src.main.Viyog

class src.main.Viyog(model, device=None)

Bases: object

Wrapper that captures first-layer activations and computes an activation-norm based score.

The wrapper finds the first convolutional layer of model (prefers an attribute named conv1 if present) and registers a forward hook that saves the layer’s output for later inspection. Use fit() to compute a training baseline (mean infinity-norm of per-sample flattened activations). Then call score() or score_loader() to compute Viyog scores on new data.

Parameters:

• model (torch.nn.Module) – The model to wrap. The forward pass must traverse at least one Conv{1,2,3}d module (or provide an attribute named conv1).

• device (torch.device | str | None, optional) – Device to use for computation. If None (default) the device is inferred from model parameters as needed.

id_norm_scores_mean

Mean per-sample infinity norm computed by fit(). None until fit completes successfully.

Type:

float or None

__init__(model, device=None)

Parameters:

• model (torch.nn.Module)

• device (Optional[torch.device | str])

Methods

Viyog_Score(num[, Temperature])	Convert centered norms into bounded scores in approximately (-1, 1).
__init__(model[, device])
close()	Remove the hook (call when you no longer need the Viyog wrapper).
fit(self, trainloader)
score(self, x[, Temperature])
score_loader(loader[, Temperature])	Helper that returns a single 1D tensor of scores for the whole loader.

close()

Remove the hook (call when you no longer need the Viyog wrapper).

Return type:

None

static Viyog_Score(num, Temperature=1000.0)

Convert centered norms into bounded scores in approximately (-1, 1).

Parameters:

• num (torch.Tensor) – Tensor (any shape) of centered norms (i.e., per-sample norm minus training mean). Must be on the same device you want the result on.

• Temperature (float, optional) – Temperature scaling factor. Default is 1000.0.

Returns:

Tensor of same shape as num with values approx in (-1, 1).

Return type:

torch.Tensor

fit(trainloader)

Compute the mean of per-sample infinity norms of the first-layer activations across the provided trainloader and store it as a Python float.

Parameters:

trainloader (torch.utils.data.DataLoader) – DataLoader providing samples to estimate the mean activation norm. Each yielded batch should be (inputs, labels) or where inputs are the first item.

Returns:

The computed mean per-sample infinity norm (stored in self.id_norm_scores_mean).

Return type:

float

score(x, Temperature=1000.0)

If x is a Tensor (batch): returns a 1D tensor of scores for that batch. If x is a DataLoader: processes entire loader and returns concatenated scores.

Parameters:

• x (torch.Tensor or torch.utils.data.DataLoader) – Input batch (tensor) or data loader to score.

• Temperature (float, optional) – Temperature scaling factor passed through to Viyog_Score().

Returns:

1D tensor of scores (device equals inferred device).

Return type:

torch.Tensor

score_loader(loader, Temperature=1000.0)

Helper that returns a single 1D tensor of scores for the whole loader.

Return type:

Tensor

Parameters:

• loader (torch.utils.data.DataLoader)

• Temperature (float)