models.soda.SODa#

class SODa(backbone: BackboneGen, neck: NeckGen, head: Head, loss_ratio: int, time_window: int = 0)[source]#

Bases: Model

Basic object detector class

Implements the basic functions for calculating losses, training the network and generating predictions. The network model is passed as a parameter when initializing.

Parameters:

backbone (BackboneGen) – Main network.
neck (NeckGen) – Feature Map Extraction Network.
head (Head) – Network for transforming feature maps into predictions.
loss_ratio (int) – The ratio of the loss for non-detection to the loss for false positives. The higher this parameter, the more guesses the network generates. This is necessary to keep the network active.
time_window (int, optional) – The size of the time window at the beginning of the sequence, which can be truncated to a random length. This ensures randomization of the length of training sequences and the ability of the network to work with streaming information. Defaults to 0.

Methods

`configure_optimizers`	Returns the configured optimizer for training the network
`forward`	Direct network pass
`loss`	Loss calculation function.
`predict`	Returns the network's predictions based on the input data
`test_step`	Network test step
`training_step`	Network training step
`validation_step`	Network validation step

Attributes

training

configure_optimizers() → Optimizer[source]#: Returns the configured optimizer for training the network

forward(X: Tensor) → tuple[Tensor, Tensor, Tensor][source]#

Direct network pass

Parameters:

X (torch.Tensor) – Input data.

Returns:

List of three tensors:

Anchors. Shape [anchor, 4].
Class predictions. Shape [ts, batch, anchor, num_classes + 1].
Box predictions. Shape [ts, batch, anchor, 4].

Return type:

tuple[torch.Tensor, torch.Tensor, torch.Tensor]

loss(preds: tuple[Tensor, Tensor, Tensor], labels: Tensor) → Tensor[source]#

Loss calculation function.

Parameters:

preds (Tuple[torch.Tensor, torch.Tensor, torch.Tensor]) –
Predictions made by a neural network. Contains three tensors:
1. anchors: Shape [anchor, 4]
2. cls_preds: Shape [ts, batch, anchor, num_classes + 1]
3. bbox_preds: Shape [ts, batch, anchor, 4]
labels (torch.Tensor) –
Tensor shape [num_labels, 5].

One label contains: class id, xlu, ylu, xrd, yrd.

Returns: