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 parser.add_argument("--oneshot", action="store_true", default=False)
84 parser.add_argument("--oneshot_input", type=str, default="head")
86 parser.add_argument("--oneshot_output", type=str, default="policy")
88 ######################################################################
90 args = parser.parse_args()
93 os.mkdir(args.result_dir)
94 except FileExistsError:
95 if not args.overwrite_results:
96 print(f"result directory {args.result_dir} already exists")
99 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
102 # torch.backends.cudnn.deterministic = True
103 # torch.backends.cudnn.benchmark = False
104 # torch.use_deterministic_algorithms(True)
105 torch.manual_seed(args.seed)
106 if torch.cuda.is_available():
107 torch.cuda.manual_seed_all(args.seed)
109 ######################################################################
113 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
115 if log_file is not None:
116 log_file.write(t + s + "\n")
124 log_string(f"args.{n} {getattr(args, n)}")
126 ######################################################################
129 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
130 # tokens that should be generated
133 def masked_inplace_autoregression(model, batch_size, input, ar_mask):
134 for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
135 i = (ar_mask.sum(0) > 0).nonzero()
137 # Needed to initialize the model's cache
138 model(mygpt.BracketedSequence(input, 0, i.min()))
139 for s in range(i.min(), i.max() + 1):
140 output = model(mygpt.BracketedSequence(input, s, 1)).x
141 logits = output[:, s]
142 if args.deterministic_synthesis:
143 t_next = logits.argmax(1)
145 dist = torch.distributions.categorical.Categorical(logits=logits)
146 t_next = dist.sample()
147 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
150 ######################################################################
153 def compute_perplexity(model, split="train"):
154 with torch.autograd.no_grad():
158 nb_samples, acc_loss = 0, 0.0
160 for input in task.batches(split=split):
161 input = input.to(device)
163 output = model(mygpt.BracketedSequence(input)).x
164 loss = F.cross_entropy(output.transpose(1, 2), input)
165 acc_loss += loss.item() * input.size(0)
166 nb_samples += input.size(0)
170 return math.exp(min(100, acc_loss / nb_samples))
173 ######################################################################
176 def oneshot_policy_loss(output, policies, mask):
177 targets = policies.permute(0, 2, 1) * mask.unsqueeze(-1)
178 output = output * mask.unsqueeze(-1)
179 return -(output.log_softmax(-1) * targets).sum() / mask.sum()
182 # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
185 def oneshot(gpt, task):
189 if args.oneshot_input == "head":
190 dim_in = args.dim_model
191 elif args.oneshot_input == "deep":
192 dim_in = args.dim_model * args.nb_blocks * 2
194 raise ValueError(f"{args.oneshot_input=}")
196 if args.oneshot_output == "policy":
198 compute_loss = oneshot_policy_loss
199 elif args.oneshot_output == "trace":
202 raise ValueError(f"{args.oneshot_output=}")
204 model = nn.Sequential(
205 nn.Linear(dim_in, args.dim_model),
207 nn.Linear(args.dim_model, args.dim_model),
209 nn.Linear(args.dim_model, 4),
212 for n_epoch in range(args.nb_epochs):
213 learning_rate = learning_rate_schedule[n_epoch]
214 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
216 acc_train_loss, nb_train_samples = 0, 0
217 for input, policies in task.policy_batches(split="train"):
219 # print(f'{input.size()=} {policies.size()=}')
220 # s = maze.stationary_densities(
223 mask = input == maze.v_empty
224 output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_input).x
225 output = model(output_gpt)
227 loss = compute_loss(output, policies, mask)
228 acc_train_loss += loss.item() * input.size(0)
229 nb_train_samples += input.size(0)
231 optimizer.zero_grad()
235 acc_test_loss, nb_test_samples = 0, 0
236 for input, policies in task.policy_batches(split="test"):
237 mask = input == maze.v_empty
238 output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_input).x
239 output = model(output_gpt)
240 loss = compute_loss(output, policies, mask)
241 acc_test_loss += loss.item() * input.size(0)
242 nb_test_samples += input.size(0)
245 f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
248 # -------------------
249 input = task.test_input[:32, : task.height * task.width]
250 targets = task.test_policies[:32].permute(0, 2, 1)
251 output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_input).x
252 output = model(output_gpt)
254 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
256 scores = scores.reshape(-1, task.height, task.width)
257 input = input.reshape(-1, task.height, task.width)
261 f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
266 # -------------------
271 ######################################################################
275 def batches(self, split="train"):
278 def vocabulary_size(self):
281 def produce_results(self, n_epoch, model):
285 ######################################################################
290 class TaskMaze(Task):
291 def map2seq(self, *m):
292 return torch.cat([x.flatten(1) for x in m], 1)
294 def seq2map(self, s):
295 s = s.reshape(s.size(0), -1, self.height, self.width)
296 return (s[:, k] for k in range(s.size(1)))
306 device=torch.device("cpu"),
308 self.batch_size = batch_size
313 train_mazes, train_paths, train_policies = maze.create_maze_data(
318 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
320 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
321 self.train_policies = train_policies.flatten(-2).to(device)
323 test_mazes, test_paths, test_policies = maze.create_maze_data(
328 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
330 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
331 self.test_policies = test_policies.flatten(-2).to(device)
333 self.nb_codes = self.train_input.max() + 1
335 def batches(self, split="train", nb_to_use=-1):
336 assert split in {"train", "test"}
337 input = self.train_input if split == "train" else self.test_input
339 input = input[:nb_to_use]
340 for batch in tqdm.tqdm(
341 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
345 def policy_batches(self, split="train", nb_to_use=-1):
346 assert split in {"train", "test"}
347 input = self.train_input if split == "train" else self.test_input
348 policies = self.train_policies if split == "train" else self.test_policies
349 input = input[:, : self.height * self.width]
350 policies = policies * (input != maze.v_wall)[:, None]
353 input = input[:nb_to_use]
354 policies = policies[:nb_to_use]
356 for batch in tqdm.tqdm(
357 zip(input.split(self.batch_size), policies.split(self.batch_size)),
359 desc=f"epoch-{split}",
363 def vocabulary_size(self):
366 def compute_error(self, model, split="train", nb_to_use=-1):
367 nb_total, nb_correct = 0, 0
368 for input in task.batches(split, nb_to_use):
369 result = input.clone()
370 ar_mask = result.new_zeros(result.size())
371 ar_mask[:, self.height * self.width :] = 1
372 result *= 1 - ar_mask
373 masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
374 mazes, paths = self.seq2map(result)
375 nb_correct += maze.path_correctness(mazes, paths).long().sum()
376 nb_total += mazes.size(0)
378 return nb_total, nb_correct
380 def produce_results(self, n_epoch, model):
381 with torch.autograd.no_grad():
385 train_nb_total, train_nb_correct = self.compute_error(
386 model, "train", nb_to_use=1000
389 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
392 test_nb_total, test_nb_correct = self.compute_error(
393 model, "test", nb_to_use=1000
396 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
399 input = self.test_input[:32]
400 result = input.clone()
401 ar_mask = result.new_zeros(result.size())
402 ar_mask[:, self.height * self.width :] = 1
403 result *= 1 - ar_mask
404 masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
406 mazes, paths = self.seq2map(input)
407 _, predicted_paths = self.seq2map(result)
409 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
412 predicted_paths=predicted_paths,
413 path_correct=maze.path_correctness(mazes, predicted_paths),
419 ######################################################################
421 log_string(f"device {device}")
425 nb_train_samples=args.nb_train_samples,
426 nb_test_samples=args.nb_test_samples,
427 batch_size=args.batch_size,
428 height=args.maze_height,
429 width=args.maze_width,
430 nb_walls=args.maze_nb_walls,
435 vocabulary_size = task.vocabulary_size()
437 log_string(f"vocabulary_size {vocabulary_size}")
439 ##############################
442 vocabulary_size=vocabulary_size,
443 dim_model=args.dim_model,
444 dim_keys=args.dim_keys,
445 dim_hidden=args.dim_hidden,
446 nb_heads=args.nb_heads,
447 nb_blocks=args.nb_blocks,
449 dropout=args.dropout,
454 nb_parameters = sum(p.numel() for p in model.parameters())
455 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
457 ######################################################################
459 nb_epochs_finished = 0
461 if args.no_checkpoint:
462 log_string(f"not trying to load checkpoint.")
466 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
467 checkpoint = torch.load(checkpoint_name)
468 nb_epochs_finished = checkpoint["nb_epochs_finished"]
469 model.load_state_dict(checkpoint["model_state"])
470 torch.set_rng_state(checkpoint["rng_state"])
471 if torch.cuda.is_available():
472 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
474 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
476 except FileNotFoundError:
477 log_string("starting from scratch.")
480 log_string("error when loading the checkpoint.")
483 ######################################################################
486 for input in task.batches(split="train"):
487 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
488 token_probas = token_count / token_count.sum()
489 entropy = -torch.xlogy(token_probas, token_probas).sum()
490 train_set_perplexity = math.exp(entropy)
492 ##############################
494 if args.learning_rate_schedule == "cos":
495 learning_rate_schedule = {}
496 for n_epoch in range(args.nb_epochs):
497 u = n_epoch / args.nb_epochs * math.pi
498 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
503 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
507 learning_rate_schedule = {}
508 learning_rate = args.learning_rate
509 for n_epoch in range(args.nb_epochs):
511 learning_rate = u[n_epoch]
512 learning_rate_schedule[n_epoch] = learning_rate
514 log_string(f"learning_rate_schedule {learning_rate_schedule}")
516 ##############################
522 ##############################
524 if nb_epochs_finished >= args.nb_epochs:
525 n_epoch = nb_epochs_finished
526 train_perplexity = compute_perplexity(model, split="train")
527 test_perplexity = compute_perplexity(model, split="test")
530 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
533 task.produce_results(n_epoch, model)
537 ##############################
539 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
540 learning_rate = learning_rate_schedule[n_epoch]
542 log_string(f"learning_rate {learning_rate}")
544 if args.optim == "sgd":
545 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
546 elif args.optim == "adam":
547 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
548 elif args.optim == "adamw":
549 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
551 raise ValueError(f"{args.optim=}")
555 nb_train_samples, acc_train_loss = 0, 0.0
557 for input in task.batches(split="train"):
558 input = input.to(device)
559 output = model(mygpt.BracketedSequence(input)).x
560 loss = F.cross_entropy(output.transpose(1, 2), input)
561 acc_train_loss += loss.item() * input.size(0)
562 nb_train_samples += input.size(0)
564 optimizer.zero_grad()
568 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
569 test_perplexity = compute_perplexity(model, split="test")
572 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
575 task.produce_results(n_epoch, model)
578 "nb_epochs_finished": n_epoch + 1,
579 "model_state": model.state_dict(),
580 "rng_state": torch.get_rng_state(),
583 if torch.cuda.is_available():
584 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
586 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
587 torch.save(checkpoint, checkpoint_name)
588 log_string(f"saved checkpoint {checkpoint_name}")
590 ######################################################################