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("--no_checkpoint", action="store_true", default=False)
69 parser.add_argument("--overwrite_results", action="store_true", default=False)
71 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
73 ##############################
76 parser.add_argument("--maze_height", type=int, default=13)
78 parser.add_argument("--maze_width", type=int, default=21)
80 parser.add_argument("--maze_nb_walls", type=int, default=15)
82 ##############################
85 parser.add_argument("--oneshot", action="store_true", default=False)
87 parser.add_argument("--oneshot_input", type=str, default="head")
89 parser.add_argument("--oneshot_output", type=str, default="policy")
91 ######################################################################
93 args = parser.parse_args()
96 os.mkdir(args.result_dir)
97 except FileExistsError:
98 if not args.overwrite_results:
99 print(f"result directory {args.result_dir} already exists")
102 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
105 # torch.backends.cudnn.deterministic = True
106 # torch.backends.cudnn.benchmark = False
107 # torch.use_deterministic_algorithms(True)
108 torch.manual_seed(args.seed)
109 if torch.cuda.is_available():
110 torch.cuda.manual_seed_all(args.seed)
112 ######################################################################
116 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
118 if log_file is not None:
119 log_file.write(t + s + "\n")
127 log_string(f"args.{n} {getattr(args, n)}")
129 ######################################################################
132 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
133 # tokens that should be generated
136 def masked_inplace_autoregression(model, batch_size, input, ar_mask):
137 for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
138 i = (ar_mask.sum(0) > 0).nonzero()
140 # Needed to initialize the model's cache
141 model(mygpt.BracketedSequence(input, 0, i.min()))
142 for s in range(i.min(), i.max() + 1):
143 output = model(mygpt.BracketedSequence(input, s, 1)).x
144 logits = output[:, s]
145 if args.deterministic_synthesis:
146 t_next = logits.argmax(1)
148 dist = torch.distributions.categorical.Categorical(logits=logits)
149 t_next = dist.sample()
150 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
153 ######################################################################
156 def compute_perplexity(model, split="train"):
157 with torch.autograd.no_grad():
161 nb_samples, acc_loss = 0, 0.0
163 for input in task.batches(split=split):
164 input = input.to(device)
166 output = model(mygpt.BracketedSequence(input)).x
167 loss = F.cross_entropy(output.transpose(1, 2), input)
168 acc_loss += loss.item() * input.size(0)
169 nb_samples += input.size(0)
173 return math.exp(min(100, acc_loss / nb_samples))
176 ######################################################################
179 def oneshot_policy_loss(mazes, output, policies, height, width):
180 masks = (mazes == maze.v_empty).unsqueeze(-1)
181 targets = policies.permute(0, 2, 1) * masks
182 output = output * masks
183 return -(output.log_softmax(-1) * targets).sum() / masks.sum()
186 def oneshot_trace_loss(mazes, output, policies, height, width):
187 masks = mazes == maze.v_empty
188 targets = maze.stationary_densities(
189 mazes.view(-1, height, width), policies.view(-1, 4, height, width)
191 targets = targets * masks
192 output = output.squeeze(-1) * masks
193 return (output - targets).abs().sum() / masks.sum()
196 def oneshot(gpt, task):
200 if args.oneshot_input == "head":
201 dim_in = args.dim_model
202 elif args.oneshot_input == "deep":
203 dim_in = args.dim_model * args.nb_blocks * 2
205 raise ValueError(f"{args.oneshot_input=}")
207 if args.oneshot_output == "policy":
209 compute_loss = oneshot_policy_loss
210 elif args.oneshot_output == "trace":
212 compute_loss = oneshot_trace_loss
214 raise ValueError(f"{args.oneshot_output=}")
216 model = nn.Sequential(
217 nn.Linear(dim_in, args.dim_model),
219 nn.Linear(args.dim_model, args.dim_model),
221 nn.Linear(args.dim_model, dim_out),
224 for n_epoch in range(args.nb_epochs):
225 learning_rate = learning_rate_schedule[n_epoch]
226 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
228 acc_train_loss, nb_train_samples = 0, 0
229 for mazes, policies in task.policy_batches(split="train"):
230 output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
231 output = model(output_gpt)
233 loss = compute_loss(mazes, output, policies, task.height, task.width)
234 acc_train_loss += loss.item() * mazes.size(0)
235 nb_train_samples += mazes.size(0)
237 optimizer.zero_grad()
241 acc_test_loss, nb_test_samples = 0, 0
242 for mazes, policies in task.policy_batches(split="test"):
243 output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
244 output = model(output_gpt)
245 loss = compute_loss(mazes, output, policies, task.height, task.width)
246 acc_test_loss += loss.item() * mazes.size(0)
247 nb_test_samples += mazes.size(0)
250 f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
253 # -------------------
254 mazes = task.test_input[:32, : task.height * task.width]
255 policies = task.test_policies[:32]
256 output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
257 output = model(output_gpt)
258 if args.oneshot_output == "policy":
259 targets = policies.permute(0, 2, 1)
261 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
263 elif args.oneshot_output == "trace":
264 targets = maze.stationary_densities(
265 mazes.view(-1, task.height, task.width),
266 policies.view(-1, 4, task.height, task.width),
270 raise ValueError(f"{args.oneshot_output=}")
272 scores = scores.reshape(-1, task.height, task.width)
273 mazes = mazes.reshape(-1, task.height, task.width)
274 targets = targets.reshape(-1, task.height, task.width)
278 f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
284 # -------------------
289 ######################################################################
293 def batches(self, split="train"):
296 def vocabulary_size(self):
299 def produce_results(self, n_epoch, model):
303 ######################################################################
308 class TaskMaze(Task):
309 def map2seq(self, *m):
310 return torch.cat([x.flatten(1) for x in m], 1)
312 def seq2map(self, s):
313 s = s.reshape(s.size(0), -1, self.height, self.width)
314 return (s[:, k] for k in range(s.size(1)))
324 device=torch.device("cpu"),
326 self.batch_size = batch_size
331 train_mazes, train_paths, train_policies = maze.create_maze_data(
336 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
338 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
339 self.train_policies = train_policies.flatten(-2).to(device)
341 test_mazes, test_paths, test_policies = maze.create_maze_data(
346 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
348 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
349 self.test_policies = test_policies.flatten(-2).to(device)
351 self.nb_codes = self.train_input.max() + 1
353 def batches(self, split="train", nb_to_use=-1):
354 assert split in {"train", "test"}
355 input = self.train_input if split == "train" else self.test_input
357 input = input[:nb_to_use]
358 for batch in tqdm.tqdm(
359 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
363 def policy_batches(self, split="train", nb_to_use=-1):
364 assert split in {"train", "test"}
365 input = self.train_input if split == "train" else self.test_input
366 policies = self.train_policies if split == "train" else self.test_policies
367 input = input[:, : self.height * self.width]
368 policies = policies * (input != maze.v_wall)[:, None]
371 input = input[:nb_to_use]
372 policies = policies[:nb_to_use]
374 for batch in tqdm.tqdm(
375 zip(input.split(self.batch_size), policies.split(self.batch_size)),
377 desc=f"epoch-{split}",
381 def vocabulary_size(self):
384 def compute_error(self, model, split="train", nb_to_use=-1):
385 nb_total, nb_correct = 0, 0
386 for input in task.batches(split, nb_to_use):
387 result = input.clone()
388 ar_mask = result.new_zeros(result.size())
389 ar_mask[:, self.height * self.width :] = 1
390 result *= 1 - ar_mask
391 masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
392 mazes, paths = self.seq2map(result)
393 nb_correct += maze.path_correctness(mazes, paths).long().sum()
394 nb_total += mazes.size(0)
396 return nb_total, nb_correct
398 def produce_results(self, n_epoch, model):
399 with torch.autograd.no_grad():
403 train_nb_total, train_nb_correct = self.compute_error(
404 model, "train", nb_to_use=1000
407 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
410 test_nb_total, test_nb_correct = self.compute_error(
411 model, "test", nb_to_use=1000
414 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
417 input = self.test_input[:32]
418 result = input.clone()
419 ar_mask = result.new_zeros(result.size())
420 ar_mask[:, self.height * self.width :] = 1
421 result *= 1 - ar_mask
422 masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
424 mazes, paths = self.seq2map(input)
425 _, predicted_paths = self.seq2map(result)
427 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
430 predicted_paths=predicted_paths,
431 path_correct=maze.path_correctness(mazes, predicted_paths),
437 ######################################################################
439 log_string(f"device {device}")
443 nb_train_samples=args.nb_train_samples,
444 nb_test_samples=args.nb_test_samples,
445 batch_size=args.batch_size,
446 height=args.maze_height,
447 width=args.maze_width,
448 nb_walls=args.maze_nb_walls,
453 vocabulary_size = task.vocabulary_size()
455 log_string(f"vocabulary_size {vocabulary_size}")
457 ##############################
460 vocabulary_size=vocabulary_size,
461 dim_model=args.dim_model,
462 dim_keys=args.dim_keys,
463 dim_hidden=args.dim_hidden,
464 nb_heads=args.nb_heads,
465 nb_blocks=args.nb_blocks,
467 dropout=args.dropout,
472 nb_parameters = sum(p.numel() for p in model.parameters())
473 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
475 ######################################################################
477 nb_epochs_finished = 0
479 if args.no_checkpoint:
480 log_string(f"not trying to load checkpoint.")
484 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
485 checkpoint = torch.load(checkpoint_name)
486 nb_epochs_finished = checkpoint["nb_epochs_finished"]
487 model.load_state_dict(checkpoint["model_state"])
488 torch.set_rng_state(checkpoint["rng_state"])
489 if torch.cuda.is_available():
490 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
492 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
494 except FileNotFoundError:
495 log_string("starting from scratch.")
498 log_string("error when loading the checkpoint.")
501 ######################################################################
504 for input in task.batches(split="train"):
505 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
506 token_probas = token_count / token_count.sum()
507 entropy = -torch.xlogy(token_probas, token_probas).sum()
508 train_set_perplexity = math.exp(entropy)
510 ##############################
512 if args.learning_rate_schedule == "cos":
513 learning_rate_schedule = {}
514 for n_epoch in range(args.nb_epochs):
515 u = n_epoch / args.nb_epochs * math.pi
516 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
521 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
525 learning_rate_schedule = {}
526 learning_rate = args.learning_rate
527 for n_epoch in range(args.nb_epochs):
529 learning_rate = u[n_epoch]
530 learning_rate_schedule[n_epoch] = learning_rate
532 log_string(f"learning_rate_schedule {learning_rate_schedule}")
534 ##############################
536 if nb_epochs_finished >= args.nb_epochs:
537 n_epoch = nb_epochs_finished
538 train_perplexity = compute_perplexity(model, split="train")
539 test_perplexity = compute_perplexity(model, split="test")
542 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
545 task.produce_results(n_epoch, model)
549 ##############################
551 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
552 learning_rate = learning_rate_schedule[n_epoch]
554 log_string(f"learning_rate {learning_rate}")
556 if args.optim == "sgd":
557 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
558 elif args.optim == "adam":
559 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
560 elif args.optim == "adamw":
561 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
563 raise ValueError(f"{args.optim=}")
567 nb_train_samples, acc_train_loss = 0, 0.0
569 for input in task.batches(split="train"):
570 input = input.to(device)
571 output = model(mygpt.BracketedSequence(input)).x
572 loss = F.cross_entropy(output.transpose(1, 2), input)
573 acc_train_loss += loss.item() * input.size(0)
574 nb_train_samples += input.size(0)
576 optimizer.zero_grad()
580 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
581 test_perplexity = compute_perplexity(model, split="test")
584 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
587 task.produce_results(n_epoch, model)
590 "nb_epochs_finished": n_epoch + 1,
591 "model_state": model.state_dict(),
592 "rng_state": torch.get_rng_state(),
595 if torch.cuda.is_available():
596 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
598 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
599 torch.save(checkpoint, checkpoint_name)
600 log_string(f"saved checkpoint {checkpoint_name}")
602 ######################################################################
607 ######################################################################