grain.autograd

A module for a variable used as a node in autograd computation graph

TODO: - support shape ops

Members

DeviceStorage alias DeviceStorage(T) = CuArray!T: Undocumented in source.
ElementType alias ElementType(V : Variable!(Elem, dim, Storage), Elem, size_t dim, alias Storage) = Elem: an alias of element type (e.g., float, double and int) of variable
HostStorage alias HostStorage(T) = T[]: CPU storage (i.e., GC dynamic array)
isDevice alias isDevice = isDeviceMemory: Undocumented in source.

gradSlice auto gradSlice(V v)
iterVariables void iterVariables(C* chain, string prefix): kind of std.algorithm.each for iterating variables inside a chain
length auto length(V v): total number of elements in variable
refIterVariables void refIterVariables(C chain, string prefix): kind of std.algorithm.each for iterating variables inside a chain
to auto to(T[] src): CUDA -> CPU memory conversion
to auto to(Src src): CPU -> CUDA memory conversion
to Variable!(T, dim, Dst) to(Variable!(T, dim, Src) src): copy variable into the other device (e.g., CPU -> CUDA or CUDA -> CPU)
uninit auto uninit(Variable!(T, dim, HostStorage) v): create new variable with uninitialized array and the same shape/strides to v on CPU
uninit auto uninit(Variable!(T, dim, DeviceStorage) v): create new variable with uninitialized array and the same shape/strides to v on CUDA
uninitVariable auto uninitVariable(uint[dim] shape, bool requiresGrad): create new variable with uninitialized array of shape on CPU/CUDA
variable auto variable(Sl sl, bool requiresGrad): a helper function to create variable object from slice
variable auto variable(T x)
auto variable(A a, bool requiresGrad): a helper function to create variable object from CPU/CUDA array
zero_ auto zero_(T[] s): fill CPU array with zero
zeros auto zeros(size_t n): create new CPU array filled with zero

BackProp struct BackProp: stores information for backpropagation
UntypedVariable struct UntypedVariable: type-erased variable used in BackProp object
Variable struct Variable(T, size_t dim, alias Storage = HostStorage, SliceKind kind = Contiguous): A variable has autograd ability with mir.ndslice.Slice like data

DimensionCount enum size_t DimensionCount(V : Variable!(Elem, dim, Storage), Elem, size_t dim, alias Storage);: a function to get the number of dimensions of variable
isHost enum bool isHost(V : Variable!(Elem, dim, Storage), Elem, size_t dim, alias Storage);: a trait to identify variable stored in CPU memory
isVariable enum bool isVariable(T);: a trait to identify variable object