def __call__(self, params, update=True):
self._prev_params2 = self._prev_params
if self._prev_params is None:
if update:
self._prev_params = nanguardt(params)
return False
delta = (self._prev_params - params).cpu().numpy()
mags = numpy.abs(delta)
dist = nanguard(numpy.sum(mags))
self._last_delta2 = self._last_delta
self._last_delta = self.delta
self.delta = delta
self.dist = dist
delta_2step = nanguard(numpy.sum(
numpy.abs(self._last_delta2 -
delta))) if self._last_delta2 is not None else 0
delta_1step = nanguard(numpy.sum(
numpy.abs(self._last_delta -
delta))) if self._last_delta is not None else 0
largest = numpy.argsort(delta)
if update:
self._prev_params = params
spaces = " " * 6
ups = ', '.join(f'{idx}={delta[idx]:0.3f}' for idx in largest[:-5:-1])
mag_std = numpy.std(mags)
sys.stdout.write(
f"converged dist: {dist:0.4f}. /param: {dist/len(params):0.3e}. two step dist: {delta_2step:0.3f}, 1step: {delta_1step:0.3f}, mag std: {mag_std}. top updates: {ups}{spaces}\n"
)
sys.stdout.flush()
return dist <= self._tol * len(params) or (
delta_2step < delta_1step / 2 and delta_2step > 0)
def from_history(self, history):
from web.util import nanguard
from web import opts
keep = []
wts = False
for idx, item in enumerate(history):
if not "items" in item:
continue
if "parent" in item:
path = "/" + item["parent"] + "/" + item["name"]
if not (item["hash"] == hashlib.sha256(path.encode(
"utf-8", ue)).hexdigest()):
raise Exception()
item["dur"] = nanguard(item.get(
"viewend", 0), f"from_history#{idx}.viewend") - nanguard(
item.get("viewstart", 0), f"from_history#{idx}.viewstart")
if nanguard(item.get("dur", 0)) < opts.minview:
if not wts and keep:
keep.pop()
wts = True
continue
wts = False
keep.append(item)
for item in keep:
self.update(item, initial=True)
def calc_next_index(self,
h,
vals,
weights,
dists_out,
debug=False,
force=False,
existing_val=None):
from web.util import nanguard
from web import opts
modelmin = self.min()
modelmax = self.max()
low_, high_ = self.softmin(vals[0], inv=True), self.softmin(vals[1])
low = max(low_, -modelmax)
high = min(high_, modelmax)
dists_out[h] = high - low
widx, lidx = self.getidx(low), self.getidx(high)
mid = existing_val
if mid is None:
mid = (low + high) / 2
midx = self.getidx(mid)
pos = nanguard((high - modelmin) / (modelmax - modelmin))
prec = opts.precision_func(pos)
delta = nanguard(len(self.model) / prec) * 2
#if low < modelmin: widx -= max((delta - 2), 1)
#if high > modelmax: lidx += max((delta - 2), 1)
#if debug: print(pos, delta, low, high, widx, lidx)
seen_enough = len(vals[0]) > opts.min_clean_wins or len(vals[0]) + len(
vals[1]) > opts.min_clean_compares
finished_enough = max(0, lidx - widx) < delta
is_goat = widx >= nanguard(len(self.model) - opts.goat_window)
info = {
"finished_enough": finished_enough,
"seen_enough": seen_enough,
"is_goat": is_goat,
"wlen": len(vals[0]),
"llen": len(vals[1]),
"prec": prec,
"midx": midx,
"lidx": lidx,
"widx": widx,
"mid": mid,
"pos": pos,
"delta": delta,
"adelta": max(0, lidx - widx),
"high": high,
"low": low,
}
if len(vals[0]) > 0 and (is_goat or len(
vals[1]) > 1) and seen_enough and finished_enough and not force:
return None, info
return midx, info
def calculate_dists(self, stats, comparisons, h):
from web.util import timing, as_pair, nanguard
from web import opts
dists = ([-50], [50])
weights = ([1], [1])
for other_hash, wins in comparisons.items():
other_val = self.getval(other_hash)
if other_val is None:
continue
if self.is_dropped(stats, other_hash):
continue
pair = as_pair(h, other_hash)
if pair in stats.too_close:
wins = tuple([
nanguard(x + sum(wins) +
opts.too_close_boost * stats.too_close[pair])
for x in wins
])
win_ratio = (wins[0]) / (wins[0] + wins[1])
if win_ratio < opts.ambiguity_threshold and win_ratio > 1 - opts.ambiguity_threshold:
continue
if wins[0] > wins[1]:
decayed_ratio = (
(wins[0] / max(1e-10, wins[0] + wins[1])) - 0.5) * 2
dists[0].append(other_val)
weights[0].append(decayed_ratio)
elif wins[0] < wins[1]:
decayed_ratio = ((
(wins[1]) / max(1e-10, wins[0] + wins[1])) - 0.5) * 2
dists[1].append(other_val)
weights[1].append(decayed_ratio)
else:
continue
return dists, weights
def softmin(self, directional_distances, inv=False):
from web.util import nanguard
from web import opts
if not len(directional_distances):
return 100
vals = nanguard(
numpy.array(directional_distances)
) #numpy.log2(numpy.maximum(directional_distances*10, 0.0001))
if inv: vals = -vals
weight = nanguard(
numpy.minimum(opts.softmin_falloff_per_unit**-vals, 2000000))
sum = numpy.sum(weight * vals)
total_weight = numpy.sum(weight)
res = nanguard(sum / total_weight)
if inv: res = -res
return res
def getprob(self, item1, item2):
import choix
from web.util import nanguard
a = self.getid(item1["hash"] if type(item1) == dict else item1)
b = self.getid(item2["hash"] if type(item2) == dict else item2)
a_new = False
b_new = False
if a >= len(self.model):
a_new = True
if b >= len(self.model):
b_new = True
#self.calculate_ranking()
if not len(self.model) or a_new or b_new:
return 0.5, 0.5
#return f"no model yet. new: {a_new}", f"no model yet. new: {b_new}"
ra, rb = choix.probabilities([a, b], self.model)
return nanguard(ra, "ra"), nanguard(rb, "rb")
def prepare_pairs(self, stats):
from web.util import nanguard
from web import opts
pairs = []
#removeme = []
#for x in self.all_items:
# if self.is_dropped(stats, x)>1 and x in self.ids:
# removeme.append(self.ids[x])
#if removeme:
# model = list(self.model)
# all_items = list(self.all_items)
# assert type(self.all_items) == list
# for idx in sorted(removeme)[::-1]:
# del all_items[idx]
# del model[idx]
# self.model=numpy.array(model)
# self.all_items =numpy.array(all_items)
#self.ids = {x: idx for idx, x in enumerate(self.all_items)}
for pair, rel_wins in stats.pair_wins.items():
if pair in stats.incomparable_pairs: continue
if (self.is_dropped(stats, pair[0]) > 1) or (self.is_dropped(
stats, pair[1]) > 1):
continue
if pair[0] not in self.ids or pair[1] not in self.ids:
continue
if pair in stats.too_close:
rel_wins = tuple([
nanguard(x + sum(rel_wins) +
opts.too_close_boost * stats.too_close[pair])
for x in rel_wins
])
if not sum(rel_wins):
continue
ratio = nanguard((rel_wins[0]) / (rel_wins[0] + rel_wins[1]))
scale = 1 #sigmoid(3*(sum(rel_wins)-1))
rel_wins = [scale * ratio, scale * (1 - ratio)]
id0, id1 = self.ids[pair[0]], self.ids[pair[1]]
if rel_wins[0]:
pairs.append((id0, id1, nanguard(rel_wins[0])))
if rel_wins[1]:
pairs.append((id1, id0, nanguard(rel_wins[1])))
return pairs
def build_sim_data(self, stats):
from web.util import nanguard
triplets = [(tuple(self.getid_sim(x) for x in key), vals)
for key, vals in stats.triplet_diffs.items()
if not any(self.is_dropped(stats, x) for x in key)]
edges = list(
itertools.chain.from_iterable(
itertools.combinations(key, 2) for key, val in triplets))
edge_ratios = list(
itertools.chain.from_iterable(
(((edge1[0], edge2[0]), nanguard(edge1[1] / (edge1[1] + edge2[1])),
nanguard(edge1[1] + edge2[1]))
for edge1, edge2 in itertools.combinations((
((vert1[0], vert2[0]), nanguard(vert1[1] + vert2[1]), vert1,
vert2)
for vert1, vert2 in itertools.combinations(zip(key, val), 2)
), 2)) for key, val in triplets))
targ = [nanguard(x[1]) for x in edge_ratios]
return edges, edge_ratios, targ
def extend_model(self, stats):
from web.util import nanguard
for pair, rel_wins in stats.pair_wins.items():
self.getid(pair[0])
self.getid(pair[1])
if len(self.model) and len(self.all_items) > len(self.model):
newlen = len(self.all_items) - len(self.model)
#gp = choix.generate_params(newlen, 0.1)
#print(gp.shape)
newvals = []
for h in self.all_items[len(self.model):]:
dists, weights = self.calculate_dists(stats,
stats.comparisons.get(h, {}),
h)
info = self.calc_next_index(h, dists, weights, {})[1]
newvals.append(nanguard(info["mid"]))
if len(newvals) != newlen:
raise Exception()
self.model = numpy.concatenate((self.model, newvals))
def calculate_nearest_neighborhood(self,
stats,
hashes_to_debug,
extra=False,
save=True):
from web import opts
from web.util import timing
with timing("calculate_nearest_neighborhood", 0.1):
distances = {h: ([-50], [50]) for h in self.all_items}
weights = {h: ([1], [1]) for h in self.all_items}
sp = {}
if save: self.searching_pool = sp
iv = {}
if save: self.inversions = iv
inversions = {}
inversion_pool = set()
dists = {}
if save:
self.distances = dists
for pair, rel_wins in stats.pair_wins.items():
if self.is_dropped(stats, pair[0]) or self.is_dropped(
stats, pair[1]):
continue
if pair in stats.incomparable_pairs: continue
if pair in stats.too_close:
rel_wins = tuple([
nanguard(x + sum(rel_wins) +
opts.too_close_boost * stats.too_close[pair])
for x in rel_wins
])
win_ratio = (rel_wins[0]) / (rel_wins[0] + rel_wins[1])
win, loss = pair[0], pair[1]
count = rel_wins[0] + rel_wins[1]
if win_ratio < 0.5:
rel_wins = rel_wins[::-1]
win, loss = pair[1], pair[0]
win_ratio = 1 - win_ratio
win_prob, inverted, details = self.check_inversion(stats, pair)
if win_prob < 0.5:
win_inversions, _ = inversions.setdefault(win, ([], []))
_, loss_inversions = inversions.setdefault(loss, ([], []))
win_inversions.append((pair, win_ratio, win_prob, count))
loss_inversions.append((pair, win_ratio, win_prob, count))
if inverted:
inversion_pool.add(pair[0])
inversion_pool.add(pair[1])
iv[pair] = inverted
decayed_ratio = (((rel_wins[0] + 1) /
(rel_wins[0] + rel_wins[1] + 2)) - 0.5) * 2
if win_ratio < opts.ambiguity_threshold:
# don't include ambiguous comparisons when tallying distances
# should reduce risk of getting in tangles
continue
wval = self.getval(win)
lval = self.getval(loss)
if wval is None or lval is None: continue
dist = wval - lval
distances[win][0].append(lval)
distances[loss][1].append(wval)
weights[win][0].append(decayed_ratio)
weights[loss][1].append(decayed_ratio)
modelmin = self.min()
modelmax = self.max()
for h, vals in distances.items():
if self.is_dropped(stats, h): continue
nextidx, _ = self.calc_next_index(h,
vals,
weights.get(h),
dists,
existing_val=self.getval(h))
if (nextidx is not None
or not stats.win_counts.get(h)) and h in self.bh2:
sp[h] = True
print(
"len(sp)",
len(sp),
"len(distances)",
len(distances),
len(self.bh2),
len([x for x in self.all_items if not stats.win_counts.get(x)]),
len([
x for x in self.all_items if not stats.win_counts.get(x)
and not self.is_dropped(stats, x)
]),
)
#q = []
if True:
for h in hashes_to_debug:
hidx = self.sorted_ids.get(h, -1)
in_pool = h in sp
in_inv_pool = h in inversion_pool
#if not in_pool or in_inv_pool: continue
dists = distances[h]
win_inversions, loss_inversions = inversions.get(h, ([], []))
wdist, ldist = self.weighted_softmin(
*dists[0]), self.weighted_softmin(*dists[1])
val = self.getval(h)
pos = (val - modelmin) / (modelmax - modelmin)
delta = int(
len(self.model) / (opts.min_target_precision +
(pos**opts.target_precision_curve) *
opts.target_precision_top))
lc = stats.loss_counts.get(h, 0)
wc = stats.win_counts.get(h, 0)
##ld = "done" if ld is None else f"{ld:4d}"
##wd = "done" if wd is None else f"{wd:4d}"
#if not wc or not lc or ld or wd: continue
##print(f"{wd} <= {wc:4d} {lc:4d} => {ld}")
rows = []
low = val - wdist
high = val + ldist
widx2, lidx2 = self.getidx(low), self.getidx(high)
#if low < modelmin: widx2 -= max((delta - 2), 1)
#if high > modelmax: lidx2 += max((delta - 2), 1)
widx, lidx = max(0, self.getidx(val - wdist)), min(
len(self.model) - 1, self.getidx(val + ldist))
midx = (widx + lidx) // 2
mval = self.sorted_model[midx]
mdist = mval - val
low = max(low, -modelmax)
high = min(high, modelmax)
vval = (low + high) / 2
vdist = vval - val
vidx = self.getidx(vval)
waidx = max(0, hidx - 1)
laidx = min(len(self.model) - 1, hidx + 1)
la_h = self.sorted_hashes[laidx]
wa_h = self.sorted_hashes[waidx]
la_val = self.getval(la_h)
wa_val = self.getval(wa_h)
wtidx = hidx - delta
ltidx = hidx + delta
wthresh_h = self.sorted_hashes[max(0, wtidx)]
lthresh_h = self.sorted_hashes[min(ltidx,
len(self.sorted_hashes) -
1)]
wthresh_val = self.getval(wthresh_h)
lthresh_val = self.getval(lthresh_h)
print()
print(f"wc={wc:2d}")
print(f"lc={lc:2d}")
print(f"lc+wc={lc+wc:2d}")
print(f"in_pool={in_pool}")
print(f"in_inv_pool={in_inv_pool}")
for pair, win_ratio, win_prob, count in win_inversions:
other = [x for x in pair if x != h][0]
other_idx = self.sorted_ids[other]
other_val = self.getval(other)
rows.append((
other_val - val, 9, "iw", other_val, other_idx,
other_idx - hidx,
f"unexpected win; win ratio: {win_ratio} ({count} samples), expected win prob: {win_prob}"
))
for pair, loss_ratio, loss_prob, count in loss_inversions:
other = [x for x in pair if x != h][0]
other_idx = self.sorted_ids[other]
other_val = self.getval(other)
rows.append((
other_val - val, 3, "il", other_val, other_idx,
other_idx - hidx,
f"unexpected loss; loss ratio: {loss_ratio} ({count} samples), expected loss prob: {loss_prob}"
))
for windist in dists[0]:
if windist == 9 or windist < 0: continue
rows.append(
(-windist, 0, "win", val - windist,
self.getidx(val - windist),
self.getidx(val - windist) - hidx, f"expected win"))
for lossdist in dists[1]:
if lossdist == 9 or lossdist < 0: continue
rows.append(
(lossdist, 12, "los", val + lossdist,
self.getidx(val + lossdist),
self.getidx(val + lossdist) - hidx, f"expected loss"))
rows.append((modelmin - val, 1, "W", modelmin, 0, 0 - hidx,
"model boundary low"))
rows.append((-wdist, 1, "w", val - wdist, widx2, widx2 - hidx,
"win boundary"))
rows.append(
(wthresh_val - val, 2, "wt", wthresh_val, wtidx,
wtidx - hidx, "search precision threshold, win side"))
rows.append((vdist, 4, "v", vval, vidx, vidx - hidx,
"midpoint in value space"))
rows.append((mdist, 5, "m", mval, midx, midx - hidx,
"midpoint in index space"))
rows.append((wa_val - val, 6, "wa", wa_val, waidx,
waidx - hidx, "prev neighbor"))
rows.append((0, 7, "", val, hidx, 0, "item"))
rows.append((la_val - val, 8, "la", la_val, laidx,
laidx - hidx, "next neighbor"))
rows.append(
(lthresh_val - val, 10, "lt", lthresh_val, ltidx,
ltidx - hidx, "search precision threshold, loss side"))
rows.append((ldist, 11, "l", val + ldist, lidx2, lidx2 - hidx,
"loss boundary"))
rows.append(
(modelmax - val, 11, "L", modelmax, len(self.model) - 1,
(len(self.model) - 1) - hidx, "model boundary high"))
maxstep = 12
for dist, step, label, val, idx, idxdist, desc in sorted(
rows, key=lambda x: (x[4], x[0])):
prefix = label.rjust(step * 2).ljust(24)
label = label.rjust(3)
print(
f" {prefix} | {label}dist={dist:7.4f} {label}val={val:7.4f} {label}idx={idx:5d} {label}idxdist={idxdist:5d} {desc}"
)
#print(sorted(dists[0]))
#print(f" wdist={-wdist:7.4f} wval={val-wdist:7.4f} widx={widx:5d} widxdist={self.getidx(val-wdist)-hidx:5d}")
#print(f"wthresh={wthresh_val-val:7.4f} wtval={wthresh_val:7.4f} wtidx={wtidx:5d} wtidxdist={wtidx-hidx:5d}")
#print(f" vdist={vdist:7.4f} vval={vval:7.4f} vidx={vidx:5d} vidxdist={vidx-hidx:5d}")
#print(f" mdist={mdist:7.4f} mval={mval:7.4f} midx={midx:5d} midxdist={midx-hidx:5d}")
#print(f" wadist={wa_val-val:7.4f} waval={wa_val:7.4f} waidx={waidx:5d} waidxdist={waidx-hidx:5d}")
#print(f" val={val:7.4f} hidx={hidx:5d}")
#print(f" ladist={la_val-val:7.4f} laval={la_val:7.4f} laidx={laidx:5d} laidxdist={laidx-hidx:5d}")
#print(f"lthresh={lthresh_val-val:7.4f} ltval={lthresh_val:7.4f} ltidx={ltidx:5d} ltidxdist={ltidx-hidx:5d}")
#print(f" ldist={ldist:7.4f} lval={val+ldist:7.4f} lidx={lidx:5d} lidxdist={self.getidx(val+ldist)-hidx:5d}")
#print(sorted(dists[1]))
#print(f"min(model): {min(self.model)}, max(model): {max(self.model)}, mean(model): {numpy.mean(self.model)}")
print(f"searching_pool: {len(sp)}, inversions: {len(iv)}")
def getval(self, h):
from web.util import nanguard
id = self.getid(h)
if id >= len(self.model):
return None
return nanguard(self.model[id])
def getidx(self, val):
from web.util import nanguard
return nanguard(numpy.searchsorted(self.sorted_model, val))
def update(self, item, initial=False):
from web.util import nanguard, as_pair, sigmoid
from web import opts
age = time.time() - (nanguard(item.get("viewend", 0), "update.viewend") /
1000)
decay = nanguard(opts.comparison_decay_func(age))
item["dur"] = nanguard(
item.get("viewend", 0) - item.get("viewstart", 0), "update.dur")
mag_decay = nanguard(
max(
min(opts.initial_mag,
opts.initial_mag / max(1, (item["dur"] / 1000))),
opts.min_mag))
if nanguard(item.get("dur", 0)) < opts.minview and not item.get("fast"):
print("\033[31mskipped due to too-low view duration", item, "\033[m")
return
if not initial:
print(
f"\033[38mvd: {item.get('dur',0)}, mag_decay: {mag_decay}, age: {age}, time_decay: {decay}\033[m"
)
pair = as_pair(*item["items"])
if "similarity" in item:
info = item["similarity"]
winner = None
least_similar = info.get("least_similar", None)
most_similar = info.get("most_similar", None)
if type(most_similar) == list:
least_similar = list(set(range(3)) - set(most_similar))[0]
most_similar = None
elif type(item.get("preference", None)) != dict:
info = {}
winner = nanguard(item.get("preference", 1) - 1)
least_similar = None
else:
info = item.get("preference", {})
winner = nanguard(info.get("prefer", 1) - 1)
least_similar = None
too_close = info.get("too_close", False)
incomparable = info.get("incomparable", False)
dislike = info.get("dislike", None)
strong = info.get("strong", None)
if type(item.get("info")) == dict and item["info"].get("t") == [
"inversions", "inversions"
]:
mag_decay = mag_decay * opts.inversion_compare_boost + sum(
self.pair_wins.get(
pair, [0, 0])) * opts.inversion_compare_relboost * mag_decay
if not dislike:
dislike = [0] * len(item["items"])
for f, dis in zip(item["items"], dislike):
if dis:
self.dislike[f["hash"]] = True
#print("dislike",f)
elif f["hash"] in self.dislike:
del self.dislike[f["hash"]]
#print("undislike",f)
if any(dislike): return
if too_close:
self.too_close[pair] = self.too_close.get(pair, 0) + nanguard(
2 * decay * (1 - sigmoid(mag_decay)))
#self.record_win(*item["items"])
#self.record_win(*item["items"][::-1])
elif incomparable:
pair = as_pair(*item["items"], strip=True)
self.incomparable_pairs[pair] = self.incomparable_pairs.get(
pair, 0) + nanguard(decay / mag_decay)
elif winner is not None:
winning = item["items"][winner]
losing = item["items"][1 - winner]
self.record_win(winning, losing, nanguard(decay), nanguard(mag_decay))
elif least_similar is not None or most_similar is not None:
sim = least_similar if least_similar is not None else most_similar
assert 0 <= sim <= 2
s1, s2 = [x for i, x in enumerate(item["items"]) if i != sim]
s3 = item["items"][sim]
self.record_similar(s1,
s2,
s3,
nanguard(decay),
nanguard(mag_decay),
invert=(most_similar is not None))