bendtorch-core

Releases (1)

• 0.1.0.0 indexed

Definitions (5,204)

src/autograd/gradient.bend

• type Slot
• type Outcome
• type State
• type Step
• def step_write
• def step_total
• def step_contribution
• def step
• def continue_step
• def conversion_error
• def decided
• def loop
• def held_checked
• def held
• def agreed
• def checked
• def accumulate
• def gradThe gradient a slot holds, or None when it never received one or was reset.
• def resetDrops a held gradient: the next contribution moves in unchanged.
• def gradientThe gradient a slot holds; zeros of shape when it never received one.
• def error_code

src/autograd/gradient_LAWS.bend

• def add_word
• def combinedThe value the specification assigns to one position after a contribution is accumulated into a slot: an Unset slot takes the contribution's word as it is, and a Held slot adds the contribution's word to its own with one F32 add, the held total first.
• def held_dimensions
• def held_observe
• def Inside
• def Mapping
• def SlotKept
• def mismatched
• def RefusalKeeps
• def Fits
• def accumulated
• def Zeros
• type MappingModel
• type RefusalModel
• type Success
• type ZerosModel
• law gradient_unset_zerosA slot that never received a contribution reads as zeros of its shape.
• law gradient_heldA slot that holds a gradient reads as that gradient.
• law accumulate_mappingOn success the slot holds a gradient with the slot's dimensions.
• law accumulate_refusal_keepsEvery refusal is a shape mismatch, decided before any write: the contribution comes back unchanged, and so does the slot.
• law accumulate_valid_successA contribution and a slot that both agree with the slot's shape always accumulate.
• law grad_unset_noneLike PyTorch's .grad, a slot that never received a contribution has no gradient: grad reads it as None, not as zeros.
• law grad_held_someA slot that holds a gradient reads as that gradient.
• law reset_unsetreset returns every slot to Unset, like zero_grad(set_to_none=True).
• law reset_then_accumulate_moves_inAfter a reset, a contribution with the slot's dimensions moves in unchanged, so a -0.0 word stays -0.0 instead of becoming +0.0 + -0.0.
• law accumulate_wrong_shape_refusedA contribution whose dimensions differ from the slot's is refused with ShapeMismatch, and the slot and the contribution come back unchanged, even when the element counts agree ([3, 2] or [6] for a [2, 3] slot).

src/autograd/gradient_PROOF.bend

• type Holds
• def lt_self
• def lt_succ_self
• def below_zero
• def below_one_more
• def lt_step
• def words_dimensions
• def same_checked
• def same_dimensions
• def same_trans
• def word_exact
• def unset_zeros
• def state_contribution
• def state_total
• def step_owner
• def step_continued
• def write_owner
• def total_owner
• def contribution_retains
• def step_retains
• def partial
• def Inv
• def StepInv
• def Dims
• def Values
• def LoopInv
• def written_true
• def written_false
• def position_inv
• def write_inv
• def total_inv
• def contribution_inv
• def step_inv
• def continue_inv
• def conversion_inv
• def decided_inv
• def loop_inv
• def initial_inv
• def finish_mapping
• def held_checked_mapping
• def held_mapping
• def agreed_mapping
• def checked_mapping
• def write_success
• def total_success
• def contribution_success
• def step_success
• def continue_success
• def decided_success
• def loop_success
• def full_loop
• def held_checked_success
• def held_success