z = GAP(X); s = sigmoid(W2 relu(W1 z)); Y = s * X per channel –
sigmoid, not softmax, so channels are gated independently.
Usage
morie_geron_senet_squeeze_excite(X, W1, W2)
Arguments
- X
(H, W, C) feature map.
- W1
(bottleneck, C) squeeze weights.
- W2
(C, bottleneck) excite weights.
Value
List with output, scale, squeeze, hidden,
reduction_ratio.
Examples
set.seed(4)
X <- array(rnorm(4 * 4 * 3), c(4, 4, 3))
r <- morie_geron_senet_squeeze_excite(X, W1 = matrix(0.3, 2, 3),
W2 = matrix(0.3, 3, 2))
r$scale
#> [1] 0.5359777 0.5359777 0.5359777
