def test_parse_valid_actions_singles(self):
actions = parser.parse_valid_actions(single())
all_actions = parser.all_actions_singles()
self.assertEqual(len(all_actions), len(actions))
for expected, actual in zip(all_actions, actions):
if is_normal_move(expected):
self.assertIsNot(actual, None)
else:
self.assertIs(actual, None)
def test_parse_valid_actions_mega(self):
actions = parser.parse_valid_actions(single_with_mega())
all_actions = parser.all_actions_singles()
self.assertEqual(len(all_actions), len(actions))
for expected, actual in zip(all_actions, actions):
if is_normal_move(expected) or is_mega_move(expected) or is_switch(
expected, exclude=[1, 5]):
self.assertIsNot(actual, None)
else:
self.assertIs(actual, None)
def extract(self, state, candidates):
rv = extract(state, self.poke_features)
candidates = list(candidates)
while len(candidates) < len(parser.all_actions_singles()):
candidates.append(None)
rv['mask'] = np.asarray([float(bool(c))
for c in candidates]).astype(config.nt())
return rv
def _action_extra(self, n):
actions = parser.all_actions_singles()
mega_mask = ag.Variable(
torch.ByteTensor([a.endswith(' mega') for a in actions
]).type(config.tt()).unsqueeze(0).unsqueeze(-1))
zmove_mask = ag.Variable(
torch.ByteTensor([a.endswith(' zmove') for a in actions
]).type(config.tt()).unsqueeze(0).unsqueeze(-1))
ultra_mask = ag.Variable(
torch.ByteTensor([a.endswith(' ultra') for a in actions
]).type(config.tt()).unsqueeze(0).unsqueeze(-1))
return (self.policy_mega(mega_mask.expand(n, -1, -1)) +
self.policy_zmove(zmove_mask.expand(n, -1, -1)) +
self.policy_ultra(ultra_mask.expand(n, -1, -1)))
result = result.get()
probs = result['probs']
probs = (1. - self._epsilon) * probs + (
self._epsilon * mask / sum(mask)).astype(config.nt())
if self._play_best_move:
action = np.argmax(probs)
else:
action = np.random.choice(len(self.candidates), p=probs)
# TODO: why not just use self.candidates?
if self.request.get('teamPreview'):
all_actions = _teampreview_actions
else:
all_actions = _singles_actions
action_string = all_actions[action]
result['candidates'] = self.candidates
result['state'] = state
result['action'] = action
result['actionString'] = action_string
result['_updates'] = block_updates
self.blocks.append(result)
rv.set(action_string)
gevent.spawn(fn)
return rv
_engine = Engine()
_singles_actions = parser.all_actions_singles()
_teampreview_actions = parser.team_preview_actions_singles()