model module

Model module.

This module is intended to provide a set of Networks for features extraction and predictions leveraging domain adaptation.

The code for the Transfer Network is taken and readjusted from the DeepDA project.

The code can be found here: https://github.com/jindongwang/transferlearning/tree/master/code/DeepDA

class model.FeaturesExtractor(input_dim: int, hidden_dim: int, output_dim: int, num_layers: int = 1, device: str = 'cpu')

Bases: Module

Features extraction Recurrent network based on Bi-LSTM. It requires a dataloader to be organized in sequences.

The structure is the following: 1. Bi-LSTM 2. LSTM 3. Linear 4. LayerNorm 5. PReLU 6. Dropout 7. Linear

__init__.

Parameters:

  • input_dim (int) – input data dimension

  • hidden_dim (int) – layers hidden dimensions

  • output_dim (int) – number of features to extract (output)

  • num_layers (int) – number of layers of the BiLSTM

  • device (str) – device to use, can be either ‘cpu’ or ‘cuda’

forward(x: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class model.Regressor(input_dim: int, hidden_dim: int = 32)

Bases: Module

Predictive linear network performing a regression task.

__init__.

Parameters:

  • input_dim (int) – input data dimension

  • hidden_dim (int) – layers hidden dimensions

forward(x: Tensor) → Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class model.TransferNet(input_dim: int, fe_hidden_dim: int = 24, fe_output_dim: int = 8, fe_num_layers: int = 1, hidden_dim: int = 32, transfer: bool = True, transfer_loss: str = 'adv', max_iter: int = 200, learning_rate: float = 0.001, weight_decay: float = 0, scheduler_step: int = 100, scheduler_gamma: float = 0.1, device: str = 'cpu', **kwargs)

Bases: Module

Main network that gathers features extracted from the base network, the predictive network and the transfer loss.

__init__.

Parameters:

  • input_dim (int) – input data dimension

  • fe_hidden_dim (int) – hidden dimension of layers of feature extraction network

  • fe_output_dim (int) – output dimension of feature extraction network

  • fe_num_layers (int) – number of layers of feature extraction Bi-LSTM network

  • hidden_dim (int) – hidden dimension of layers of predictive network

  • transfer (bool) – true to enable transfer

  • transfer_loss (str) – transfer loss to use. Available values are: * ‘adv’: adversarial * ‘daan’: DAAN * ‘mmd’: MMD * ‘coral’: CORAL * ‘bnm’: BNM

  • max_iter (int) – maximum number of iterations of the Lambda Scheduler used in DAAN and Adversarial losses

  • learning_rate (float) – learning rate

  • weight_decay (float) – optimizer weight decay

  • scheduler_step (int) – step size for the step scheduler

  • scheduler_gamma (float) – gamma for the step scheduler

  • device (str) – ‘cpu’ or ‘cuda’

  • kwargs –

epoch_based_processing(*args, **kwargs)

Handle DAAN loss in a distinct way. Update the DAAN dynamic factor during training time.

Parameters:

  • args –

  • kwargs –

forward(source: Tensor, target: Tensor, source_label: float64)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

get_parameters(initial_lr: float = 1.0)

Define specific parameters and adjust the learning rate for transfer losses. Typically, the learning rate for the base network is slowed.

Parameters:

initial_lr (float) – initial learning rate

predict(x: Tensor) → Tensor

Output a prediction

Parameters:

x (torch.Tensor) – input tensor