bendtorch-core
Search uses release 0.1.0.0 · 5,204 definitions
Definitions (5,204)
- type OneContractWhat a VJP that sends one contribution promises: it goes to node a with the target's dimensions and holds cell(i) at every in-bounds i.
- law tensor_relu_vjp_valuesd relu(a)/da: with the saved input x (input's shape) and the gradient g (shape), the Relu VJP sends the pointwise ReluBackward of x and g, reduced to the input's shape, to a, and nothing else.
- law tensor_mse_vjp_valuesd mse_loss(a, b)/da and d/db: with the saved differences e = right - left and d = left - right (both of the broadcast shape `diff`) and the gradient g (shape), all replayed, the MseLoss VJP sends the pointwise MseGrad of g and d, reduced to the left operand's shape, to a, then the pointwise MseGrad of g and e, reduced to the right operand's shape, to b.
- law tensor_cross_entropy_vjp_valuesd cross_entropy(a)/da: with the saved log_softmax lp and one-hot rows (both of the input's shape) and the gradient g (shape), all replayed, the CrossEntropy VJP sends two contributions, both to a with the input's dimensions: the pointwise OneHotGrad of g and the one-hot rows, then the pointwise SoftmaxGrad of g and lp, each reduced to the input's shape.
- def cell
- def cells_snoc
- def cells_left`cells` reads the left input only at left_map(plan, k), which is in bounds, so a left history that lists the same words gives the same cells.
- def product_replayed
- def pair_contractTwo delivered contributions, each a replayed tensor with its target's dimensions whose cells hold the stated words, meet the pair contract.
- def permuted_atA permuted copy holds, at every in-bounds index, the source's word at permute_source: its words are `permuted_words`, which are the words of the copy loop's history.
- def factors_snoc
- def first_factorsThe pairs the first product g @ y^T reads, on the copy of g and the transposed copy of y, are the pairs right_transposed_factors names on g and y themselves.
- def second_factorsThe pairs the second product x^T @ g reads, on the transposed copy of x and the copy of g's first half, are the pairs left_transposed_factors names on x and g themselves.
- def MatmulGoalThe record the Matmul VJP law concludes, for the plans the law names.
- def product_cellOne product's cell: the contribution reads the replayed product, and the product's cell is the ordered total over the copies, which first_factors or second_factors turns into the stated pairs.
- def facts_left
- def facts_right
- def facts_compatible
- def matmul_values
- def nt_first_factorsThe pairs g @ W reads on the Second halves of the clones of g and W are the pairs on g and W themselves.
- def nt_second_factorsThe pairs g^T @ x reads on the Second half of the First half of g and the Second half of x are the pairs on g and x themselves.
- def NTGoalThe record the MatmulNT VJP law concludes, for the plans the law names.
- def nt_values
- def frozen_factorsThe pairs g^T @ x reads on the Second half of g and the Second half of x are the pairs on g and x themselves.
- def FrozenGoal
- def frozen_values
- def same_refl
- def frozen_skipsWith a not b, Frozen{a, MatmulNT{a, b, ..}} runs the right product alone.
- def swapped_matrixThe axes transpose_dynamic(t, 0, 1) permutes a rank-2 t by are [1, 0].
- def nt_wordFactor k of cell (row, column) of a `Kept` right operand of dimensions [columns, inner] is the word right^T holds at k * columns + column.
- def tn_wordFactor k of cell (row, column) of a `Swapped` left operand of dimensions [inner, rows] is the word left^T holds at row * inner + k.
- def nt_factors
- def tn_factors
- def one_contractA delivered contribution, a replayed tensor with its target's dimensions whose cells hold the stated words, meets the one-contribution contract.