+ return escape.nb_codes
+
+ def thinking_autoregression(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
+ ):
+ result = self.test_input[:250].clone()
+ t = torch.arange(result.size(1), device=result.device)[None, :]
+
+ state_len = self.height * self.width
+ index_action = state_len
+ index_reward = state_len + 1
+ index_lookahead_reward = state_len + 2
+ it_len = state_len + 3 # state / action / reward / lookahead_reward
+
+ result[:, it_len:] = -1
+
+ def ar(result, ar_mask, logit_biases=None):
+ ar_mask = ar_mask.expand_as(result)
+ result *= 1 - ar_mask
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis=deterministic_synthesis,
+ logit_biases=logit_biases,
+ device=self.device,
+ progress_bar_desc=None,
+ )
+
+ # Generate iteration after iteration
+
+ 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)
+
+ snapshots = []
+
+ for u in tqdm.tqdm(
+ range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
+ ):
+ # Re-generate the lookahead_reward pessimistically in the
+ # previous iterations
+ ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
+ ar(result, ar_mask, logit_biases=-optimistic_bias)
+ snapshots.append(result[:10].detach().clone())
+
+ # Generate the state
+ ar_mask = (t >= u).long() * (t < u + state_len).long()
+ ar(result, ar_mask)
+ snapshots.append(result[:10].detach().clone())
+
+ # Re-generate the lookahead_reward optimistically in the
+ # previous iterations
+ ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
+ ar(result, ar_mask, logit_biases=optimistic_bias)
+ snapshots.append(result[:10].detach().clone())
+
+ # Generate the action and reward
+ ar_mask = (t >= u + index_action).long() * (t <= u + index_reward).long()
+ ar(result, ar_mask)
+ snapshots.append(result[:10].detach().clone())
+
+ 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, lookahead=True
+ )
+ str = escape.episodes2str(
+ lr, s, a, r, unicode=True, ansi_colors=True
+ )
+ f.write(str)
+ f.write("\n\n")
+
+ # Saving the generated sequences
+
+ s, a, r, lr = escape.seq2episodes(result, self.height, self.width)
+ str = escape.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
+
+ filename = os.path.join(result_dir, f"test_thinking_seq_{n_epoch:04d}.txt")
+ with open(filename, "w") as f:
+ f.write(str)
+ logger(f"wrote {filename}")