- # 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()
+ result = self.test_input[:250].clone()
+ # Erase all the content but that of the first iteration
+ result[:, self.world.it_len :] = -1
+ # Set the lookahead_reward of the firs to UNKNOWN
+ result[:, self.world.index_lookahead_reward] = self.world.lookahead_reward2code(
+ 2
+ )
+
+ t = torch.arange(result.size(1), device=result.device)[None, :]
+
+ for u in tqdm.tqdm(
+ range(0, result.size(1), self.world.it_len),
+ desc="thinking",
+ ):
+ # Generate the next state but keep the initial one, the
+ # lookahead_reward of previous iterations are set to
+ # UNKNOWN
+ if u > 0:
+ result[
+ :, u + self.world.index_lookahead_reward
+ ] = self.world.lookahead_reward2code(2)
+ ar_mask = (t >= u + self.world.index_states).long() * (
+ t < u + self.world.index_states + self.world.state_len
+ ).long()
+ ar(result, ar_mask)
+
+ # Generate the action and reward with lookahead_reward to +1
+ result[
+ :, u + self.world.index_lookahead_reward
+ ] = self.world.lookahead_reward2code(1)
+ ar_mask = (t >= u + self.world.index_reward).long() * (
+ t <= u + self.world.index_action
+ ).long()
+ ar(result, ar_mask)
+
+ # Set the lookahead_reward to UNKNOWN for the next iterations
+ result[
+ :, u + self.world.index_lookahead_reward
+ ] = self.world.lookahead_reward2code(2)
+
+ 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 = self.world.seq2episodes(
+ s[n : n + 1],
+ )
+ str = self.world.episodes2str(
+ lr, s, a, r, unicode=True, ansi_colors=True
+ )
+ f.write(str)
+ f.write("\n\n")