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=0):
135 if args.random_regression_order:
136 order = torch.rand(x.size(), device=x.device)
137 order[:, :fixed_len] = torch.arange(-fixed_len, 0, device=x.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, reverse=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))
150 v = u.new(u.size()).scatter_(1, order, u)
152 v = u.gather(1, order)
153 v = v.reshape(v.size()[:2] + x.size()[2:])
157 def shuffle(x, fixed_len):
158 order = generation_order(x, fixed_len)
159 return reorder(x, order), order
162 def eval_mygpt(model, input, mode="standard", fixed_len=0):
163 x, order = shuffle(input, fixed_len)
164 x = model(mygpt.BracketedSequence(x), mode=mode, order=order).x
165 return reorder(x, order, reverse=True)
168 ######################################################################
170 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
171 # tokens that should be generated
174 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
175 for input, ar_mask, order in zip(
176 input.split(batch_size), ar_mask.split(batch_size), order.split(batch_size)
178 i = (ar_mask.sum(0) > 0).nonzero()
180 # Needed to initialize the model's cache
181 model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
182 for s in range(i.min(), i.max() + 1):
183 output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
184 logits = output[:, s]
185 if args.deterministic_synthesis:
186 t_next = logits.argmax(1)
188 dist = torch.distributions.categorical.Categorical(logits=logits)
189 t_next = dist.sample()
190 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
193 ######################################################################
196 def compute_perplexity(model, task, fixed_len, split="train"):
197 with torch.autograd.no_grad():
201 nb_samples, acc_loss = 0, 0.0
203 for input in task.batches(split=split):
204 input = input.to(device)
205 output = eval_mygpt(model, input, fixed_len=fixed_len)
206 loss = F.cross_entropy(output.transpose(1, 2), input)
207 acc_loss += loss.item() * input.size(0)
208 nb_samples += input.size(0)
212 return math.exp(min(100, acc_loss / nb_samples))
215 ######################################################################
218 def oneshot_policy_loss(mazes, output, policies, height, width):
219 masks = (mazes == maze.v_empty).unsqueeze(-1)
220 targets = policies.permute(0, 2, 1) * masks
221 output = output * masks
222 return -(output.log_softmax(-1) * targets).sum() / masks.sum()
225 def oneshot_trace_loss(mazes, output, policies, height, width):
226 masks = mazes == maze.v_empty
227 targets = maze.stationary_densities(
228 mazes.view(-1, height, width), policies.view(-1, 4, height, width)
230 targets = targets * masks
231 output = output.squeeze(-1) * masks
232 return (output - targets).abs().sum() / masks.sum()
235 def oneshot(gpt, task):
239 if args.oneshot_input == "head":
240 dim_in = args.dim_model
241 elif args.oneshot_input == "deep":
242 dim_in = args.dim_model * args.nb_blocks * 2
244 raise ValueError(f"{args.oneshot_input=}")
246 if args.oneshot_output == "policy":
248 compute_loss = oneshot_policy_loss
249 elif args.oneshot_output == "trace":
251 compute_loss = oneshot_trace_loss
253 raise ValueError(f"{args.oneshot_output=}")
255 model = nn.Sequential(
256 nn.Linear(dim_in, args.dim_model),
258 nn.Linear(args.dim_model, args.dim_model),
260 nn.Linear(args.dim_model, dim_out),
263 for n_epoch in range(args.nb_epochs):
264 learning_rate = learning_rate_schedule[n_epoch]
265 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
267 acc_train_loss, nb_train_samples = 0, 0
268 for mazes, policies in task.policy_batches(split="train"):
269 output_gpt = eval_mygpt(
270 gpt, mazes, mode=args.oneshot_input, fixed_len=task.height * task.width
272 output = model(output_gpt)
274 loss = compute_loss(mazes, output, policies, task.height, task.width)
275 acc_train_loss += loss.item() * mazes.size(0)
276 nb_train_samples += mazes.size(0)
278 optimizer.zero_grad()
282 acc_test_loss, nb_test_samples = 0, 0
283 for mazes, policies in task.policy_batches(split="test"):
284 output_gpt = eval_mygpt(
285 gpt, mazes, mode=args.oneshot_input, fixed_len=task.height * task.width
287 output = model(output_gpt)
288 loss = compute_loss(mazes, output, policies, task.height, task.width)
289 acc_test_loss += loss.item() * mazes.size(0)
290 nb_test_samples += mazes.size(0)
293 f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
296 # -------------------
297 mazes = task.test_input[:32, : task.height * task.width]
298 policies = task.test_policies[:32]
299 output_gpt = eval_mygpt(
300 gpt, mazes, mode=args.oneshot_input, fixed_len=task.height * task.width
302 output = model(output_gpt)
303 if args.oneshot_output == "policy":
304 targets = policies.permute(0, 2, 1)
306 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
308 elif args.oneshot_output == "trace":
309 targets = maze.stationary_densities(
310 mazes.view(-1, task.height, task.width),
311 policies.view(-1, 4, task.height, task.width),
315 raise ValueError(f"{args.oneshot_output=}")
317 scores = scores.reshape(-1, task.height, task.width)
318 mazes = mazes.reshape(-1, task.height, task.width)
319 targets = targets.reshape(-1, task.height, task.width)
321 f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png"
324 os.path.join(args.result_dir, filename),
329 log_string(f"wrote {filename}")
331 # -------------------
336 ######################################################################
340 def batches(self, split="train", nb_to_use=-1, desc=None):
343 def vocabulary_size(self):
346 def produce_results(self, n_epoch, model):
350 ######################################################################
355 class TaskMaze(Task):
356 def map2seq(self, *m):
357 return torch.cat([x.flatten(1) for x in m], 1)
359 def seq2map(self, s):
360 s = s.reshape(s.size(0), -1, self.height, self.width)
361 return (s[:, k] for k in range(s.size(1)))
371 device=torch.device("cpu"),
373 self.batch_size = batch_size
378 train_mazes, train_paths, train_policies = maze.create_maze_data(
383 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
385 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
386 self.train_policies = train_policies.flatten(-2).to(device)
388 test_mazes, test_paths, test_policies = maze.create_maze_data(
393 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
395 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
396 self.test_policies = test_policies.flatten(-2).to(device)
398 self.nb_codes = self.train_input.max() + 1
400 def batches(self, split="train", nb_to_use=-1, desc=None):
401 assert split in {"train", "test"}
402 input = self.train_input if split == "train" else self.test_input
404 input = input[:nb_to_use]
406 desc = f"epoch-{split}"
407 for batch in tqdm.tqdm(
408 input.split(self.batch_size), dynamic_ncols=True, desc=desc
412 def policy_batches(self, split="train", nb_to_use=-1, desc=None):
413 assert split in {"train", "test"}
414 input = self.train_input if split == "train" else self.test_input
415 policies = self.train_policies if split == "train" else self.test_policies
416 input = input[:, : self.height * self.width]
417 policies = policies * (input != maze.v_wall)[:, None]
420 input = input[:nb_to_use]
421 policies = policies[:nb_to_use]
424 desc = f"epoch-{split}"
425 for batch in tqdm.tqdm(
426 zip(input.split(self.batch_size), policies.split(self.batch_size)),
432 def vocabulary_size(self):
435 def compute_error(self, model, split="train", nb_to_use=-1):
436 nb_total, nb_correct = 0, 0
437 for input in task.batches(split, nb_to_use):
438 result = input.clone()
439 ar_mask = result.new_zeros(result.size())
440 ar_mask[:, self.height * self.width :] = 1
441 result *= 1 - ar_mask
442 x, order = shuffle(result, self.height * self.width)
443 masked_inplace_autoregression(
444 model, self.batch_size, x, ar_mask, order=order
446 result = reorder(x, order, reverse=True)
447 mazes, paths = self.seq2map(result)
448 nb_correct += maze.path_correctness(mazes, paths).long().sum()
449 nb_total += mazes.size(0)
451 return nb_total, nb_correct
453 def produce_results(self, n_epoch, model):
454 with torch.autograd.no_grad():
458 train_nb_total, train_nb_correct = self.compute_error(
459 model, "train", nb_to_use=1000
462 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
465 test_nb_total, test_nb_correct = self.compute_error(
466 model, "test", nb_to_use=1000
469 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
472 input = self.test_input[:32]
473 result = input.clone()
474 ar_mask = result.new_zeros(result.size())
475 ar_mask[:, self.height * self.width :] = 1
476 result *= 1 - ar_mask
477 x, order = shuffle(result, self.height * self.width)
478 masked_inplace_autoregression(
479 model, self.batch_size, x, ar_mask, order=order
481 result = reorder(x, order, reverse=True)
483 mazes, paths = self.seq2map(input)
484 _, predicted_paths = self.seq2map(result)
485 filename = f"result_{n_epoch:04d}.png"
487 os.path.join(args.result_dir, filename),
490 predicted_paths=predicted_paths,
491 path_correct=maze.path_correctness(mazes, predicted_paths),
493 log_string(f"wrote {filename}")
498 ######################################################################
500 log_string(f"device {device}")
504 nb_train_samples=args.nb_train_samples,
505 nb_test_samples=args.nb_test_samples,
506 batch_size=args.batch_size,
507 height=args.maze_height,
508 width=args.maze_width,
509 nb_walls=args.maze_nb_walls,
514 vocabulary_size = task.vocabulary_size()
516 log_string(f"vocabulary_size {vocabulary_size}")
518 ##############################
521 vocabulary_size=vocabulary_size,
522 dim_model=args.dim_model,
523 dim_keys=args.dim_keys,
524 dim_hidden=args.dim_hidden,
525 nb_heads=args.nb_heads,
526 nb_blocks=args.nb_blocks,
528 dropout=args.dropout,
533 nb_parameters = sum(p.numel() for p in model.parameters())
534 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
536 ######################################################################
538 nb_epochs_finished = 0
540 if args.no_checkpoint:
541 log_string(f"not trying to load checkpoint.")
545 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
546 checkpoint = torch.load(checkpoint_name)
547 nb_epochs_finished = checkpoint["nb_epochs_finished"]
548 model.load_state_dict(checkpoint["model_state"])
549 torch.set_rng_state(checkpoint["rng_state"])
550 if torch.cuda.is_available():
551 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
553 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
555 except FileNotFoundError:
556 log_string("starting from scratch.")
559 log_string("error when loading the checkpoint.")
562 ######################################################################
565 for input in task.batches(split="train"):
566 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
567 token_probas = token_count / token_count.sum()
568 entropy = -torch.xlogy(token_probas, token_probas).sum()
569 train_set_perplexity = math.exp(entropy)
571 ##############################
573 if args.learning_rate_schedule == "cos":
574 learning_rate_schedule = {}
575 for n_epoch in range(args.nb_epochs):
576 u = n_epoch / args.nb_epochs * math.pi
577 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
582 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
586 learning_rate_schedule = {}
587 learning_rate = args.learning_rate
588 for n_epoch in range(args.nb_epochs):
590 learning_rate = u[n_epoch]
591 learning_rate_schedule[n_epoch] = learning_rate
593 log_string(f"learning_rate_schedule {learning_rate_schedule}")
595 ##############################
597 if nb_epochs_finished >= args.nb_epochs:
598 n_epoch = nb_epochs_finished
599 train_perplexity = compute_perplexity(
600 model, task, fixed_len=task.height * task.width, split="train"
602 test_perplexity = compute_perplexity(
603 model, task, fixed_len=task.height * task.width, split="test"
607 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
610 task.produce_results(n_epoch, model)
612 ##############################
614 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
615 learning_rate = learning_rate_schedule[n_epoch]
617 log_string(f"learning_rate {learning_rate}")
619 if args.optim == "sgd":
620 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
621 elif args.optim == "adam":
622 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
623 elif args.optim == "adamw":
624 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
626 raise ValueError(f"{args.optim=}")
630 nb_train_samples, acc_train_loss = 0, 0.0
632 for input in task.batches(split="train"):
633 input = input.to(device)
635 model, input, mode=args.oneshot_input, fixed_len=task.height * task.width
637 loss = F.cross_entropy(output.transpose(1, 2), input)
638 acc_train_loss += loss.item() * input.size(0)
639 nb_train_samples += input.size(0)
641 optimizer.zero_grad()
645 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
646 test_perplexity = compute_perplexity(
647 model, task, fixed_len=task.height * task.width, split="test"
651 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
654 task.produce_results(n_epoch, model)
657 "nb_epochs_finished": n_epoch + 1,
658 "model_state": model.state_dict(),
659 "rng_state": torch.get_rng_state(),
662 if torch.cuda.is_available():
663 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
665 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
666 torch.save(checkpoint, checkpoint_name)
667 log_string(f"saved checkpoint {checkpoint_name}")
669 ######################################################################
674 ######################################################################