def test_train(self):
q = MemoryOnlyHashQ('test', 3)
algo = AlgoQLearning(q)
# train on 10 games
for _ in range(100):
ep = Episode(algo, PolicyRandom())
ep.run(StateTest(random.choice([True, False])))
# check if policy is optimal
policy = PolicyExploit(q)
# print(q.get_all(StateTest().inputs()))
for cells, optimal_action in [
([0, 0, 0], 1),
([2, 0, 0], 1),
([0, 0, 2], 1),
]:
state = StateTest()
state.cells = cells
self.assertEqual(policy.play(state)[0], optimal_action, state)
def episodes_from_fns(fns, limit=None, split='dev'):
use_scripts = (scripted_tell in fns) or (scripted_tell_before_peek in fns)
if scripted_tell_after_peek in fns:
use_scripts = True
run_from = codraw_data.get_scenes_and_scripts_with_peek(split)
elif use_scripts:
run_from = codraw_data.get_scenes_and_scripts(split)
else:
run_from = codraw_data.get_scenes(split)
if limit is not None:
run_from = run_from[:limit]
sims = []
with torch.no_grad():
for run_from_single in run_from:
if use_scripts:
episode = Episode.run_script(run_from_single, fns)
else:
episode = Episode.run(run_from_single, fns)
yield episode
def collect_episodes(fns,
dg,
scenes=codraw_data.get_scenes('dev'),
batch_size=16,
utterance_penalty=0.25,
gamma=0.99,
uninformative_penalty=0.3):
with torch.no_grad():
episodes = []
for scene in np.random.choice(scenes, batch_size):
ep = Episode.run(scene, fns)
episodes.append(ep)
example_batch = examples_from_episodes(
episodes,
dg=dg,
utterance_penalty=utterance_penalty,
gamma=gamma,
uninformative_penalty=uninformative_penalty,
)
return episodes, example_batch
