from functools import reduce import inspect import numpy as np import operator as op def sigmoid(x): return 1.0 / (1 + np.exp(-x)) CHOOSE_CACHE = {} def choose_using_cache(n, r): if n not in CHOOSE_CACHE: CHOOSE_CACHE[n] = {} if r not in CHOOSE_CACHE[n]: CHOOSE_CACHE[n][r] = choose(n, r, use_cache=False) return CHOOSE_CACHE[n][r] def choose(n, r, use_cache=True): if use_cache: return choose_using_cache(n, r) if n < r: return 0 if r == 0: return 1 denom = reduce(op.mul, range(1, r + 1), 1) numer = reduce(op.mul, range(n, n - r, -1), 1) return numer // denom def get_num_args(function): return len(get_parameters(function)) def get_parameters(function): return inspect.signature(function).parameters # Just to have a less heavyweight name for this extremely common operation # # We may wish to have more fine-grained control over division by zero behavior # in the future (separate specifiable values for 0/0 and x/0 with x != 0), # but for now, we just allow the option to handle indeterminate 0/0. def clip(a, min_a, max_a): if a < min_a: return min_a elif a > max_a: return max_a return a def clip_in_place(array, min_val=None, max_val=None): if max_val is not None: array[array > max_val] = max_val if min_val is not None: array[array < min_val] = min_val return array def fdiv(a, b, zero_over_zero_value=None): if zero_over_zero_value is not None: out = np.full_like(a, zero_over_zero_value) where = np.logical_or(a != 0, b != 0) else: out = None where = True return np.true_divide(a, b, out=out, where=where) def binary_search(function, target, lower_bound, upper_bound, tolerance=1e-4): lh = lower_bound rh = upper_bound while abs(rh - lh) > tolerance: mh = np.mean([lh, rh]) lx, mx, rx = [function(h) for h in (lh, mh, rh)] if lx == target: return lx if rx == target: return rx if lx <= target and rx >= target: if mx > target: rh = mh else: lh = mh elif lx > target and rx < target: lh, rh = rh, lh else: return None return mh