3 # Any copyright is dedicated to the Public Domain.
4 # https://creativecommons.org/publicdomain/zero/1.0/
6 # Written by Francois Fleuret <francois@fleuret.org>
8 # torch.backends.cuda.matmul.allow_tf23
9 # torch.autocast(torch.bfloat16)
11 import math, sys, argparse, time, tqdm, itertools, os
13 import torch, torchvision
15 from torch.nn import functional as F
17 import mygpt, tensorstack
19 ######################################################################
21 if torch.cuda.is_available():
22 device = torch.device("cuda")
23 torch.backends.cuda.matmul.allow_tf32 = True
25 device = torch.device("cpu")
27 ######################################################################
29 parser = argparse.ArgumentParser(description="A maze shortest path solving with a GPT.")
31 parser.add_argument("--log_filename", type=str, default="train.log")
33 parser.add_argument("--result_dir", type=str, default="results_default")
35 parser.add_argument("--seed", type=int, default=0)
37 parser.add_argument("--nb_epochs", type=int, default=25)
39 parser.add_argument("--nb_train_samples", type=int, default=200000)
41 parser.add_argument("--nb_test_samples", type=int, default=50000)
43 parser.add_argument("--batch_size", type=int, default=25)
45 parser.add_argument("--optim", type=str, default="adam")
47 parser.add_argument("--learning_rate", type=float, default=1e-3)
50 "--learning_rate_schedule", type=str, default="10: 2e-4,20: 4e-5,30: 8e-6"
53 parser.add_argument("--dim_model", type=int, default=512)
55 parser.add_argument("--dim_keys", type=int, default=64)
57 parser.add_argument("--dim_hidden", type=int, default=2048)
59 parser.add_argument("--nb_heads", type=int, default=8)
61 parser.add_argument("--nb_blocks", type=int, default=12)
63 parser.add_argument("--dropout", type=float, default=0.1)
65 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
67 parser.add_argument("--random_regression_order", action="store_true", default=False)
69 parser.add_argument("--noncausal_prompt", action="store_true", default=False)
71 parser.add_argument("--no_checkpoint", action="store_true", default=False)
73 parser.add_argument("--overwrite_results", action="store_true", default=False)
75 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
77 ##############################
80 parser.add_argument("--maze_height", type=int, default=13)
82 parser.add_argument("--maze_width", type=int, default=21)
84 parser.add_argument("--maze_nb_walls", type=int, default=15)
86 ##############################
89 parser.add_argument("--oneshot", action="store_true", default=False)
91 parser.add_argument("--oneshot_input", type=str, default="head")
93 parser.add_argument("--oneshot_output", type=str, default="trace")
95 ######################################################################
97 args = parser.parse_args()
100 os.mkdir(args.result_dir)
101 except FileExistsError:
102 if not args.overwrite_results:
103 print(f"result directory {args.result_dir} already exists")
106 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
109 # torch.backends.cudnn.deterministic = True
110 # torch.backends.cudnn.benchmark = False
111 # torch.use_deterministic_algorithms(True)
112 torch.manual_seed(args.seed)
113 if torch.cuda.is_available():
114 torch.cuda.manual_seed_all(args.seed)
116 ######################################################################
120 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
122 if log_file is not None:
123 log_file.write(t + s + "\n")
131 log_string(f"args.{n} {getattr(args, n)}")
133 ######################################################################
136 def reorder(x, order, reverse=False): # x is NxTxD1x...xDk, order is NxT'
137 u = x.reshape(x.size()[:2] + (-1,))
138 order = order.unsqueeze(-1).expand(-1, -1, u.size(-1))
140 v = u.new(u.size()).scatter_(1, order, u)
142 v = u.gather(1, order)
143 v = v.reshape(v.size()[:2] + x.size()[2:])
147 def shuffle(x, prompt_len):
148 if args.random_regression_order:
149 order = torch.rand(x.size(), device=x.device)
150 order[:, :prompt_len] = torch.arange(-prompt_len, 0, device=x.device)
151 order = order.sort(1).indices
154 torch.arange(x.size(1), device=x.device).unsqueeze(0).expand(x.size(0), -1)
156 return reorder(x, order), order
159 def eval_mygpt(model, input, mode="standard", prompt_len=0):
160 x, order = shuffle(input, prompt_len)
161 x = model(mygpt.BracketedSequence(x), mode=mode, order=order).x
162 return reorder(x, order, reverse=True)
165 ######################################################################
167 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
168 # tokens that should be generated
171 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
172 for input, ar_mask, order in zip(
173 input.split(batch_size), ar_mask.split(batch_size), order.split(batch_size)
175 i = (ar_mask.sum(0) > 0).nonzero()
177 # Needed to initialize the model's cache
178 model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
179 for s in range(i.min(), i.max() + 1):
180 output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
181 logits = output[:, s]
182 if args.deterministic_synthesis:
183 t_next = logits.argmax(1)
185 dist = torch.distributions.categorical.Categorical(logits=logits)
186 t_next = dist.sample()
187 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
190 ######################################################################
193 def compute_perplexity(model, task, prompt_len, split="train"):
194 with torch.autograd.no_grad():
198 nb_samples, acc_loss = 0, 0.0
200 for input in task.batches(split=split):
201 input = input.to(device)
202 output = eval_mygpt(model, input, prompt_len=prompt_len)
203 if args.noncausal_prompt:
204 d = input.size(1) // 2
205 loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
207 loss = F.cross_entropy(output.transpose(1, 2), input)
208 acc_loss += loss.item() * input.size(0)
209 nb_samples += input.size(0)
213 return math.exp(min(100, acc_loss / nb_samples))
216 ######################################################################
219 def oneshot_policy_loss(mazes, output, policies, height, width):
220 masks = (mazes == maze.v_empty).unsqueeze(-1)
221 targets = policies.permute(0, 2, 1) * masks
222 output = output * masks
223 return -(output.log_softmax(-1) * targets).sum() / masks.sum()
226 def oneshot_trace_loss(mazes, output, policies, height, width):
227 masks = mazes == maze.v_empty
228 targets = maze.stationary_densities(
229 mazes.view(-1, height, width), policies.view(-1, 4, height, width)
231 targets = targets * masks
232 output = output.squeeze(-1) * masks
233 return (output - targets).abs().sum() / masks.sum()
236 def oneshot(model, learning_rate_scheduler, task):
239 mazes = task.test_input[:32].clone()
240 mazes[:, task.height * task.width :] = 0
241 output = eval_mygpt(model, mazes, prompt_len=task.height * task.width)
242 output = F.softmax(output, dim=2)
243 print(f"{output.size()=}")
244 proba_path = output[:, task.height * task.width :, 4].reshape(
245 -1, task.height, task.width
247 mazes = mazes[:, : task.height * task.width].reshape(-1, task.height, task.width)
248 # targets = targets.reshape(-1, task.height, task.width)
249 filename = f"oneshot.png"
251 os.path.join(args.result_dir, filename),
253 score_paths=proba_path,
254 # score_truth=targets,
256 log_string(f"wrote {filename}")
259 def oneshot_old(gpt, learning_rate_scheduler, task):
263 if args.oneshot_input == "head":
264 dim_in = args.dim_model
265 elif args.oneshot_input == "deep":
266 dim_in = args.dim_model * args.nb_blocks * 2
268 raise ValueError(f"{args.oneshot_input=}")
270 if args.oneshot_output == "policy":
272 compute_loss = oneshot_policy_loss
273 elif args.oneshot_output == "trace":
275 compute_loss = oneshot_trace_loss
277 raise ValueError(f"{args.oneshot_output=}")
279 model = nn.Sequential(
280 nn.Linear(dim_in, args.dim_model),
282 nn.Linear(args.dim_model, args.dim_model),
284 nn.Linear(args.dim_model, dim_out),
287 learning_rate_scheduler.reset()
289 for n_epoch in range(args.nb_epochs):
290 learning_rate = learning_rate_scheduler.get_learning_rate()
291 log_string(f"learning_rate {n_epoch} {learning_rate}")
293 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
295 acc_train_loss, nb_train_samples = 0, 0
296 for mazes, policies in task.policy_batches(split="train"):
297 output_gpt = eval_mygpt(
298 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
300 output = model(output_gpt)
302 loss = compute_loss(mazes, output, policies, task.height, task.width)
303 acc_train_loss += loss.item() * mazes.size(0)
304 nb_train_samples += mazes.size(0)
306 optimizer.zero_grad()
310 learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
312 acc_test_loss, nb_test_samples = 0, 0
313 for mazes, policies in task.policy_batches(split="test"):
314 output_gpt = eval_mygpt(
315 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
317 output = model(output_gpt)
318 loss = compute_loss(mazes, output, policies, task.height, task.width)
319 acc_test_loss += loss.item() * mazes.size(0)
320 nb_test_samples += mazes.size(0)
323 f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
326 # -------------------
327 mazes = task.test_input[:32, : task.height * task.width]
328 policies = task.test_policies[:32]
329 output_gpt = eval_mygpt(
330 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
332 output = model(output_gpt)
333 if args.oneshot_output == "policy":
334 targets = policies.permute(0, 2, 1)
336 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
338 elif args.oneshot_output == "trace":
339 targets = maze.stationary_densities(
340 mazes.view(-1, task.height, task.width),
341 policies.view(-1, 4, task.height, task.width),
345 raise ValueError(f"{args.oneshot_output=}")
347 scores = scores.reshape(-1, task.height, task.width)
348 mazes = mazes.reshape(-1, task.height, task.width)
349 targets = targets.reshape(-1, task.height, task.width)
351 f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png"
354 os.path.join(args.result_dir, filename),
359 log_string(f"wrote {filename}")
361 # -------------------
366 ######################################################################
369 class LearningRateScheduler:
370 def get_learning_rate(self):
373 def update(self, nb_finished_epochs, loss):
382 def set_state(self, state):
384 for k, v in state.items():
388 class StepWiseScheduler(LearningRateScheduler):
389 def __init__(self, schedule):
390 self.nb_finished_epochs = 0
391 self.schedule = schedule
393 def get_learning_rate(self):
394 return self.schedule[self.nb_finished_epochs]
396 def update(self, nb_finished_epochs, loss):
397 self.nb_finished_epochs = nb_finished_epochs
400 self.nb_finished_epochs = 0
403 return {"nb_finished_epochs": self.nb_finished_epochs}
406 class AutoScheduler(LearningRateScheduler):
407 def __init__(self, learning_rate_init, growth=1.0, degrowth=0.2):
408 self.learning_rate_init = learning_rate_init
409 self.learning_rate = learning_rate_init
411 self.degrowth = degrowth
412 self.pred_loss = None
414 def get_learning_rate(self):
415 return self.learning_rate
417 def update(self, nb_finished_epochs, loss):
418 if self.pred_loss is not None:
419 if loss >= self.pred_loss:
420 self.learning_rate *= self.degrowth
422 self.learning_rate *= self.growth
423 self.pred_loss = loss
426 self.learning_rate = self.learning_rate_init
430 "learning_rate_init": self.learning_rate_init,
431 "pred_loss": self.pred_loss,
435 ######################################################################
439 def batches(self, split="train", nb_to_use=-1, desc=None):
442 def vocabulary_size(self):
445 def produce_results(self, n_epoch, model):
449 ######################################################################
454 class TaskMaze(Task):
455 def map2seq(self, *m):
456 return torch.cat([x.flatten(1) for x in m], 1)
458 def seq2map(self, s):
459 s = s.reshape(s.size(0), -1, self.height, self.width)
460 return (s[:, k] for k in range(s.size(1)))
470 device=torch.device("cpu"),
472 self.batch_size = batch_size
477 train_mazes, train_paths, train_policies = maze.create_maze_data(
482 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
484 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
485 self.train_policies = train_policies.flatten(-2).to(device)
487 test_mazes, test_paths, test_policies = maze.create_maze_data(
492 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
494 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
495 self.test_policies = test_policies.flatten(-2).to(device)
497 self.nb_codes = self.train_input.max() + 1
499 def batches(self, split="train", nb_to_use=-1, desc=None):
500 assert split in {"train", "test"}
501 input = self.train_input if split == "train" else self.test_input
503 input = input[:nb_to_use]
505 desc = f"epoch-{split}"
506 for batch in tqdm.tqdm(
507 input.split(self.batch_size), dynamic_ncols=True, desc=desc
511 def policy_batches(self, split="train", nb_to_use=-1, desc=None):
512 assert split in {"train", "test"}
513 input = self.train_input if split == "train" else self.test_input
514 policies = self.train_policies if split == "train" else self.test_policies
515 input = input[:, : self.height * self.width]
516 policies = policies * (input != maze.v_wall)[:, None]
519 input = input[:nb_to_use]
520 policies = policies[:nb_to_use]
523 desc = f"epoch-{split}"
524 for batch in tqdm.tqdm(
525 zip(input.split(self.batch_size), policies.split(self.batch_size)),
531 def vocabulary_size(self):
534 def compute_error(self, model, split="train", nb_to_use=-1):
535 nb_total, nb_correct = 0, 0
536 for input in task.batches(split, nb_to_use):
537 result = input.clone()
538 ar_mask = result.new_zeros(result.size())
539 ar_mask[:, self.height * self.width :] = 1
540 result *= 1 - ar_mask
541 x, order = shuffle(result, self.height * self.width)
542 masked_inplace_autoregression(
543 model, self.batch_size, x, ar_mask, order=order
545 result = reorder(x, order, reverse=True)
546 mazes, paths = self.seq2map(result)
547 nb_correct += maze.path_correctness(mazes, paths).long().sum()
548 nb_total += mazes.size(0)
550 return nb_total, nb_correct
552 def produce_results(self, n_epoch, model):
553 with torch.autograd.no_grad():
557 train_nb_total, train_nb_correct = self.compute_error(
558 model, "train", nb_to_use=1000
561 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
564 test_nb_total, test_nb_correct = self.compute_error(
565 model, "test", nb_to_use=1000
568 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
571 input = self.test_input[:32]
572 result = input.clone()
573 ar_mask = result.new_zeros(result.size())
574 ar_mask[:, self.height * self.width :] = 1
575 result *= 1 - ar_mask
576 x, order = shuffle(result, self.height * self.width)
577 masked_inplace_autoregression(
578 model, self.batch_size, x, ar_mask, order=order
580 result = reorder(x, order, reverse=True)
582 mazes, paths = self.seq2map(input)
583 _, predicted_paths = self.seq2map(result)
584 filename = f"result_{n_epoch:04d}.png"
586 os.path.join(args.result_dir, filename),
589 predicted_paths=predicted_paths,
590 path_correct=maze.path_correctness(mazes, predicted_paths),
592 log_string(f"wrote {filename}")
597 ######################################################################
599 log_string(f"device {device}")
603 nb_train_samples=args.nb_train_samples,
604 nb_test_samples=args.nb_test_samples,
605 batch_size=args.batch_size,
606 height=args.maze_height,
607 width=args.maze_width,
608 nb_walls=args.maze_nb_walls,
613 vocabulary_size = task.vocabulary_size()
615 log_string(f"vocabulary_size {vocabulary_size}")
617 ##############################
620 def noncausal_prompt_amm_generator(d):
621 q = torch.arange(d)[:, None]
622 k = torch.arange(d)[None, :]
623 s = args.maze_height * args.maze_width
624 return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
628 def noncausal_prompt_oneshot_amm_generator(d):
629 q = torch.arange(d)[:, None]
630 k = torch.arange(d)[None, :]
631 s = args.maze_height * args.maze_width
637 amm_generator = noncausal_prompt_oneshot_amm_generator
638 elif args.noncausal_prompt:
639 amm_generator = noncausal_prompt_amm_generator
644 vocabulary_size=vocabulary_size,
645 dim_model=args.dim_model,
646 dim_keys=args.dim_keys,
647 dim_hidden=args.dim_hidden,
648 nb_heads=args.nb_heads,
649 nb_blocks=args.nb_blocks,
651 dropout=args.dropout,
652 amm_generator=amm_generator,
657 nb_parameters = sum(p.numel() for p in model.parameters())
658 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
660 ######################################################################
662 if args.learning_rate_schedule == "auto":
663 learning_rate_scheduler = AutoScheduler(args.learning_rate)
665 elif args.learning_rate_schedule == "cos":
667 for n_epoch in range(args.nb_epochs):
668 u = n_epoch / args.nb_epochs * math.pi
669 schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
670 learning_rate_scheduler = StepWiseScheduler(schedule)
671 log_string(f"learning_rate_schedule {schedule}")
677 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
682 learning_rate = args.learning_rate
683 for n_epoch in range(args.nb_epochs):
685 learning_rate = u[n_epoch]
686 schedule[n_epoch] = learning_rate
687 learning_rate_scheduler = StepWiseScheduler(schedule)
688 log_string(f"learning_rate_schedule {schedule}")
690 ######################################################################
692 nb_epochs_finished = 0
694 if args.no_checkpoint:
695 log_string(f"not trying to load checkpoint.")
699 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
700 checkpoint = torch.load(checkpoint_name)
701 nb_epochs_finished = checkpoint["nb_epochs_finished"]
702 model.load_state_dict(checkpoint["model_state"])
703 learning_rate_scheduler.set_state(checkpoint["learning_rate_scheduler_state"])
704 torch.set_rng_state(checkpoint["rng_state"])
705 if torch.cuda.is_available():
706 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
708 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
710 except FileNotFoundError:
711 log_string("starting from scratch.")
714 # log_string("error when loading the checkpoint.")
717 ######################################################################
720 oneshot(model, learning_rate_scheduler, task)
723 ######################################################################
726 for input in task.batches(split="train"):
727 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
728 token_probas = token_count / token_count.sum()
729 entropy = -torch.xlogy(token_probas, token_probas).sum()
730 train_set_perplexity = math.exp(entropy)
732 ##############################
734 if nb_epochs_finished >= args.nb_epochs:
735 n_epoch = nb_epochs_finished
736 train_perplexity = compute_perplexity(
737 model, task, prompt_len=task.height * task.width, split="train"
739 test_perplexity = compute_perplexity(
740 model, task, prompt_len=task.height * task.width, split="test"
744 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
747 task.produce_results(n_epoch, model)
749 ##############################
751 learning_rate_scheduler.reset()
753 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
754 learning_rate = learning_rate_scheduler.get_learning_rate()
755 log_string(f"learning_rate {n_epoch} {learning_rate}")
757 if args.optim == "sgd":
758 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
759 elif args.optim == "adam":
760 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
761 elif args.optim == "adamw":
762 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
764 raise ValueError(f"{args.optim=}")
768 nb_train_samples, acc_train_loss = 0, 0.0
770 for input in task.batches(split="train"):
771 input = input.to(device)
772 output = eval_mygpt(model, input, prompt_len=task.height * task.width)
773 if args.noncausal_prompt:
774 d = input.size(1) // 2
775 loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
777 loss = F.cross_entropy(output.transpose(1, 2), input)
778 acc_train_loss += loss.item() * input.size(0)
779 nb_train_samples += input.size(0)
781 optimizer.zero_grad()
785 learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
787 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
788 test_perplexity = compute_perplexity(
789 model, task, prompt_len=task.height * task.width, split="test"
793 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
796 task.produce_results(n_epoch, model)
799 "nb_epochs_finished": n_epoch + 1,
800 "model_state": model.state_dict(),
801 "learning_rate_scheduler_state": learning_rate_scheduler.get_state(),
802 "rng_state": torch.get_rng_state(),
805 if torch.cuda.is_available():
806 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
808 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
809 torch.save(checkpoint, checkpoint_name)
810 log_string(f"saved checkpoint {checkpoint_name}")
812 ######################################################################