mmcls.models.LabelSmoothLoss

class mmcls.models.LabelSmoothLoss(label_smooth_val, num_classes=None, mode='original', reduction='mean', loss_weight=1.0)

Initializer for the label smoothed cross entropy loss.

Refers to Rethinking the Inception Architecture for Computer Vision

This decreases gap between output scores and encourages generalization. Labels provided to forward can be one-hot like vectors (NxC) or class indices (Nx1). And this accepts linear combination of one-hot like labels from mixup or cutmix except multi-label task.

Parameters

• label_smooth_val (float) – The degree of label smoothing.

• num_classes (int, optional) – Number of classes. Defaults to None.

• mode (str) – Refers to notes, Options are ‘original’, ‘classy_vision’, ‘multi_label’. Defaults to ‘original’

• reduction (str) – The method used to reduce the loss. Options are “none”, “mean” and “sum”. Defaults to ‘mean’.

• loss_weight (float) – Weight of the loss. Defaults to 1.0.

Notes

if the mode is “original”, this will use the same label smooth method as the original paper as:

\[(1-\epsilon)\delta_{k, y} + \frac{\epsilon}{K}\]

where epsilon is the label_smooth_val, K is the num_classes and delta(k,y) is Dirac delta, which equals 1 for k=y and 0 otherwise.

if the mode is “classy_vision”, this will use the same label smooth method as the facebookresearch/ClassyVision repo as:

\[\frac{\delta_{k, y} + \epsilon/K}{1+\epsilon}\]

if the mode is “multi_label”, this will accept labels from multi-label task and smoothing them as:

\[(1-2\epsilon)\delta_{k, y} + \epsilon\]