dropout2d

→Tensor

dropout2d(x: Tensor, p: float = 0.5, training: bool = True)

source

Channel-wise dropout for 2-D (spatial) feature maps.

Drops entire 2-D feature maps of an $(N, C, H, W)$ tensor with probability p and rescales the survivors by $1/(1-p)$ . This is the dropout variant of choice for convolutional networks (Tompson et al. 2015 "SpatialDropout"): adjacent pixels within a feature map are spatially correlated, so elementwise dropout removes very little information. Channel-wise masking forces independence between feature maps instead.

Parameters

xTensor

Input tensor of shape

(N, C, H, W)

pfloat= 0.5

Channel-drop probability in

[0, 1)

(default 0.5).

trainingbool= True

When False, identity (default True).

Returns

Tensor

Same shape and dtype as x.

Notes

For each batch element $n$ and channel $c$ , draw $m_{n,c} \sim \mathrm{Bernoulli}(1 - p)$ and broadcast over the spatial dimensions:

y_{n, c, h, w} = \frac{m_{n, c}}{1 - p} \cdot x_{n, c, h, w}.

Examples

>>> import lucid
>>> from lucid.nn.functional import dropout2d
>>> x = lucid.ones((2, 8, 4, 4))
>>> y = dropout2d(x, p=0.5, training=True)