bendtorch-core

Search uses release 0.1.0.0 · 5,204 definitions

Releases (1)

0.1.0.0 indexed

Definitions (5,204)

src/autograd/reduce.bend

def error_code

src/autograd/reduce_fiber_proof.bend

def add_comm
def lt_succ_eq
type Splitt = quotient * size + remainder with remainder < size.
def split_step
def split_next
def split
def split_divmodThe quotient and remainder of a split are Nat.div and Nat.mod.
def div_of
def mod_of
type Decomposedt = (t / size) * size + t % size, and t % size < size.
def decomposed
def decompose
def lt_irrefl
def le_add_cong
def le_trans
def mul_le_mono
def quotient_belowBelow a * size, the quotient by size is below a.
def positive_of
def div_below
def eq_add_cancela + x = a + y exactly when x = y.
def ne_from
def mask_atOn a kept row of left stride `size`, the position with digit d and walk w below size lands on t exactly when d is t's quotient and w its remainder.
def mask_off_decided
def mask_off
def kept
def maskedThe contribution offsets base + walk_right(p) of the positions p = cursor ..
def visitedThe contribution offsets `Reduce.fiber` reads, in order.
def append_nil
def kept_append

src/autograd/reduce_LAWS.bend

def owner
def succeeded
def word_at
def add_word
def matched
def contribute
def fiberThe specified accumulation: start at +0.0 and visit contribution positions 0, 1, ..., count - 1 in ascending order; each position whose independent broadcast projection, Projection.offset_spec, has target_offset equal to target_index adds the word at its source_offset with one F32 add.
def planned_fiber
def Listedvalues lists the contribution's words: the checked observation at every U32 position is the word at that position.
def MappingWith
def Mapping
def ZeroValues
type Retained
type MappingModel
type Success
type Length
type Zeros
law reduce_retains_inputThe contribution owner comes back with the same dimensions and observations.
law reduce_mappingOn success the output has the target's dimensions, and for every list of words that lists the contribution, each in-bounds target position holds the ordered fiber sum of those words.
law listed_lengthA list of words that lists the contribution has exactly the contribution's logical length.
law reduce_valid_successA target whose right-aligned broadcast with the contribution shape is the contribution shape itself always succeeds.
law reduce_empty_zerosReducing an empty contribution gives zeros at every target position, not an empty tensor: for example [0, 3] reduced to [1, 3] is three zeros.
law reduce_shape_errorWhen the broadcast planner rejects the target and contribution shapes, reduce_to fails with exactly that planner error and returns the contribution owner.
law reduce_not_reducibleWhen the planned broadcast of the target is not the contribution's shape (for example a [2, 3] target for a [3] contribution), reduce_to fails with NotReducible and returns the contribution owner.

bendtorch-core

Releases (1)

Definitions (5,204)

src/autograd/reduce.bend

src/autograd/reduce_fiber_proof.bend

src/autograd/reduce_LAWS.bend

src/autograd/reduce_PROOF.bend