bend-ml-tensor-fork

Releases (8)

• 0.1.6.0 indexed
• 0.1.5.0 unindexed
• 0.1.4.0 unindexed
• 0.1.3.0 unindexed
• 0.1.2.0 unindexed
• 0.1.1.1 unindexed
• 0.1.1.0 unindexed
• 0.1.0.0 unindexed

Definitions (63)

main.bend

• type B2
• type B3
• type B4
• def blocks1
• def join2
• def blocks2
• def join3
• def blocks3
• def join4
• def blocks4
• def par1
• def par2
• def par3
• def par4
• def depth_gothe smallest d with 2^d >= n (TA.cap_depth's answer) by doubling a U32: log2(n) steps, where cap_depth compares Nats (a share of every product's time at 512 x 1536)
• def depth
• type PE
• def pk_b2
• def pk_leaflines i0 ..
• def pk_of
• def pk_done block below 16384 cells (MNIST's 100 x 784 and 128 x 784 pack faster in one), else 2^d
• def pk_par
• def tg_c
• def rblockrows i0 ..
• def cbrowscolumns j0 ..
• def rpar
• def cpar
• def t8_tree
• def t8_go
• def gemm_fork
• def gemm_rowstwo or more rows: the shared tree when d > 0, tensor-array's sequential cells at d = 0
• def gemm_any2
• def gemm_anyTA.gemm_any's arguments and result; one row runs the scalar loop at every depth, as in tensor-array 0.1.12.0
• def matmulC = A · B (n x k) · (k x m)
• def matmul_accC + A · B, into C
• def matmul_ntC = A · Bᵀ (n x k) · (m x k)ᵀ: weights with one row per output
• def matmul_nt_accC + A · Bᵀ, into C
• def matmul_tnC = Aᵀ · B (k x n)ᵀ · (k x m): a weight gradient, Xᵀ · dY
• def matmul_tn_accC + Aᵀ · B, into C: dW += Xᵀ · dY
• def mp_s3
• def mp_s2
• def mp_step
• def mp_go
• def mm_precise2
• def matmul_precise
• type SE
• def sr3a row's loss term log(max(p, 1e-12)) and its gradient: cea_row3..cea_row0 without the running loss
• def sr2
• def sr1
• def sr0
• def ce_put
• def ce_row
• def ce_step
• def ce_loop
• def ce_leafrows i0 ..
• def ce_sumthe loss over k terms, in row order from +0.0, as cea_rows adds them
• type CL
• def lab_fita label past the row gives p = 0 and no one-hot cell, as c does: kept as c so that it fits a U32
• def lab_takeup to n labels into a, counted in k; the rest stays a list
• def ce_fin2after the blocks: the loss from t, then any labels past the n rows, in order, then tensor-array's mean
• def ce_par3
• def softmax_ce2
• def softmax_cemean cross-entropy and its gradient, TA.Mat.softmax_ce's arguments and bits, with the rows in up to 2^par blocks (par = 0: tensor-array's own loop)