), but normalized

HAN layer. Arguments --------- meta_paths : list of metapaths, each as a list of edge types in_size : input feature dimension

When you want to ignore a feature

Denoising AutoEncoder (DAE) Loss. Tries to reconstruct the input with an MLP, and penalizes the difference. Intended for use in pretraining

GAT using different linear mappings for each node and edge type todo: compare to this relational GAT model: https://openreview.net/pdf?id=Bk

GCN using different linear mappings for each node and edge type

r""" https://arxiv.org/pdf/1708.02002.pdf

Graph Attention Network as described in https://arxiv.org/pdf/1710.10903.pdf

Graph Convolutional Network as described in https://arxiv.org/pdf/1609.02907.pdf

Base class for all GNN models

Samples elements randomly from a given list of indices for imbalanced dataset Arguments: num_samples (int, optional): number of sampl

Model that ignores relational structure and just inits the nodes, pools them, and computes the output. Inspired by https://arxiv.org/pdf/1905

Relational version of Graph Attention Network

Relational Graph Convolutional Network as described in https://arxiv.org/abs/1703.06103

Straightforward MLP model for tabular data, loosely based on github.com/fastai/fastai/blob/master/fastai/tabular layer_sizes can contain int

Base class for all tabular models

r""" Loss function that augments the cross entropy loss with an encouragement for the TabNet masks to be sparse