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("--no_checkpoint", action="store_true", default=False)
71 parser.add_argument("--overwrite_results", action="store_true", default=False)
73 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
75 ##############################
78 parser.add_argument("--maze_height", type=int, default=13)
80 parser.add_argument("--maze_width", type=int, default=21)
82 parser.add_argument("--maze_nb_walls", type=int, default=15)
84 ##############################
87 parser.add_argument("--oneshot", action="store_true", default=False)
89 parser.add_argument("--oneshot_input", type=str, default="head")
91 parser.add_argument("--oneshot_output", type=str, default="trace")
93 ######################################################################
95 args = parser.parse_args()
98 os.mkdir(args.result_dir)
99 except FileExistsError:
100 if not args.overwrite_results:
101 print(f"result directory {args.result_dir} already exists")
104 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
107 # torch.backends.cudnn.deterministic = True
108 # torch.backends.cudnn.benchmark = False
109 # torch.use_deterministic_algorithms(True)
110 torch.manual_seed(args.seed)
111 if torch.cuda.is_available():
112 torch.cuda.manual_seed_all(args.seed)
114 ######################################################################
118 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
120 if log_file is not None:
121 log_file.write(t + s + "\n")
129 log_string(f"args.{n} {getattr(args, n)}")
131 ######################################################################
134 def generation_order(x, fixed_len):
135 if args.random_regression_order:
136 order = torch.rand(x.size(), device=x.device)
137 order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=order.device)
138 order = order.sort(1).indices
141 torch.arange(x.size(1), device=x.device).unsqueeze(0).expand(x.size(0), -1)
146 def reorder(x, order, back=False): # x is NxTxD1x...xDk, order is NxT'
147 u = x.reshape(x.size()[:2] + (-1,))
148 order = order.unsqueeze(-1).expand(-1, -1, u.size(-1))
151 v.scatter_(1, order, u)
153 v = u.gather(1, order)
154 v = v.reshape(v.size()[:2] + x.size()[2:])
158 def shuffle(x, fixed_len):
159 order = generation_order(x, fixed_len)
160 return reorder(x, order), order
163 ######################################################################
165 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
166 # tokens that should be generated
169 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
170 for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
171 i = (ar_mask.sum(0) > 0).nonzero()
173 # Needed to initialize the model's cache
174 model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
175 for s in range(i.min(), i.max() + 1):
176 output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
177 logits = output[:, s]
178 if args.deterministic_synthesis:
179 t_next = logits.argmax(1)
181 dist = torch.distributions.categorical.Categorical(logits=logits)
182 t_next = dist.sample()
183 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
186 ######################################################################
189 def compute_perplexity(model, fixed_len, split="train"):
190 with torch.autograd.no_grad():
194 nb_samples, acc_loss = 0, 0.0
196 for input in task.batches(split=split):
197 input = input.to(device)
198 x, order = shuffle(input, fixed_len)
199 x = model(mygpt.BracketedSequence(x), order=order).x
200 output = reorder(x, order, back=True)
201 loss = F.cross_entropy(output.transpose(1, 2), input)
202 acc_loss += loss.item() * input.size(0)
203 nb_samples += input.size(0)
207 return math.exp(min(100, acc_loss / nb_samples))
210 ######################################################################
213 def oneshot_policy_loss(mazes, output, policies, height, width):
214 masks = (mazes == maze.v_empty).unsqueeze(-1)
215 targets = policies.permute(0, 2, 1) * masks
216 output = output * masks
217 return -(output.log_softmax(-1) * targets).sum() / masks.sum()
220 def oneshot_trace_loss(mazes, output, policies, height, width):
221 masks = mazes == maze.v_empty
222 targets = maze.stationary_densities(
223 mazes.view(-1, height, width), policies.view(-1, 4, height, width)
225 targets = targets * masks
226 output = output.squeeze(-1) * masks
227 return (output - targets).abs().sum() / masks.sum()
230 def oneshot(gpt, task):
234 if args.oneshot_input == "head":
235 dim_in = args.dim_model
236 elif args.oneshot_input == "deep":
237 dim_in = args.dim_model * args.nb_blocks * 2
239 raise ValueError(f"{args.oneshot_input=}")
241 if args.oneshot_output == "policy":
243 compute_loss = oneshot_policy_loss
244 elif args.oneshot_output == "trace":
246 compute_loss = oneshot_trace_loss
248 raise ValueError(f"{args.oneshot_output=}")
250 model = nn.Sequential(
251 nn.Linear(dim_in, args.dim_model),
253 nn.Linear(args.dim_model, args.dim_model),
255 nn.Linear(args.dim_model, dim_out),
258 for n_epoch in range(args.nb_epochs):
259 learning_rate = learning_rate_schedule[n_epoch]
260 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
262 acc_train_loss, nb_train_samples = 0, 0
263 for mazes, policies in task.policy_batches(split="train"):
264 x, order = shuffle(mazes, task.height * task.width)
265 x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
266 output_gpt = reorder(x, order, back=True)
267 output = model(output_gpt)
269 loss = compute_loss(mazes, output, policies, task.height, task.width)
270 acc_train_loss += loss.item() * mazes.size(0)
271 nb_train_samples += mazes.size(0)
273 optimizer.zero_grad()
277 acc_test_loss, nb_test_samples = 0, 0
278 for mazes, policies in task.policy_batches(split="test"):
279 x, order = shuffle(mazes, task.height * task.width)
280 x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
281 output_gpt = reorder(x, order, back=True)
282 output = model(output_gpt)
283 loss = compute_loss(mazes, output, policies, task.height, task.width)
284 acc_test_loss += loss.item() * mazes.size(0)
285 nb_test_samples += mazes.size(0)
288 f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
291 # -------------------
292 mazes = task.test_input[:32, : task.height * task.width]
293 policies = task.test_policies[:32]
294 x, order = shuffle(mazes, task.height * task.width)
295 x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
296 output_gpt = reorder(x, order, back=True)
297 output = model(output_gpt)
298 if args.oneshot_output == "policy":
299 targets = policies.permute(0, 2, 1)
301 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
303 elif args.oneshot_output == "trace":
304 targets = maze.stationary_densities(
305 mazes.view(-1, task.height, task.width),
306 policies.view(-1, 4, task.height, task.width),
310 raise ValueError(f"{args.oneshot_output=}")
312 scores = scores.reshape(-1, task.height, task.width)
313 mazes = mazes.reshape(-1, task.height, task.width)
314 targets = targets.reshape(-1, task.height, task.width)
318 f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
324 # -------------------
329 ######################################################################
333 def batches(self, split="train", nb_to_use=-1, desc=None):
336 def vocabulary_size(self):
339 def produce_results(self, n_epoch, model):
343 ######################################################################
348 class TaskMaze(Task):
349 def map2seq(self, *m):
350 return torch.cat([x.flatten(1) for x in m], 1)
352 def seq2map(self, s):
353 s = s.reshape(s.size(0), -1, self.height, self.width)
354 return (s[:, k] for k in range(s.size(1)))
364 device=torch.device("cpu"),
366 self.batch_size = batch_size
371 train_mazes, train_paths, train_policies = maze.create_maze_data(
376 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
378 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
379 self.train_policies = train_policies.flatten(-2).to(device)
381 test_mazes, test_paths, test_policies = maze.create_maze_data(
386 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
388 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
389 self.test_policies = test_policies.flatten(-2).to(device)
391 self.nb_codes = self.train_input.max() + 1
393 def batches(self, split="train", nb_to_use=-1, desc=None):
394 assert split in {"train", "test"}
395 input = self.train_input if split == "train" else self.test_input
397 input = input[:nb_to_use]
399 desc = f"epoch-{split}"
400 for batch in tqdm.tqdm(
401 input.split(self.batch_size), dynamic_ncols=True, desc=desc
405 def policy_batches(self, split="train", nb_to_use=-1, desc=None):
406 assert split in {"train", "test"}
407 input = self.train_input if split == "train" else self.test_input
408 policies = self.train_policies if split == "train" else self.test_policies
409 input = input[:, : self.height * self.width]
410 policies = policies * (input != maze.v_wall)[:, None]
413 input = input[:nb_to_use]
414 policies = policies[:nb_to_use]
417 desc = f"epoch-{split}"
418 for batch in tqdm.tqdm(
419 zip(input.split(self.batch_size), policies.split(self.batch_size)),
425 def vocabulary_size(self):
428 def compute_error(self, model, split="train", nb_to_use=-1):
429 nb_total, nb_correct = 0, 0
430 for input in task.batches(split, nb_to_use):
431 result = input.clone()
432 ar_mask = result.new_zeros(result.size())
433 ar_mask[:, self.height * self.width :] = 1
434 result *= 1 - ar_mask
435 x, order = shuffle(result, self.height * self.width)
436 masked_inplace_autoregression(
437 model, self.batch_size, x, ar_mask, order=order
439 result = reorder(x, order, back=True)
440 mazes, paths = self.seq2map(result)
441 nb_correct += maze.path_correctness(mazes, paths).long().sum()
442 nb_total += mazes.size(0)
444 return nb_total, nb_correct
446 def produce_results(self, n_epoch, model):
447 with torch.autograd.no_grad():
451 train_nb_total, train_nb_correct = self.compute_error(
452 model, "train", nb_to_use=1000
455 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
458 test_nb_total, test_nb_correct = self.compute_error(
459 model, "test", nb_to_use=1000
462 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
465 input = self.test_input[:32]
466 result = input.clone()
467 ar_mask = result.new_zeros(result.size())
468 ar_mask[:, self.height * self.width :] = 1
469 result *= 1 - ar_mask
470 masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
472 mazes, paths = self.seq2map(input)
473 _, predicted_paths = self.seq2map(result)
475 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
478 predicted_paths=predicted_paths,
479 path_correct=maze.path_correctness(mazes, predicted_paths),
485 ######################################################################
487 log_string(f"device {device}")
491 nb_train_samples=args.nb_train_samples,
492 nb_test_samples=args.nb_test_samples,
493 batch_size=args.batch_size,
494 height=args.maze_height,
495 width=args.maze_width,
496 nb_walls=args.maze_nb_walls,
501 vocabulary_size = task.vocabulary_size()
503 log_string(f"vocabulary_size {vocabulary_size}")
505 ##############################
508 vocabulary_size=vocabulary_size,
509 dim_model=args.dim_model,
510 dim_keys=args.dim_keys,
511 dim_hidden=args.dim_hidden,
512 nb_heads=args.nb_heads,
513 nb_blocks=args.nb_blocks,
515 dropout=args.dropout,
520 nb_parameters = sum(p.numel() for p in model.parameters())
521 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
523 ######################################################################
525 nb_epochs_finished = 0
527 if args.no_checkpoint:
528 log_string(f"not trying to load checkpoint.")
532 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
533 checkpoint = torch.load(checkpoint_name)
534 nb_epochs_finished = checkpoint["nb_epochs_finished"]
535 model.load_state_dict(checkpoint["model_state"])
536 torch.set_rng_state(checkpoint["rng_state"])
537 if torch.cuda.is_available():
538 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
540 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
542 except FileNotFoundError:
543 log_string("starting from scratch.")
546 log_string("error when loading the checkpoint.")
549 ######################################################################
552 for input in task.batches(split="train"):
553 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
554 token_probas = token_count / token_count.sum()
555 entropy = -torch.xlogy(token_probas, token_probas).sum()
556 train_set_perplexity = math.exp(entropy)
558 ##############################
560 if args.learning_rate_schedule == "cos":
561 learning_rate_schedule = {}
562 for n_epoch in range(args.nb_epochs):
563 u = n_epoch / args.nb_epochs * math.pi
564 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
569 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
573 learning_rate_schedule = {}
574 learning_rate = args.learning_rate
575 for n_epoch in range(args.nb_epochs):
577 learning_rate = u[n_epoch]
578 learning_rate_schedule[n_epoch] = learning_rate
580 log_string(f"learning_rate_schedule {learning_rate_schedule}")
582 ##############################
584 if nb_epochs_finished >= args.nb_epochs:
585 n_epoch = nb_epochs_finished
586 train_perplexity = compute_perplexity(
587 model, fixed_len=task.height * task.width, split="train"
589 test_perplexity = compute_perplexity(
590 model, fixed_len=task.height * task.width, split="test"
594 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
597 task.produce_results(n_epoch, model)
599 ##############################
601 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
602 learning_rate = learning_rate_schedule[n_epoch]
604 log_string(f"learning_rate {learning_rate}")
606 if args.optim == "sgd":
607 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
608 elif args.optim == "adam":
609 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
610 elif args.optim == "adamw":
611 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
613 raise ValueError(f"{args.optim=}")
617 nb_train_samples, acc_train_loss = 0, 0.0
619 for input in task.batches(split="train"):
620 input = input.to(device)
621 x, order = shuffle(input, task.height * task.width)
622 x = model(mygpt.BracketedSequence(x), order=order).x
623 output = reorder(x, order, back=True)
624 loss = F.cross_entropy(output.transpose(1, 2), input)
625 acc_train_loss += loss.item() * input.size(0)
626 nb_train_samples += input.size(0)
628 optimizer.zero_grad()
632 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
633 test_perplexity = compute_perplexity(
634 model, fixed_len=task.height * task.width, split="test"
638 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
641 task.produce_results(n_epoch, model)
644 "nb_epochs_finished": n_epoch + 1,
645 "model_state": model.state_dict(),
646 "rng_state": torch.get_rng_state(),
649 if torch.cuda.is_available():
650 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
652 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
653 torch.save(checkpoint, checkpoint_name)
654 log_string(f"saved checkpoint {checkpoint_name}")
656 ######################################################################
661 ######################################################################