+######################################################################
+
+import qmlp
+
+
+class QMLP(Task):
+ ######################
+
+ def __init__(
+ self,
+ nb_train_samples,
+ nb_test_samples,
+ batch_size,
+ result_dir,
+ logger=None,
+ device=torch.device("cpu"),
+ ):
+ super().__init__()
+
+ self.device = device
+ self.batch_size = batch_size
+ self.nb_samples_per_mlp = 256
+
+ if logger is not None:
+ logger(
+ f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
+ )
+
+ seq, q_test_set, test_error = qmlp.generate_sequence_and_test_set(
+ nb_mlps=nb_train_samples + nb_test_samples,
+ nb_samples=self.nb_samples_per_mlp,
+ device=self.device,
+ batch_size=64,
+ nb_epochs=250,
+ nb_mlps_per_batch=1024,
+ )
+
+ self.train_input = seq[:nb_train_samples]
+ self.train_q_test_set = q_test_set[:nb_train_samples]
+ self.train_ref_test_errors = test_error[:nb_train_samples]
+ self.test_input = seq[nb_train_samples:]
+ self.test_q_test_set = q_test_set[nb_train_samples:]
+ self.test_ref_test_errors = test_error[nb_train_samples:]
+
+ filename = os.path.join(result_dir, f"train_errors_ref.dat")
+ with open(filename, "w") as f:
+ for e in self.train_ref_test_errors:
+ f.write(f"{e}\n")
+
+ filename = os.path.join(result_dir, f"test_errors_ref.dat")
+ with open(filename, "w") as f:
+ for e in self.test_ref_test_errors:
+ f.write(f"{e}\n")
+
+ self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
+
+ def batches(self, split="train"):
+ assert split in {"train", "test"}
+ input = self.train_input if split == "train" else self.test_input
+ for batch in tqdm.tqdm(
+ input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
+ ):
+ yield batch
+
+ def vocabulary_size(self):
+ return self.nb_codes
+
+ def produce_results(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis
+ ):
+ correct = self.test_input[:1000]
+ result = correct.clone()
+ ar_mask = (
+ torch.arange(result.size(1), device=result.device)
+ > self.nb_samples_per_mlp * 3 + 1
+ ).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,
+ )
+
+ q_train_set = result[:, : self.nb_samples_per_mlp * 3]
+ q_params = result[:, self.nb_samples_per_mlp * 3 + 1 :]
+ error_test = qmlp.evaluate_q_params(q_params, self.test_q_test_set)
+
+ filename = os.path.join(result_dir, f"test_errors_{n_epoch:04d}.dat")
+ with open(filename, "w") as f:
+ for e in error_test:
+ f.write(f"{e}\n")
+
+
+######################################################################
+
+import escape
+
+
+class Escape(Task):
+ def __init__(
+ self,
+ nb_train_samples,
+ nb_test_samples,
+ batch_size,
+ height,
+ width,
+ T,
+ nb_walls,
+ logger=None,
+ device=torch.device("cpu"),
+ ):
+ super().__init__()
+
+ self.batch_size = batch_size
+ self.device = device
+ self.height = height
+ self.width = width
+
+ states, actions, rewards = escape.generate_episodes(
+ nb_train_samples + nb_test_samples, height, width, T, nb_walls
+ )
+ seq = escape.episodes2seq(states, actions, rewards)
+ # seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3]
+ self.train_input = seq[:nb_train_samples].to(self.device)
+ self.test_input = seq[nb_train_samples:].to(self.device)
+
+ 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 batch
+
+ def vocabulary_size(self):
+ 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_lookahead_reward = 0
+ index_states = 1
+ index_action = state_len + 1
+ index_reward = state_len + 2
+ it_len = state_len + 3 # lookahead_reward / state / action / reward
+
+ result[:, it_len:] = -1
+
+ 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[:10].detach().clone())
+
+ # 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)
+
+ for u in tqdm.tqdm(
+ range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
+ ):
+ lr, _, _, _ = escape.seq2episodes(result[:, :u], self.height, self.width)
+
+ # Generate the lookahead_reward and state
+ ar_mask = (t % it_len == index_lookahead_reward).long() * (
+ t <= u + index_lookahead_reward
+ ).long()
+ 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")
+
+ # Saving the generated sequences
+
+ lr, s, a, r = 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}")
+
+ def produce_results(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
+ ):
+ result = self.test_input[:250].clone()
+
+ # Saving the ground truth
+
+ lr, s, a, r = 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_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.height * self.width + 3
+ ).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 = 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_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
+ )