The LSTM-based subnetwork that is used in TFN for text

The subnetwork that is used in TFN for video and audio in the pre-fusion stage

Transformer encoder consisting of *args.encoder_layers* layers. Each layer is a :class:`TransformerEncoderLayer`. Args: embed_tok

Computes and stores the average and current value

Adapted from https://github.com/wzell/cmd/blob/master/models/domain_regularizer.py

Contrastive Predictive Coding: score computation. See https://arxiv.org/pdf/1807.03748.pdf. Args: x_size (int): embeddin

Compute the Modality Mutual Information Lower Bound (MMILB) given bimodal representations. Args: x_size (int): embedding size of input mod

Residual autoencoder using fc layers layers should be something like [128, 64, 32] eg:[128,64,32]-> add: [(input_dim, 128), (128, 64)

The subnetwork that is used in TFN for video and audio in the pre-fusion stage

Data processing using albumentation same as torchvision transforms

Multi-headed attention. See "Attention Is All You Need" for more details.

Computes and stores the minimum loss value and its epoch index

This module produces sinusoidal positional embeddings of any length. Padding symbols are ignored, but it is necessary to specify whether padding

Encoder layer block. In the original paper each operation (multi-head attention or FFN) is postprocessed with: `dropout -> add residual -> lay

Cuda out of memory

Computes and stores the minimum loss value and its epoch index

Computes and stores the minimum loss value and its epoch index