grain.functions.unary

A module for unary functions

TODO: support cudnn functions (see PDF manual in .deb for detail https://developer.nvidia.com/cudnn) - activation (e.g., clipped-relu, elu), cudnnActivationForward/Backward - (non-log) softmax, cudnnSoftmaxForward/Backward - sqrt not, cudnnOpTensor - transform (e.g., contiguous or permute strides), cudnnTransformTensor - reshape (i.e., view), ...??? - reduce (sum, prod, min, max, amax, avg), cudnnReduceTensor - pool (max, average), cudnnPoolingForward/Backward - dropout, cudnnDropoutForward/Backward

Members

Aliases

AvgPool alias AvgPool(T, size_t poolDims) = Pool!(false, T, poolDims): Undocumented in source.
MaxPool alias MaxPool(T, size_t poolDims) = Pool!(true, T, poolDims): Undocumented in source.

Functions

logsumexp logsumexp(S x): Undocumented in source. Be warned that the author may not have intended to support it.
ndStrided auto ndStrided(S s, ptrdiff_t[dim] strides): n-dimensional strided
sumNdStrided void sumNdStrided(S src, D dst, ptrdiff_t[dim] strides): Undocumented in source. Be warned that the author may not have intended to support it.
unaryFunc void unaryFunc(Variable!(float, dim, DeviceStorage) x): wrapper of CUDA kernel unary functions
unaryPow void unaryPow(Variable!(float, dim, DeviceStorage) x, float power): wrapper of CUDA kernel pow function
unpad auto unpad(S s, size_t[N] lengths): only both is supported

Manifest constants

CUDNN_ACTIVATION_IMPL_MIXIN enum CUDNN_ACTIVATION_IMPL_MIXIN;

Structs

Abs struct Abs(T, size_t dim): y = abs x
Cos struct Cos(T, size_t dim): y = cos x
Exp struct Exp(T, size_t dim): y = exp x
Log struct Log(T, size_t dim): y = exp x
LogSoftmax struct LogSoftmax(T, size_t dim = 2)
Neg struct Neg(T, size_t dim): y = -x
Pool struct Pool(bool isMax, T, size_t poolDims): max pooling function
PoolRefImpl struct PoolRefImpl(alias poolFun): reference implementaion of pooling function
Pow struct Pow(T, size_t dim): y = pow x
ReLU struct ReLU(T, size_t dim): TODO implement scale with cudnnScaleTensor rectified linear unit nonlinearity (using cuDNN)
Reciprocal struct Reciprocal(T, size_t dim): y = 1 / x
Scale struct Scale(T, size_t dim): y = alpha * x
Sigmoid struct Sigmoid(T, size_t dim): sigmoid function
Sin struct Sin(T, size_t dim): y = sin x
Tan struct Tan(T, size_t dim): y = tan x
Tanh struct Tanh(T, size_t dim): hyperbolic tangent