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 random_order(result, fixed_len):
135 if args.random_regression_order:
136 order = torch.rand(result.size(), device=result.device)
137 order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=order.device)
138 return order.sort(1).indices
140 return torch.arange(result.size(1)).unsqueeze(0).expand(result.size(0), -1)
143 def shuffle(x, order, reorder=False):
145 order = order.unsqueeze(-1).expand(-1, -1, x.size(-1))
148 y.scatter_(1, order, x)
151 return x.gather(1, order)
154 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
155 # tokens that should be generated
158 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
159 for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
160 i = (ar_mask.sum(0) > 0).nonzero()
162 # Needed to initialize the model's cache
163 model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
164 for s in range(i.min(), i.max() + 1):
165 output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
166 logits = output[:, s]
167 if args.deterministic_synthesis:
168 t_next = logits.argmax(1)
170 dist = torch.distributions.categorical.Categorical(logits=logits)
171 t_next = dist.sample()
172 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
175 ######################################################################
178 def compute_perplexity(model, split="train"):
179 with torch.autograd.no_grad():
183 nb_samples, acc_loss = 0, 0.0
185 for input in task.batches(split=split):
186 input = input.to(device)
187 order = random_order(input, task.height * task.width)
188 input = shuffle(input, order)
189 output = model(mygpt.BracketedSequence(input), order=order).x
190 loss = F.cross_entropy(output.transpose(1, 2), input)
191 acc_loss += loss.item() * input.size(0)
192 nb_samples += input.size(0)
196 return math.exp(min(100, acc_loss / nb_samples))
199 ######################################################################
202 def oneshot_policy_loss(mazes, output, policies, height, width):
203 masks = (mazes == maze.v_empty).unsqueeze(-1)
204 targets = policies.permute(0, 2, 1) * masks
205 output = output * masks
206 return -(output.log_softmax(-1) * targets).sum() / masks.sum()
209 def oneshot_trace_loss(mazes, output, policies, height, width):
210 masks = mazes == maze.v_empty
211 targets = maze.stationary_densities(
212 mazes.view(-1, height, width), policies.view(-1, 4, height, width)
214 targets = targets * masks
215 output = output.squeeze(-1) * masks
216 return (output - targets).abs().sum() / masks.sum()
219 def oneshot(gpt, task):
223 if args.oneshot_input == "head":
224 dim_in = args.dim_model
225 elif args.oneshot_input == "deep":
226 dim_in = args.dim_model * args.nb_blocks * 2
228 raise ValueError(f"{args.oneshot_input=}")
230 if args.oneshot_output == "policy":
232 compute_loss = oneshot_policy_loss
233 elif args.oneshot_output == "trace":
235 compute_loss = oneshot_trace_loss
237 raise ValueError(f"{args.oneshot_output=}")
239 model = nn.Sequential(
240 nn.Linear(dim_in, args.dim_model),
242 nn.Linear(args.dim_model, args.dim_model),
244 nn.Linear(args.dim_model, dim_out),
247 for n_epoch in range(args.nb_epochs):
248 learning_rate = learning_rate_schedule[n_epoch]
249 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
251 acc_train_loss, nb_train_samples = 0, 0
252 for mazes, policies in task.policy_batches(split="train"):
253 order = random_order(mazes, task.height * task.width)
254 x = shuffle(mazes, order)
255 x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
256 output_gpt = shuffle(x, order, reorder=True)
257 output = model(output_gpt)
259 loss = compute_loss(mazes, output, policies, task.height, task.width)
260 acc_train_loss += loss.item() * mazes.size(0)
261 nb_train_samples += mazes.size(0)
263 optimizer.zero_grad()
267 acc_test_loss, nb_test_samples = 0, 0
268 for mazes, policies in task.policy_batches(split="test"):
269 order = random_order(mazes, task.height * task.width)
270 x = shuffle(mazes, order)
271 x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
272 output_gpt = shuffle(x, order, reorder=True)
273 output = model(output_gpt)
274 loss = compute_loss(mazes, output, policies, task.height, task.width)
275 acc_test_loss += loss.item() * mazes.size(0)
276 nb_test_samples += mazes.size(0)
279 f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
282 # -------------------
283 mazes = task.test_input[:32, : task.height * task.width]
284 policies = task.test_policies[:32]
285 order = random_order(mazes, task.height * task.width)
286 x = shuffle(mazes, order)
287 x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
288 output_gpt = shuffle(x, order, reorder=True)
289 output = model(output_gpt)
290 if args.oneshot_output == "policy":
291 targets = policies.permute(0, 2, 1)
293 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
295 elif args.oneshot_output == "trace":
296 targets = maze.stationary_densities(
297 mazes.view(-1, task.height, task.width),
298 policies.view(-1, 4, task.height, task.width),
302 raise ValueError(f"{args.oneshot_output=}")
304 scores = scores.reshape(-1, task.height, task.width)
305 mazes = mazes.reshape(-1, task.height, task.width)
306 targets = targets.reshape(-1, task.height, task.width)
310 f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
316 # -------------------
321 ######################################################################
325 def batches(self, split="train", nb_to_use=-1, desc=None):
328 def vocabulary_size(self):
331 def produce_results(self, n_epoch, model):
335 ######################################################################
340 class TaskMaze(Task):
341 def map2seq(self, *m):
342 return torch.cat([x.flatten(1) for x in m], 1)
344 def seq2map(self, s):
345 s = s.reshape(s.size(0), -1, self.height, self.width)
346 return (s[:, k] for k in range(s.size(1)))
356 device=torch.device("cpu"),
358 self.batch_size = batch_size
363 train_mazes, train_paths, train_policies = maze.create_maze_data(
368 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
370 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
371 self.train_policies = train_policies.flatten(-2).to(device)
373 test_mazes, test_paths, test_policies = maze.create_maze_data(
378 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
380 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
381 self.test_policies = test_policies.flatten(-2).to(device)
383 self.nb_codes = self.train_input.max() + 1
385 def batches(self, split="train", nb_to_use=-1, desc=None):
386 assert split in {"train", "test"}
387 input = self.train_input if split == "train" else self.test_input
389 input = input[:nb_to_use]
391 desc = f"epoch-{split}"
392 for batch in tqdm.tqdm(
393 input.split(self.batch_size), dynamic_ncols=True, desc=desc
397 def policy_batches(self, split="train", nb_to_use=-1, desc=None):
398 assert split in {"train", "test"}
399 input = self.train_input if split == "train" else self.test_input
400 policies = self.train_policies if split == "train" else self.test_policies
401 input = input[:, : self.height * self.width]
402 policies = policies * (input != maze.v_wall)[:, None]
405 input = input[:nb_to_use]
406 policies = policies[:nb_to_use]
409 desc = f"epoch-{split}"
410 for batch in tqdm.tqdm(
411 zip(input.split(self.batch_size), policies.split(self.batch_size)),
417 def vocabulary_size(self):
420 def compute_error(self, model, split="train", nb_to_use=-1):
421 nb_total, nb_correct = 0, 0
422 for input in task.batches(split, nb_to_use):
423 result = input.clone()
424 ar_mask = result.new_zeros(result.size())
425 ar_mask[:, self.height * self.width :] = 1
426 result *= 1 - ar_mask
427 order = random_order(result, self.height * self.width)
428 masked_inplace_autoregression(
429 model, self.batch_size, result, ar_mask, order=order
431 result = shuffle(result, order, reorder=True)
432 mazes, paths = self.seq2map(result)
433 nb_correct += maze.path_correctness(mazes, paths).long().sum()
434 nb_total += mazes.size(0)
436 return nb_total, nb_correct
438 def produce_results(self, n_epoch, model):
439 with torch.autograd.no_grad():
443 train_nb_total, train_nb_correct = self.compute_error(
444 model, "train", nb_to_use=1000
447 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
450 test_nb_total, test_nb_correct = self.compute_error(
451 model, "test", nb_to_use=1000
454 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
457 input = self.test_input[:32]
458 result = input.clone()
459 ar_mask = result.new_zeros(result.size())
460 ar_mask[:, self.height * self.width :] = 1
461 result *= 1 - ar_mask
462 masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
464 mazes, paths = self.seq2map(input)
465 _, predicted_paths = self.seq2map(result)
467 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
470 predicted_paths=predicted_paths,
471 path_correct=maze.path_correctness(mazes, predicted_paths),
477 ######################################################################
479 log_string(f"device {device}")
483 nb_train_samples=args.nb_train_samples,
484 nb_test_samples=args.nb_test_samples,
485 batch_size=args.batch_size,
486 height=args.maze_height,
487 width=args.maze_width,
488 nb_walls=args.maze_nb_walls,
493 vocabulary_size = task.vocabulary_size()
495 log_string(f"vocabulary_size {vocabulary_size}")
497 ##############################
500 vocabulary_size=vocabulary_size,
501 dim_model=args.dim_model,
502 dim_keys=args.dim_keys,
503 dim_hidden=args.dim_hidden,
504 nb_heads=args.nb_heads,
505 nb_blocks=args.nb_blocks,
507 dropout=args.dropout,
512 nb_parameters = sum(p.numel() for p in model.parameters())
513 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
515 ######################################################################
517 nb_epochs_finished = 0
519 if args.no_checkpoint:
520 log_string(f"not trying to load checkpoint.")
524 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
525 checkpoint = torch.load(checkpoint_name)
526 nb_epochs_finished = checkpoint["nb_epochs_finished"]
527 model.load_state_dict(checkpoint["model_state"])
528 torch.set_rng_state(checkpoint["rng_state"])
529 if torch.cuda.is_available():
530 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
532 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
534 except FileNotFoundError:
535 log_string("starting from scratch.")
538 log_string("error when loading the checkpoint.")
541 ######################################################################
544 for input in task.batches(split="train"):
545 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
546 token_probas = token_count / token_count.sum()
547 entropy = -torch.xlogy(token_probas, token_probas).sum()
548 train_set_perplexity = math.exp(entropy)
550 ##############################
552 if args.learning_rate_schedule == "cos":
553 learning_rate_schedule = {}
554 for n_epoch in range(args.nb_epochs):
555 u = n_epoch / args.nb_epochs * math.pi
556 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
561 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
565 learning_rate_schedule = {}
566 learning_rate = args.learning_rate
567 for n_epoch in range(args.nb_epochs):
569 learning_rate = u[n_epoch]
570 learning_rate_schedule[n_epoch] = learning_rate
572 log_string(f"learning_rate_schedule {learning_rate_schedule}")
574 ##############################
576 if nb_epochs_finished >= args.nb_epochs:
577 n_epoch = nb_epochs_finished
578 train_perplexity = compute_perplexity(model, split="train")
579 test_perplexity = compute_perplexity(model, split="test")
582 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
585 task.produce_results(n_epoch, model)
587 ##############################
589 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
590 learning_rate = learning_rate_schedule[n_epoch]
592 log_string(f"learning_rate {learning_rate}")
594 if args.optim == "sgd":
595 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
596 elif args.optim == "adam":
597 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
598 elif args.optim == "adamw":
599 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
601 raise ValueError(f"{args.optim=}")
605 nb_train_samples, acc_train_loss = 0, 0.0
607 for input in task.batches(split="train"):
608 input = input.to(device)
609 order = random_order(input, task.height * task.width)
610 input = shuffle(input, order)
611 output = model(mygpt.BracketedSequence(input), order=order).x
612 loss = F.cross_entropy(output.transpose(1, 2), input)
613 acc_train_loss += loss.item() * input.size(0)
614 nb_train_samples += input.size(0)
616 optimizer.zero_grad()
620 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
621 test_perplexity = compute_perplexity(model, split="test")
624 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
627 task.produce_results(n_epoch, model)
630 "nb_epochs_finished": n_epoch + 1,
631 "model_state": model.state_dict(),
632 "rng_state": torch.get_rng_state(),
635 if torch.cuda.is_available():
636 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
638 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
639 torch.save(checkpoint, checkpoint_name)
640 log_string(f"saved checkpoint {checkpoint_name}")
642 ######################################################################
647 ######################################################################