+
+import greed
+
+
+class Greed(Task):
+ def __init__(
+ self,
+ nb_train_samples,
+ nb_test_samples,
+ batch_size,
+ height,
+ width,
+ T,
+ nb_walls,
+ nb_coins,
+ logger=None,
+ device=torch.device("cpu"),
+ ):
+ super().__init__()
+
+ self.batch_size = batch_size
+ self.device = device
+
+ self.world = greed.GreedWorld(height, width, T, nb_walls, nb_coins)
+
+ states, actions, rewards = self.world.generate_episodes(
+ nb_train_samples + nb_test_samples
+ )
+ seq = self.world.episodes2seq(states, actions, rewards)
+ self.train_input = seq[:nb_train_samples].to(self.device)
+ self.test_input = seq[nb_train_samples:].to(self.device)
+
+ def wipe_lookahead_rewards(self, batch):
+ t = torch.arange(batch.size(1), device=batch.device)[None, :]
+ u = torch.randint(batch.size(1), (batch.size(0), 1), device=batch.device)
+ lr_mask = (t <= u).long() * (
+ t % self.world.it_len == self.world.index_lookahead_reward
+ ).long()
+
+ return (
+ lr_mask * self.world.lookahead_reward2code(greed.REWARD_UNKNOWN)
+ + (1 - lr_mask) * batch
+ )
+
+ def batches(self, split="train", nb_to_use=-1, desc=None):
+ assert split in {"train", "test"}
+ input = self.train_input if split == "train" else self.test_input
+ if nb_to_use > 0:
+ input = input[:nb_to_use]
+ if desc is None:
+ desc = f"epoch-{split}"
+ for batch in tqdm.tqdm(
+ input.split(self.batch_size), dynamic_ncols=True, desc=desc
+ ):
+ yield self.wipe_lookahead_rewards(batch)
+
+ def vocabulary_size(self):
+ return self.world.nb_codes
+
+ def thinking_autoregression(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
+ ):
+ snapshots = []
+
+ 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,
+ )
+ warnings.warn("keeping thinking snapshots", RuntimeWarning)
+ snapshots.append(result[:100].detach().clone())
+
+ # Generate iteration after iteration
+
+ 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(
+ greed.REWARD_UNKNOWN
+ )
+
+ 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(greed.REWARD_UNKNOWN)
+ 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(greed.REWARD_PLUS)
+ 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(greed.REWARD_UNKNOWN)
+
+ filename = os.path.join(result_dir, f"test_thinking_compute_{n_epoch:04d}.txt")
+ with open(filename, "w") as f:
+ for n in range(snapshots[0].size(0)):
+ 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")
+
+ # Saving the generated sequences
+
+ lr, s, a, r = self.world.seq2episodes(result)
+ str = self.world.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}")
+
+ def produce_results(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
+ ):
+ result = self.wipe_lookahead_rewards(self.test_input[:250].clone())
+
+ # Saving the ground truth
+
+ lr, s, a, r = self.world.seq2episodes(
+ result,
+ )
+ str = self.world.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
+
+ filename = os.path.join(result_dir, f"test_true_seq_{n_epoch:04d}.txt")
+ with open(filename, "w") as f:
+ f.write(str)
+ logger(f"wrote {filename}")
+
+ # Re-generating from the first frame
+
+ ar_mask = (
+ torch.arange(result.size(1), device=result.device) >= self.world.it_len
+ ).long()[None, :]
+ ar_mask = ar_mask.expand_as(result)
+ result *= 1 - ar_mask # paraaaaanoiaaaaaaa
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ device=self.device,
+ )
+
+ # Saving the generated sequences
+
+ lr, s, a, r = self.world.seq2episodes(
+ result,
+ )
+ str = self.world.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
+
+ filename = os.path.join(result_dir, f"test_seq_{n_epoch:04d}.txt")
+ with open(filename, "w") as f:
+ f.write(str)
+ logger(f"wrote {filename}")
+
+ self.thinking_autoregression(
+ n_epoch, model, result_dir, logger, deterministic_synthesis, nmax
+ )
+
+
+######################################################################