holocron.nn¶

Non-linear activations¶

class holocron.nn.Mish[source]¶: Implements the Mish activation module from https://arxiv.org/pdf/1908.08681.pdf

class holocron.nn.NLReLU(inplace=False)[source]¶

Implements the Natural-Logarithm ReLU activation module from https://arxiv.org/pdf/1908.03682.pdf

Parameters:: inplace (bool) – should the operation be performed inplace

class holocron.nn.FocalLoss(gamma=2, **kwargs)[source]¶

Implementation of Focal Loss as described in “Focal Loss for Dense Object Detection”

Parameters:

gamma (float, optional) – exponent parameter of the focal loss
weight (torch.Tensor[K], optional) – class weight for loss computation
ignore_index (int, optional) – specifies target value that is ignored and do not contribute to gradient
reduction (str, optional) – type of reduction to apply to the final loss

class holocron.nn.MultiLabelCrossEntropy(**kwargs)[source]¶

Implementation of the cross-entropy loss for multi-label targets

Parameters:

weight (torch.Tensor[K], optional) – class weight for loss computation
ignore_index (int, optional) – specifies target value that is ignored and do not contribute to gradient
reduction (str, optional) – type of reduction to apply to the final loss

class holocron.nn.LabelSmoothingCrossEntropy(eps=0.1, **kwargs)[source]¶

Implementation of the cross-entropy loss with label smoothing on hard target as described in “Attention Is All You Need”

Parameters:

eps (float, optional) – smoothing factor
weight (torch.Tensor[K], optional) – class weight for loss computation
ignore_index (int, optional) – specifies target value that is ignored and do not contribute to gradient
reduction (str, optional) – type of reduction to apply to the final loss

class holocron.nn.MixupLoss(criterion)[source]¶

Parameters:: criterion (callable) – initial criterion to be used on normal sample & targets

class holocron.nn.ConcatDownsample2d(scale_factor)[source]¶

Implements a loss-less downsampling operation described in https://pjreddie.com/media/files/papers/YOLO9000.pdf by stacking adjacent information on the channel dimension.