def add(self, data):
if data.status == Status.INTERESTING:
return
if data.status < self.best_status:
return
if data.status > self.best_status:
self.best_status = data.status
self.reset()
self.fresh_examples.append(data)
if len(self) > self.pool_size:
pop_random(self.random, self.used_examples or self.fresh_examples)
assert self.pool_size == len(self)
def add(self, data):
if data.status == Status.INTERESTING:
return
if data.status < self.best_status:
return
if data.status > self.best_status:
self.best_status = data.status
self.reset()
self.fresh_examples.append(data)
if len(self) > self.pool_size:
pop_random(self.random, self.used_examples or self.fresh_examples)
assert self.pool_size == len(self)
def maybe_discard_one(self):
if len(self) > self.pool_size:
if self.used_examples or self.fresh_examples:
pop_random(self.random, self.used_examples
or self.fresh_examples)
else:
# Get an arbitrary label from those with the longest pools,
# discard the lowest-scored example from that pool,
# and discard the label if it had no other examples.
label, pool = max(self.scored_examples.items(),
key=lambda kv: len(kv[1]))
pool.pop()
if not pool:
self.scored_examples.pop(label)
assert len(self) == self.pool_size
def select(self):
if self.fresh_examples:
result = pop_random(self.random, self.fresh_examples)
self.used_examples.append(result)
return result
else:
return self.random.choice(self.used_examples)
def select(self):
if self.fresh_examples:
result = pop_random(self.random, self.fresh_examples)
self.used_examples.append(result)
return result
else:
return self.random.choice(self.used_examples)
def select(self):
# If we have feedback from targeted PBT, choose a label and then an example
# from that pool. Each example is twice as likely as the next-highest-ranked.
if self.scored_examples: # pragma: no cover
# For some reason this clause is often showing up as uncovered,
# though tests/cover/test_targeting.py definitely executes it :-/
pool = self.random.choice(list(self.scored_examples.values()))
stop_at = self.random.random() * self.weights[len(pool) - 1]
return pool[bisect(self.weights, stop_at)]
# Otherwise, prefer first-time mutations to previously-mutated examples
if self.fresh_examples:
result = pop_random(self.random, self.fresh_examples)
self.used_examples.append(result)
return result
else:
return self.random.choice(self.used_examples)
def selection_order(depth, n):
pending = LazySequenceCopy(range(n))
while pending:
yield pop_random(random, pending)
