reshape.v, transpose, broadcast.v, stack.v, split.v exist as core tensor ops
but autograd has no derivative gates for them. Building models with non-trivial graph
shapes cannot backprop through these operations.
Task
Add backward gates for:
reshape / view — gradient reshapes back to original shape
transpose / permute — gradient transposes back
concat — gradient splits along the concatenation axis
stack — gradient unstacks
Files
autograd/ — new gate files
reshape.v,transpose,broadcast.v,stack.v,split.vexist as core tensor opsbut autograd has no derivative gates for them. Building models with non-trivial graph
shapes cannot backprop through these operations.
Task
Add backward gates for:
reshape/view— gradient reshapes back to original shapetranspose/permute— gradient transposes backconcat— gradient splits along the concatenation axisstack— gradient unstacksFiles
autograd/— new gate files