- # Put a lookahead reward to +1, sample the action and reward
- result[:, u - 1] = (1) + 1 + escape.first_lookahead_rewards_code
- ar_mask = (t >= state_len).long() * (t < state_len + 2).long()
- ar_mask = ar_mask[None, :]
- ar_mask = ar_mask.expand_as(result)
- result *= 1 - ar_mask
- ar()
+ optimistic_bias = result.new_zeros(escape.nb_codes, device=result.device)
+ optimistic_bias[escape.lookahead_reward2code(-1)] = -math.log(1e1)
+ optimistic_bias[escape.lookahead_reward2code(1)] = math.log(1e1)
+
+ for u in tqdm.tqdm(
+ range(self.it_len, result.size(1) - self.it_len + 1, self.it_len),
+ desc="thinking",
+ ):
+ # Generate the lookahead_reward and state
+ ar_mask = (t % self.it_len == self.index_lookahead_reward).long() * (
+ t <= u + self.index_lookahead_reward
+ ).long()
+ ar(result, ar_mask)
+
+ # Generate the lookahead_reward and state
+ ar_mask = (t >= u + self.index_states).long() * (
+ t < u + self.index_states + self.state_len
+ ).long()
+ ar(result, ar_mask)
+
+ # Re-generate the lookahead_reward
+ ar_mask = (t % self.it_len == self.index_lookahead_reward).long() * (
+ t <= u + self.index_lookahead_reward
+ ).long()
+ ar(result, ar_mask, logit_biases=optimistic_bias)
+
+ # Generate the action and reward
+ ar_mask = (t >= u + self.index_action).long() * (
+ t <= u + self.index_reward
+ ).long()
+ ar(result, ar_mask)
+
+ filename = os.path.join(result_dir, f"test_thinking_compute_{n_epoch:04d}.txt")
+ with open(filename, "w") as f:
+ for n in range(10):
+ for s in snapshots:
+ lr, s, a, r = escape.seq2episodes(
+ s[n : n + 1], self.height, self.width
+ )
+ str = escape.episodes2str(
+ lr, s, a, r, unicode=True, ansi_colors=True
+ )
+ f.write(str)
+ f.write("\n\n")