bendtorch

Releases (6)

• 1.3.0.0 indexed
• 1.2.0.0 unindexed
• 1.1.1.0 unindexed
• 1.1.0.0 unindexed
• 1.0.1.0 unindexed
• 1.0.0.0 unindexed

Definitions (1,049)

tensor.bend

• def all_mapevery position to position 0: the whole-tensor reductions
• def sum_allthe sum of every element: a tensor of shape [] (one number)
• def sum_all_bwd
• def mean_all
• def mean_all_bwd
• def reshape
• def flat_numeln * 1 == n, so a shape and its flattening hold the same count
• def flatten
• def permutedthe shape with axes reordered by p: axis d of the result is axis p[d] of s
• def count
• def all_onceevery axis 0..n-1 exactly once, from axis i
• def is_permp is a permutation of the axes of a rank-n tensor
• def perm_axesreading s at every position of the permuted shape: axis d has s's size and s's stride of axis p[d]; the last first
• def perm_map
• def permute
• def permute_bwdthe gradient of permute: dy put back where each number came from
• def swap_perm0..n-1 with i and j swapped
• def transpose_p
• def narrow_okaxis d exists and start + len fits in it
• def narrow_map
• def narrowpositions st ..
• def narrow_bwdthe gradient of narrow: dy in its window, zeros elsewhere
• def samethe same sizes on every axis
• def cat_eqa and b have the same rank, and the same sizes on every axis but d
• def cat_ok
• def cat_shape
• def cat_map_awhere a's and b's positions go in the result
• def cat_map_b
• def t3_cat
• def cat_b
• def cata then b along axis d
• type CatGthe gradients of cat: dz's window of a, and dz's window of b (dz comes back); returned as TP<a, b> after dz
• def catg_c
• def catg_b
• def cat_bwd
• type Geowhere the operands of out[b, i, j] = sum over t of X[b·xb + i·xi + t·xt] · Y[b·yb + t·yt + j·yj] are
• type Dot
• def Dot.then
• def dot_lacc + the k products from x[xo] and y[yo], stepping xt and yt; rx and ry hold the first pair, already read
• def x_off
• def y_off
• def geo_k
• def geo_xt
• def geo_yt
• def bmm_lout[j] for cnt positions from j
• def u
• def bmm
• def bmm_dxdX = dZ · Yᵀ per batch: [bs, n, m] · [bs, k, m]ᵀ = [bs, n, k]; dZ and Y come back
• def bmm_dydY = Xᵀ · dZ per batch: [bs, n, k]ᵀ · [bs, n, m] = [bs, k, m]; X and dZ come back