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(mazes, output, policies, height, width):
177 masks = (mazes == maze.v_empty).unsqueeze(-1)
178 targets = policies.permute(0, 2, 1) * masks
179 output = output * masks
180 return -(output.log_softmax(-1) * targets).sum() / masks.sum()
183 def oneshot_trace_loss(mazes, output, policies, height, width):
184 masks = mazes == maze.v_empty
185 targets = maze.stationary_densities(
186 mazes.view(-1, height, width), policies.view(-1, 4, height, width)
188 targets = targets * masks
189 output = output.squeeze(-1) * masks
190 return (output - targets).abs().sum() / masks.sum()
193 def oneshot(gpt, task):
197 if args.oneshot_input == "head":
198 dim_in = args.dim_model
199 elif args.oneshot_input == "deep":
200 dim_in = args.dim_model * args.nb_blocks * 2
202 raise ValueError(f"{args.oneshot_input=}")
204 if args.oneshot_output == "policy":
206 compute_loss = oneshot_policy_loss
207 elif args.oneshot_output == "trace":
209 compute_loss = oneshot_trace_loss
211 raise ValueError(f"{args.oneshot_output=}")
213 model = nn.Sequential(
214 nn.Linear(dim_in, args.dim_model),
216 nn.Linear(args.dim_model, args.dim_model),
218 nn.Linear(args.dim_model, dim_out),
221 for n_epoch in range(args.nb_epochs):
222 learning_rate = learning_rate_schedule[n_epoch]
223 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
225 acc_train_loss, nb_train_samples = 0, 0
226 for mazes, policies in task.policy_batches(split="train"):
228 # print(f'{mazes.size()=} {policies.size()=}')
229 # s = maze.stationary_densities(
232 output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
233 output = model(output_gpt)
235 loss = compute_loss(mazes, output, policies, task.height, task.width)
236 acc_train_loss += loss.item() * mazes.size(0)
237 nb_train_samples += mazes.size(0)
239 optimizer.zero_grad()
243 acc_test_loss, nb_test_samples = 0, 0
244 for mazes, policies in task.policy_batches(split="test"):
245 output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
246 output = model(output_gpt)
247 loss = compute_loss(mazes, output, policies, task.height, task.width)
248 acc_test_loss += loss.item() * mazes.size(0)
249 nb_test_samples += mazes.size(0)
252 f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
255 # -------------------
256 mazes = task.test_input[:32, : task.height * task.width]
257 policies = task.test_policies[:32]
258 output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
259 output = model(output_gpt)
260 if args.oneshot_output == "policy":
261 targets = policies.permute(0, 2, 1)
263 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
265 elif args.oneshot_output == "trace":
266 targets = maze.stationary_densities(
267 mazes.view(-1, task.height, task.width),
268 policies.view(-1, 4, task.height, task.width),
272 raise ValueError(f"{args.oneshot_output=}")
274 scores = scores.reshape(-1, task.height, task.width)
275 mazes = mazes.reshape(-1, task.height, task.width)
276 targets = targets.reshape(-1, task.height, task.width)
280 f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
286 # -------------------
291 ######################################################################
295 def batches(self, split="train"):
298 def vocabulary_size(self):
301 def produce_results(self, n_epoch, model):
305 ######################################################################
310 class TaskMaze(Task):
311 def map2seq(self, *m):
312 return torch.cat([x.flatten(1) for x in m], 1)
314 def seq2map(self, s):
315 s = s.reshape(s.size(0), -1, self.height, self.width)
316 return (s[:, k] for k in range(s.size(1)))
326 device=torch.device("cpu"),
328 self.batch_size = batch_size
333 train_mazes, train_paths, train_policies = maze.create_maze_data(
338 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
340 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
341 self.train_policies = train_policies.flatten(-2).to(device)
343 test_mazes, test_paths, test_policies = maze.create_maze_data(
348 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
350 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
351 self.test_policies = test_policies.flatten(-2).to(device)
353 self.nb_codes = self.train_input.max() + 1
355 def batches(self, split="train", nb_to_use=-1):
356 assert split in {"train", "test"}
357 input = self.train_input if split == "train" else self.test_input
359 input = input[:nb_to_use]
360 for batch in tqdm.tqdm(
361 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
365 def policy_batches(self, split="train", nb_to_use=-1):
366 assert split in {"train", "test"}
367 input = self.train_input if split == "train" else self.test_input
368 policies = self.train_policies if split == "train" else self.test_policies
369 input = input[:, : self.height * self.width]
370 policies = policies * (input != maze.v_wall)[:, None]
373 input = input[:nb_to_use]
374 policies = policies[:nb_to_use]
376 for batch in tqdm.tqdm(
377 zip(input.split(self.batch_size), policies.split(self.batch_size)),
379 desc=f"epoch-{split}",
383 def vocabulary_size(self):
386 def compute_error(self, model, split="train", nb_to_use=-1):
387 nb_total, nb_correct = 0, 0
388 for input in task.batches(split, nb_to_use):
389 result = input.clone()
390 ar_mask = result.new_zeros(result.size())
391 ar_mask[:, self.height * self.width :] = 1
392 result *= 1 - ar_mask
393 masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
394 mazes, paths = self.seq2map(result)
395 nb_correct += maze.path_correctness(mazes, paths).long().sum()
396 nb_total += mazes.size(0)
398 return nb_total, nb_correct
400 def produce_results(self, n_epoch, model):
401 with torch.autograd.no_grad():
405 train_nb_total, train_nb_correct = self.compute_error(
406 model, "train", nb_to_use=1000
409 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
412 test_nb_total, test_nb_correct = self.compute_error(
413 model, "test", nb_to_use=1000
416 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
419 input = self.test_input[:32]
420 result = input.clone()
421 ar_mask = result.new_zeros(result.size())
422 ar_mask[:, self.height * self.width :] = 1
423 result *= 1 - ar_mask
424 masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
426 mazes, paths = self.seq2map(input)
427 _, predicted_paths = self.seq2map(result)
429 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
432 predicted_paths=predicted_paths,
433 path_correct=maze.path_correctness(mazes, predicted_paths),
439 ######################################################################
441 log_string(f"device {device}")
445 nb_train_samples=args.nb_train_samples,
446 nb_test_samples=args.nb_test_samples,
447 batch_size=args.batch_size,
448 height=args.maze_height,
449 width=args.maze_width,
450 nb_walls=args.maze_nb_walls,
455 vocabulary_size = task.vocabulary_size()
457 log_string(f"vocabulary_size {vocabulary_size}")
459 ##############################
462 vocabulary_size=vocabulary_size,
463 dim_model=args.dim_model,
464 dim_keys=args.dim_keys,
465 dim_hidden=args.dim_hidden,
466 nb_heads=args.nb_heads,
467 nb_blocks=args.nb_blocks,
469 dropout=args.dropout,
474 nb_parameters = sum(p.numel() for p in model.parameters())
475 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
477 ######################################################################
479 nb_epochs_finished = 0
481 if args.no_checkpoint:
482 log_string(f"not trying to load checkpoint.")
486 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
487 checkpoint = torch.load(checkpoint_name)
488 nb_epochs_finished = checkpoint["nb_epochs_finished"]
489 model.load_state_dict(checkpoint["model_state"])
490 torch.set_rng_state(checkpoint["rng_state"])
491 if torch.cuda.is_available():
492 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
494 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
496 except FileNotFoundError:
497 log_string("starting from scratch.")
500 log_string("error when loading the checkpoint.")
503 ######################################################################
506 for input in task.batches(split="train"):
507 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
508 token_probas = token_count / token_count.sum()
509 entropy = -torch.xlogy(token_probas, token_probas).sum()
510 train_set_perplexity = math.exp(entropy)
512 ##############################
514 if args.learning_rate_schedule == "cos":
515 learning_rate_schedule = {}
516 for n_epoch in range(args.nb_epochs):
517 u = n_epoch / args.nb_epochs * math.pi
518 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
523 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
527 learning_rate_schedule = {}
528 learning_rate = args.learning_rate
529 for n_epoch in range(args.nb_epochs):
531 learning_rate = u[n_epoch]
532 learning_rate_schedule[n_epoch] = learning_rate
534 log_string(f"learning_rate_schedule {learning_rate_schedule}")
536 ##############################
542 ##############################
544 if nb_epochs_finished >= args.nb_epochs:
545 n_epoch = nb_epochs_finished
546 train_perplexity = compute_perplexity(model, split="train")
547 test_perplexity = compute_perplexity(model, split="test")
550 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
553 task.produce_results(n_epoch, model)
557 ##############################
559 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
560 learning_rate = learning_rate_schedule[n_epoch]
562 log_string(f"learning_rate {learning_rate}")
564 if args.optim == "sgd":
565 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
566 elif args.optim == "adam":
567 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
568 elif args.optim == "adamw":
569 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
571 raise ValueError(f"{args.optim=}")
575 nb_train_samples, acc_train_loss = 0, 0.0
577 for input in task.batches(split="train"):
578 input = input.to(device)
579 output = model(mygpt.BracketedSequence(input)).x
580 loss = F.cross_entropy(output.transpose(1, 2), input)
581 acc_train_loss += loss.item() * input.size(0)
582 nb_train_samples += input.size(0)
584 optimizer.zero_grad()
588 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
589 test_perplexity = compute_perplexity(model, split="test")
592 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
595 task.produce_results(n_epoch, model)
598 "nb_epochs_finished": n_epoch + 1,
599 "model_state": model.state_dict(),
600 "rng_state": torch.get_rng_state(),
603 if torch.cuda.is_available():
604 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
606 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
607 torch.save(checkpoint, checkpoint_name)
608 log_string(f"saved checkpoint {checkpoint_name}")
610 ######################################################################