mlbend-generative
Search uses release 0.1.1.0 · 36 definitions
Definitions (36)
- def mix32murmur3's fmix32 finaliser, over U32 arithmetic that wraps
- def unxor13the 13-shift undone: xor the shifted value twice more
- def mix32_invthe inverse of mix32, in closed form: the 16-shift is an involution on 32 bits; the 13-shift inverts by the triple xor; the multiplies invert by their modular inverses mod 2^32
- def hash_ofthe hash of one counter: the step salt times the golden ratio, the index added beside it
- def uniforma uniform in (0, 1]: the low 24 bits of the hash plus one, exactly representable, so log(u) is never log(0)
- def gauss_bodyone Gaussian of a counter, by parity: the even index of the pair brings the radius, the odd one the angle
- def gauss_at
- def noise_fromThe noise of counter s, left numbers from index `from` on.
- def noise_ofthe noise of counter s: m numbers, the index 0 first
- law noise_lengthLAW 1 (noise_length): m numbers asked for are m numbers built.
- def cos_f
- def alpha_atᾱ_t of the cosine schedule over `steps` steps (t = 1 ..
- def beta_atβ_t, the ratio the reverse step removes: 1 - ᾱ_t/ᾱ_{t-1}, clamped at 0.99 — the cosine schedule's last ratio is 1 in F32 (ᾱ_T ~ 1e-15), and the reverse step's 1/sqrt(1-β) would amplify any prediction error
- def alphas_fromThe schedule from step `from` on, `left` entries.
- def alphasᾱ_1 ..
- law alphas_lengthLAW 2 (alphas_length): a schedule of t entries is t entries.
- def lin_beta_atβ_t of the linear schedule over `steps` steps (t = 1 ..
- def lin_alphas_fromThe schedule ᾱ_1 ..
- def lin_alphasᾱ_1 ..
- law lin_alphas_lengthLAW 4 (lin_alphas_length): a linear schedule of t entries is t entries.
- def freq_ofthe frequency of rung p: a decade ladder 0.001 ..
- def temb_fromThe embedding of step t over `steps`, from rung `p` on, `left` rungs — each rung a sine and a cosine.
- def temb_ofthe embedding of step t: k rungs, the lowest frequency first
- law temb_lengthLAW 3 (temb_length): an embedding of k rungs is 2k numbers.
- def groups_ofHow many entries the walk emits over n entries.
- def strided_fromEvery k-th entry of xs, the (skip+1)-th first: drop `skip`, emit one, drop k-1, emit one, and so on to the end of the list.
- def stridedevery k-th entry of xs, the k-th first
- law strided_lengthLAW 5 (strided_length): every k-th entry of n entries is exactly the number the count says.
- def sigma_of
- type StepThe five coefficients of one step: the four roots and the noise scale.
- def step_of
- def ddim_elemone element of the step: x0 folded into x_prev
- def head_or_zerothe head of a list, or 0.0 when it is empty: the missing eps or z of a step whose z is not drawn at all (eta = 0 passes Nil, and sigma·0.0 is 0.0)
- def tail_or_nilthe rest of a list, or empty: the tail beside head_or_zero above
- def ddimOne step over the whole sample: elementwise, and the shape follows x — n elements in, n elements out (LAW 6).
- law ddim_lengthLAW 6 (ddim_length): a step over n elements is n elements.