# Written by Francois Fleuret <francois@fleuret.org>
-import math, os, tqdm
+import math, os, tqdm, warnings
import torch, torchvision
ar_mask,
deterministic_synthesis,
forbidden_tokens=None,
+ logit_biases=None,
progress_bar_desc="autoregression",
device=torch.device("cpu"),
):
for input, ar_mask in batches:
model.masked_inplace_autoregression(
- input, ar_mask, forbidden_tokens, deterministic_synthesis
+ input,
+ ar_mask,
+ deterministic_synthesis,
+ forbidden_tokens,
+ logit_biases,
)
model.train(t)
class Task:
- def batches(self, split="train"):
+ def batches(self, split="train", nb_to_use=-1, desc=None):
pass
def vocabulary_size(self):
class TaskFromFile(Task):
- def tensorize(self, pairs):
+ def tensorize(self, pairs, shuffle):
len_max = max([len(x[0]) for x in pairs])
input = torch.cat(
0,
).to("cpu")
+ if shuffle:
+ i = torch.randperm(input.size(0))
+ input = input[i].contiguous()
+ pred_mask = pred_mask[i].contiguous()
+
return input, pred_mask
# trim all the tensors in the tuple z to remove as much token from
def __init__(
self,
- filename,
+ train_filename,
+ test_filename,
nb_train_samples,
nb_test_samples,
batch_size,
+ shuffle=False,
device=torch.device("cpu"),
):
self.batch_size = batch_size
self.device = device
- pairs = []
- with open(filename, "r") as f:
- for _ in range(nb_train_samples + nb_test_samples):
- sequence = f.readline().strip()
- pred_mask = f.readline().strip()
- assert len(sequence) == len(pred_mask)
- assert set(pred_mask) == {"0", "1", "2"}, f"{set(pred_mask)}"
- pairs.append((sequence, pred_mask))
-
- symbols = ["#"] + list(set("".join([x[0] for x in pairs])) - set(["#"]))
+ def read_file(filename, nb=-1):
+ pairs = []
+ with open(filename, "r") as f:
+ while True:
+ sequence = f.readline().strip()
+ if not sequence:
+ break
+ pred_mask = f.readline().strip()
+ assert len(sequence) == len(pred_mask)
+ assert set(pred_mask).issubset({"0", "1", "2"}), f"{set(pred_mask)}"
+ pairs.append((sequence, pred_mask))
+ if len(pairs) == nb:
+ break
+
+ if nb > 0:
+ pairs = pairs[:nb]
+ assert len(pairs) == nb
+
+ return pairs
+
+ train_pairs = read_file(train_filename, nb_train_samples)
+ test_pairs = read_file(test_filename, nb_test_samples)
+
+ symbols = ["#"] + list(
+ set("".join([x[0] for x in train_pairs + test_pairs])) - set(["#"])
+ )
self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
self.id2char = dict([(n, c) for c, n in self.char2id.items()])
self.train_input, self.train_pred_masks = self.tensorize(
- pairs[:nb_train_samples]
+ train_pairs, shuffle=shuffle
+ )
+ self.test_input, self.test_pred_masks = self.tensorize(
+ test_pairs, shuffle=shuffle
)
- self.test_input, self.test_pred_masks = self.tensorize(pairs[nb_train_samples:])
def batches(self, split="train", nb_to_use=-1, desc=None):
assert split in {"train", "test"}
logger(f"----------------------------------------------------------")
- for e in self.tensor2str(result[:10]):
+ for e in self.tensor2str(result[:50]):
logger(f"test_before {e}")
masked_inplace_autoregression(
logger(f"----------------------------------------------------------")
- for e, c in zip(self.tensor2str(result[:10]), self.tensor2str(correct[:10])):
+ for e, c in zip(self.tensor2str(result[:50]), self.tensor2str(correct[:50])):
logger(f"test_after {e}")
logger(f"correct {c}")
self.train_input = self.tensorize(self.train_descr)
self.test_input = self.tensorize(self.test_descr)
- def batches(self, split="train"):
+ 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
for batch in tqdm.tqdm(
def compute_error(
self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
):
+ model_device = next(model.parameters()).device
nb_total, nb_correct = 0, 0
count = torch.zeros(
self.width * self.height,
self.width * self.height,
- device=self.device,
+ device=model_device,
dtype=torch.int64,
)
for input in self.batches(split, nb_to_use):
+ input = input.to(model_device)
result = input.clone()
ar_mask = result.new_zeros(result.size())
ar_mask[:, self.height * self.width :] = 1
eol = " " if j < count.size(1) - 1 else "\n"
f.write(f"{count[i,j]}{eol}")
- input = self.test_input[:48]
+ input = self.test_input[:48].to(next(model.parameters()).device)
result = input.clone()
ar_mask = result.new_zeros(result.size())
ar_mask[:, self.height * self.width :] = 1
device=self.device,
)
+ #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+ for label, input in [
+ ("train", self.train_input[:32]),
+ ("test", self.test_input[:32]),
+ ]:
+ output = model(BracketedSequence(input)).x
+ output = output.log_softmax(dim=-1)
+ filename = os.path.join(
+ result_dir, f"stack_with_crossentropy_{n_epoch:04d}_{label}.txt"
+ )
+ with open(filename, "w") as f:
+ for n in range(input.size(0)):
+ s = stack.seq_to_str(
+ input[n], nb_stacks=self.nb_stacks, nb_digits=self.nb_digits
+ )
+ for t, k, w in zip(range(input[n].size(0)), input[n], s.split(" ")):
+ u = (
+ " " * (10 - len(w))
+ + w
+ + " "
+ + str(output[n][t][k].exp().item())
+ + "\n"
+ )
+ f.write(u)
+ f.write("\n")
+ logger(f"wrote {filename}")
+ #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+
for n in range(result.size(0)):
logger(
f"test_after {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
self.t_nul = self.token2id["#"]
self.t_true = self.token2id["true"]
self.t_false = self.token2id["false"]
- self.t_pipe = self.token2id["|"]
+ # self.t_pipe = self.token2id["|"]
# Tokenize the train and test sets
self.train_input = self.str2tensor(self.train_descr)
None if len(self.play_descr) == 0 else self.str2tensor(self.play_descr)
)
- def batches(self, split="train"):
+ 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
for batch in tqdm.tqdm(
self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
- def batches(self, split="train"):
+ 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
for batch in tqdm.tqdm(
######################################################################
+
+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
+ )
+
+
+######################################################################
+######################################################################
+
+import world
+
+
+class World(Task):
+ def save_image(self, input, result_dir, filename, logger):
+ img = world.sample2img(input.to("cpu"), self.height, self.width)
+ image_name = os.path.join(result_dir, filename)
+ torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=8, padding=2)
+ logger(f"wrote {image_name}")
+
+ def __init__(
+ self,
+ nb_train_samples,
+ nb_test_samples,
+ batch_size,
+ result_dir=None,
+ logger=None,
+ device=torch.device("cpu"),
+ ):
+ super().__init__()
+
+ self.batch_size = batch_size
+ self.device = device
+ self.height = 6
+ self.width = 8
+
+ self.train_input = world.generate(
+ nb_train_samples, height=self.height, width=self.width
+ )
+ self.train_ar_mask = (
+ (torch.arange(self.train_input.size(1)) > self.train_input.size(1) // 2)
+ .long()[None, :]
+ .expand_as(self.train_input)
+ )
+
+ self.test_input = world.generate(
+ nb_test_samples, height=self.height, width=self.width
+ )
+ self.test_ar_mask = (
+ (torch.arange(self.test_input.size(1)) > self.test_input.size(1) // 2)
+ .long()[None, :]
+ .expand_as(self.test_input)
+ )
+
+ self.train_input, self.train_ar_mask = self.train_input.to(
+ device
+ ), self.train_ar_mask.to(device)
+ self.test_input, self.test_ar_mask = self.test_input.to(
+ device
+ ), self.test_ar_mask.to(device)
+
+ self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
+
+ if result_dir is not None:
+ self.save_image(
+ self.train_input[:96], result_dir, f"world_train.png", logger
+ )
+
+ 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 produce_results(
+ self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
+ ):
+ def compute_accuracy(input, ar_mask, logger=None):
+ input, ar_mask = input[:nmax], ar_mask[:nmax]
+ result = input.clone() * (1 - ar_mask)
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ progress_bar_desc=None,
+ device=self.device,
+ )
+
+ nb_total, nb_correct = (
+ input.size(0),
+ (input == result).long().min(dim=1).values.sum(),
+ )
+
+ return nb_total, nb_correct
+
+ train_nb_total, train_nb_correct = compute_accuracy(
+ self.train_input, self.train_ar_mask
+ )
+
+ logger(
+ f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
+ )
+
+ test_nb_total, test_nb_correct = compute_accuracy(
+ self.test_input, self.test_ar_mask, logger
+ )
+
+ logger(
+ f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
+ )
+
+ main_test_accuracy = test_nb_correct / test_nb_total
+ logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
+
+ ##############################
+
+ input, ar_mask = self.test_input[:96], self.test_ar_mask[:96]
+ result = input.clone() * (1 - ar_mask)
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ progress_bar_desc=None,
+ device=self.device,
+ )
+
+ self.save_image(
+ result[:96], result_dir, f"world_result_{n_epoch:04d}.png", logger
+ )
+
+ return main_test_accuracy
+
+ def store_new_quizzes(self, new_quizzes, for_train=True):
+ input = self.train_input if for_train else self.test_input
+
+ nb_current = input.size(0)
+ nb_new = new_quizzes.size(0)
+ if nb_new >= nb_current:
+ input[...] = new_quizzes[:nb_current]
+ else:
+ nb_kept = nb_current - nb_new
+ input[:nb_kept] = input[-nb_kept:].clone()
+ input[nb_kept:] = new_quizzes
+
+ def create_new_quizzes(
+ self, n_epoch, result_dir, logger, nb, models, other_models, nb_runs
+ ):
+ new_quizzes = torch.empty(
+ nb, self.height * self.width * 2 + 1, device=self.device, dtype=torch.int64
+ )
+ ar_mask = torch.full(new_quizzes.size(), 1, device=self.device)
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ new_quizzes,
+ ar_mask,
+ deterministic_synthesis=False,
+ progress_bar_desc="new quizzes",
+ device=self.device,
+ )
+
+ input = (
+ new_quizzes[:, None, :]
+ .expand(-1, nb_runs, -1)
+ .clone()
+ .reshape(-1, new_quizzes.size(-1))
+ )
+ result = input.clone()
+
+ ar_mask = (
+ (torch.arange(result.size(1), device=self.device) > result.size(1) // 2)
+ .long()[None, :]
+ .expand_as(result)
+ )
+
+ dispatch = torch.randint(len(other_models), (result.size(0),))
+
+ for n, m in enumerate(other_models):
+ masked_inplace_autoregression(
+ m,
+ self.batch_size,
+ result[dispatch == n],
+ ar_mask[dispatch == n],
+ deterministic_synthesis=False,
+ progress_bar_desc=None,
+ device=self.device,
+ )
+
+ nb_correct = (
+ (input == result).long().min(dim=-1).values.reshape(-1, nb_runs).sum(dim=-1)
+ )
+
+ return new_quizzes, nb_correct