def acts(self, states):
obs = torch.cat(
tuple(Translator.encode_board(s).unsqueeze(0) for s in states), 0)
valids = [
torch.tensor(list(
map(Translator.encode_move_idx, s.micro_legal_moves)),
dtype=torch.long) for s in states
]
policy = self.net.forward(obs)[0]
prob = [policy[i, valids[i]] for i in range(len(states))]
prob = [prob[i] / prob[i].sum() for i in range(len(states))]
idxs = [
valids[i][dist.Categorical(prob[i]).sample()]
for i in range(len(states))
]
return [
Translator.decode_move(idxs[i].item(), states[i])
for i in range(len(states))
]
def extract_episodes(self):
collected = collect_episodes(
ReinforcementAI(self.judge), episode_cnt, episode_length)
episodes = []
for c in collected:
obs = torch.cat(
tuple(Translator.encode_board(b).unsqueeze(0)
for b in c['boards']), 0)
idx = torch.tensor(
tuple(Translator.encode_move_idx(m) for m in c['moves']),
dtype=torch.long)
prob = self.judge.forward(obs)[0][
torch.arange(idx.size()[0], dtype=torch.long), idx]
ext = c['extrinsic']
episodes += [
{'obs': obs, 'idx': idx, 'prob': prob, 'ext': ext,
'over': c['boards'][-1].is_game_over()}]
return episodes