+ ar(result, ar_mask)
+
+ # Generate the lookahead_reward and state
+ ar_mask = (t >= u + index_states).long() * (
+ t < u + index_states + state_len
+ ).long()
+ ar(result, ar_mask)
+
+ # Re-generate the lookahead_reward
+ ar_mask = (t % it_len == index_lookahead_reward).long() * (
+ t <= u + index_lookahead_reward
+ ).long()
+ ar(result, ar_mask, logit_biases=optimistic_bias)
+
+ # Generate the action and reward
+ ar_mask = (t >= u + index_action).long() * (t <= u + 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")