+ self.test_descr = self.grid_factory.generate_samples(
+ nb=nb_test_samples, fraction_play=0.0, progress_bar=lambda r: tqdm.tqdm(r)
+ )
+
+ if fraction_play > 0:
+ self.play_descr = self.grid_factory.generate_samples(
+ nb=25, fraction_play=1.0, progress_bar=lambda r: tqdm.tqdm(r)
+ )
+ else:
+ self.play_descr = []
+
+ # Build the tokenizer
+ tokens = set()
+ for d in [self.train_descr, self.test_descr, self.play_descr]:
+ for s in d:
+ for t in s.strip().split(" "):
+ tokens.add(t)
+ # make this set a sorted list to get the same tensors given
+ # the same descr
+ tokens = list(tokens)
+ tokens.sort()
+ tokens = ["#"] + tokens
+ self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
+ self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
+ self.t_nul = self.token2id["#"]
+ self.t_true = self.token2id["true"]
+ self.t_false = self.token2id["false"]
+ self.t_pipe = self.token2id["|"]
+
+ # Tokenize the train and test sets
+ self.train_input = self.str2tensor(self.train_descr)
+ self.test_input = self.str2tensor(self.test_descr)
+ self.play_input = (
+ None if len(self.play_descr) == 0 else self.str2tensor(self.play_descr)
+ )
+
+ 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 self.trim(batch)
+
+ def vocabulary_size(self):
+ return len(self.token2id)
+
+ def produce_results(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis
+ ):
+ correct = self.test_input[:1000]
+ result = correct.clone()
+ ar_mask = torch.logical_or(result == self.t_true, result == self.t_false).long()
+ result *= 1 - ar_mask # paraaaaanoiaaaaaaa
+
+ logger(f"----------------------------------------------------------")
+
+ for e in self.tensor2str(result[:10]):
+ logger(f"test_before {e}")
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ device=self.device,
+ )
+
+ logger(f"----------------------------------------------------------")
+
+ for e in self.tensor2str(result[:10]):
+ logger(f"test_after {e}")
+
+ logger(f"----------------------------------------------------------")
+
+ nb_total = ar_mask.sum().item()
+ nb_correct = ((correct == result).long() * ar_mask).sum().item()
+
+ logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
+ logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
+
+ if self.play_input is not None:
+ result = self.play_input.clone()
+ ar_mask = (result == self.t_pipe).long().cumsum(dim=1).clamp(max=1)
+ result *= 1 - ar_mask # paraaaaanoiaaaaaaa
+
+ logger(f"----------------------------------------------------------")
+
+ for e in self.tensor2str(result[:10]):
+ logger(f"play_before {e}")
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ device=self.device,
+ )
+
+ logger(f"----------------------------------------------------------")
+
+ for e in self.tensor2str(result[:10]):
+ logger(f"play_after {e}")
+
+ logger(f"----------------------------------------------------------")
+
+
+######################################################################
+
+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,
+ 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
+ )
+ seq = escape.episodes2seq(states, actions, rewards, lookahead_delta=T)
+ # 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)
+
+ self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
+
+ 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 self.nb_codes
+
+ def thinking_autoregression(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
+ ):
+ result = self.test_input[:100].clone()
+ t = torch.arange(result.size(1), device=result.device)
+ state_len = self.height * self.width
+ iteration_len = state_len + 3
+
+ def ar():
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ device=self.device,
+ )
+
+ for u in range(
+ iteration_len, result.size(1) - iteration_len + 1, iteration_len
+ ):
+ # Put a lookahead reward to -1, sample the next state
+ result[:, u - 1] = (-1) + 1 + escape.first_lookahead_rewards_code
+ ar_mask = (t >= u).long() * (t < u + state_len).long()
+ ar_mask = ar_mask[None, :]
+ ar_mask = ar_mask.expand_as(result)
+ result *= 1 - ar_mask
+ ar()
+
+ # 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()
+
+ # Saving the generated sequences
+
+ s, a, r, lr = escape.seq2episodes(
+ result, self.height, self.width, lookahead=True
+ )
+ str = escape.episodes2str(
+ s, a, r, lookahead_rewards=lr, 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[:100].clone()
+
+ # Saving the ground truth
+
+ s, a, r, lr = escape.seq2episodes(
+ result, self.height, self.width, lookahead=True
+ )
+ str = escape.episodes2str(
+ s, a, r, lookahead_rewards=lr, 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
+
+ s, a, r, lr = escape.seq2episodes(
+ result, self.height, self.width, lookahead=True
+ )
+ str = escape.episodes2str(
+ s, a, r, lookahead_rewards=lr, 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
+ )