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, os
13 import torch, torchvision
15 from torch.nn import functional as F
20 ######################################################################
22 if torch.cuda.is_available():
23 device = torch.device("cuda")
24 torch.backends.cuda.matmul.allow_tf32 = True
26 device = torch.device("cpu")
28 ######################################################################
30 parser = argparse.ArgumentParser(
31 description="An implementation of GPT with cache.",
32 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
39 help="sandbox, picoclvr, mnist, maze, snake, stack, expr, world",
42 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
44 parser.add_argument("--result_dir", type=str, default=None)
46 parser.add_argument("--seed", type=int, default=0)
48 parser.add_argument("--nb_epochs", type=int, default=None)
50 parser.add_argument("--batch_size", type=int, default=None)
52 parser.add_argument("--nb_train_samples", type=int, default=None)
54 parser.add_argument("--nb_test_samples", type=int, default=None)
56 parser.add_argument("--optim", type=str, default="adam")
58 parser.add_argument("--learning_rate", type=float, default=1e-4)
60 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
62 parser.add_argument("--model", type=str, default="37M")
64 parser.add_argument("--dim_model", type=int, default=None)
66 parser.add_argument("--dim_keys", type=int, default=None)
68 parser.add_argument("--dim_hidden", type=int, default=None)
70 parser.add_argument("--nb_heads", type=int, default=None)
72 parser.add_argument("--nb_blocks", type=int, default=None)
74 parser.add_argument("--dropout", type=float, default=0.1)
76 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
78 parser.add_argument("--no_checkpoint", action="store_true", default=False)
80 parser.add_argument("--overwrite_results", action="store_true", default=False)
82 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
84 ##############################
87 parser.add_argument("--sandbox_level", type=int, default=0)
89 parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
91 parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
93 parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
95 ##############################
98 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
100 parser.add_argument("--picoclvr_height", type=int, default=12)
102 parser.add_argument("--picoclvr_width", type=int, default=16)
104 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
106 ##############################
109 parser.add_argument("--maze_height", type=int, default=23)
111 parser.add_argument("--maze_width", type=int, default=39)
113 parser.add_argument("--maze_nb_walls", type=int, default=45)
115 ##############################
118 parser.add_argument("--snake_height", type=int, default=6)
120 parser.add_argument("--snake_width", type=int, default=8)
122 parser.add_argument("--snake_nb_colors", type=int, default=5)
124 parser.add_argument("--snake_length", type=int, default=200)
126 ##############################
129 parser.add_argument("--stack_nb_steps", type=int, default=100)
131 parser.add_argument("--stack_nb_stacks", type=int, default=3)
133 parser.add_argument("--stack_nb_digits", type=int, default=3)
135 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
137 ##############################
140 parser.add_argument("--expr_nb_variables", type=int, default=5)
142 parser.add_argument("--expr_sequence_length", type=int, default=40)
144 parser.add_argument("--expr_operand_max", type=int, default=9)
146 parser.add_argument("--expr_result_max", type=int, default=99)
148 parser.add_argument("--expr_input_file", type=str, default=None)
150 ##############################
153 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
155 ######################################################################
157 args = parser.parse_args()
159 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
161 if args.result_dir is None:
162 args.result_dir = f"results_{args.task}"
164 ######################################################################
166 default_task_args = {
170 "nb_train_samples": 100000,
171 "nb_test_samples": 10000,
176 "nb_train_samples": 250000,
177 "nb_test_samples": 10000,
182 "nb_train_samples": 250000,
183 "nb_test_samples": 10000,
188 "nb_train_samples": 250000,
189 "nb_test_samples": 10000,
194 "nb_train_samples": 250000,
195 "nb_test_samples": 10000,
200 "nb_train_samples": 100000,
201 "nb_test_samples": 1000,
206 "nb_train_samples": 1000000,
207 "nb_test_samples": 10000,
212 "nb_train_samples": 25000,
213 "nb_test_samples": 1000,
217 if args.task in default_task_args:
218 for k, v in default_task_args[args.task].items():
219 if getattr(args, k) is None:
222 ######################################################################
224 default_model_args = {
255 if args.model in default_model_args:
256 for k, v in default_model_args[args.model].items():
257 if getattr(args, k) is None:
260 raise ValueError(f"Unknown model {args.model}")
262 ######################################################################
265 os.mkdir(args.result_dir)
266 except FileExistsError:
267 if not args.overwrite_results:
268 print(f"result directory {args.result_dir} already exists")
271 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
274 # torch.backends.cudnn.deterministic = True
275 # torch.backends.cudnn.benchmark = False
276 # torch.use_deterministic_algorithms(True)
277 torch.manual_seed(args.seed)
278 if torch.cuda.is_available():
279 torch.cuda.manual_seed_all(args.seed)
281 ######################################################################
285 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
287 if log_file is not None:
288 log_file.write(t + s + "\n")
296 log_string(f"args.{n} {getattr(args, n)}")
299 ######################################################################
302 def picoclvr_pruner_horizontal_green(p):
303 return not ("green" in p and ("left" in p or "right" in p))
306 picoclvr_pruner_train = (
307 picoclvr_pruner_horizontal_green
308 if args.picocvlr_prune_properties in {"train+eval"}
312 picoclvr_pruner_eval = (
313 (lambda p: not picoclvr_pruner_horizontal_green(p))
314 if args.picocvlr_prune_properties in {"train+eval", "eval"}
318 ######################################################################
320 if args.task == "sandbox":
321 if args.sandbox_level == 0:
322 problem = tasks.ProblemLevel0(
323 nb_sentences=args.sandbox_levels_nb_items,
324 len_prompt=args.sandbox_levels_len_source,
325 len_result=args.sandbox_levels_len_result,
327 elif args.sandbox_level == 1:
328 problem = tasks.ProblemLevel1(
329 nb_operators=args.sandbox_levels_nb_items,
330 len_source=args.sandbox_levels_len_source,
331 len_result=args.sandbox_levels_len_result,
333 elif args.sandbox_level == 2:
334 problem = tasks.ProblemLevel2(
335 len_source=args.sandbox_levels_len_source,
336 len_result=args.sandbox_levels_len_result,
339 raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
341 task = tasks.SandBox(
343 # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
344 nb_train_samples=args.nb_train_samples,
345 nb_test_samples=args.nb_test_samples,
346 batch_size=args.batch_size,
351 elif args.task == "picoclvr":
352 task = tasks.PicoCLVR(
353 nb_train_samples=args.nb_train_samples,
354 nb_test_samples=args.nb_test_samples,
355 batch_size=args.batch_size,
356 height=args.picoclvr_height,
357 width=args.picoclvr_width,
358 nb_colors=args.picoclvr_nb_colors,
361 pruner_train=picoclvr_pruner_train,
362 pruner_eval=picoclvr_pruner_eval,
365 elif args.task == "mnist":
367 nb_train_samples=args.nb_train_samples,
368 nb_test_samples=args.nb_test_samples,
369 batch_size=args.batch_size,
373 elif args.task == "maze":
375 nb_train_samples=args.nb_train_samples,
376 nb_test_samples=args.nb_test_samples,
377 batch_size=args.batch_size,
378 height=args.maze_height,
379 width=args.maze_width,
380 nb_walls=args.maze_nb_walls,
384 elif args.task == "snake":
386 nb_train_samples=args.nb_train_samples,
387 nb_test_samples=args.nb_test_samples,
388 batch_size=args.batch_size,
389 height=args.snake_height,
390 width=args.snake_width,
391 nb_colors=args.snake_nb_colors,
392 length=args.snake_length,
393 prompt_length=args.snake_length // 2,
397 elif args.task == "stack":
399 nb_train_samples=args.nb_train_samples,
400 nb_test_samples=args.nb_test_samples,
401 batch_size=args.batch_size,
403 nb_steps=args.stack_nb_steps,
404 nb_stacks=args.stack_nb_stacks,
405 nb_digits=args.stack_nb_digits,
406 fraction_values_for_train=args.stack_fraction_values_for_train,
410 elif args.task == "expr":
412 nb_train_samples=args.nb_train_samples,
413 nb_test_samples=args.nb_test_samples,
414 nb_variables=args.expr_nb_variables,
415 sequence_length=args.expr_sequence_length,
416 operand_max=args.expr_operand_max,
417 result_max=args.expr_result_max,
418 batch_size=args.batch_size,
422 elif args.task == "world":
424 nb_train_samples=args.nb_train_samples,
425 nb_test_samples=args.nb_test_samples,
426 batch_size=args.batch_size,
427 vqae_nb_epochs=args.world_vqae_nb_epochs,
433 raise ValueError(f"Unknown task {args.task}")
435 ######################################################################
437 log_string(f"device {device}")
439 vocabulary_size = task.vocabulary_size()
441 log_string(f"vocabulary_size {vocabulary_size}")
443 ##############################
446 vocabulary_size=vocabulary_size,
447 dim_model=args.dim_model,
448 dim_keys=args.dim_keys,
449 dim_hidden=args.dim_hidden,
450 nb_heads=args.nb_heads,
451 nb_blocks=args.nb_blocks,
453 dropout=args.dropout,
458 nb_parameters = sum(p.numel() for p in model.parameters())
459 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
461 ######################################################################
463 nb_epochs_finished = 0
465 if args.no_checkpoint:
466 log_string(f"not trying to load checkpoint.")
470 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
471 checkpoint = torch.load(checkpoint_name)
472 nb_epochs_finished = checkpoint["nb_epochs_finished"]
473 model.load_state_dict(checkpoint["model_state"])
474 torch.set_rng_state(checkpoint["rng_state"])
475 if torch.cuda.is_available():
476 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
478 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
480 except FileNotFoundError:
481 log_string("starting from scratch.")
484 log_string("error when loading the checkpoint.")
487 ######################################################################
489 if args.task == "expr" and args.expr_input_file is not None:
490 task.produce_results(
495 args.deterministic_synthesis,
496 args.expr_input_file,
501 ######################################################################
503 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
505 # Compute the entropy of the training tokens
508 for input in task.batches(split="train"):
509 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
510 token_probas = token_count / token_count.sum()
511 entropy = -torch.xlogy(token_probas, token_probas).sum()
512 train_set_perplexity = math.exp(entropy)
514 ##############################
516 # A bit of paranoia never hurts
521 for input in task.batches(split="train"):
522 assert input.dim() == 2 and input.dtype == torch.int64
524 train_examples[x.sum().item()] = x
526 nb_total, nb_collisions = 0, 0
527 for input in task.batches(split="test"):
528 assert input.dim() == 2 and input.dtype == torch.int64
531 y = train_examples.get(x.sum().item())
533 if x.size() == y.size() and (x - y).abs().sum() == 0:
539 f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
542 ##############################
544 if args.learning_rate_schedule == "cos":
545 learning_rate_schedule = {}
546 for n_epoch in range(args.nb_epochs):
547 u = n_epoch / args.nb_epochs * math.pi
548 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
553 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
557 learning_rate_schedule = {}
558 learning_rate = args.learning_rate
559 for n_epoch in range(args.nb_epochs):
561 learning_rate = u[n_epoch]
562 learning_rate_schedule[n_epoch] = learning_rate
564 log_string(f"learning_rate_schedule {learning_rate_schedule}")
566 ##############################
570 if nb_epochs_finished >= nb_epochs:
571 task.produce_results(
576 args.deterministic_synthesis,
579 for n_epoch in range(nb_epochs_finished, nb_epochs):
580 learning_rate = learning_rate_schedule[n_epoch]
582 log_string(f"learning_rate {learning_rate}")
584 if args.optim == "sgd":
585 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
586 elif args.optim == "adam":
587 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
588 elif args.optim == "adamw":
589 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
591 raise ValueError(f"Unknown optimizer {args.optim}.")
595 nb_train_samples, acc_train_loss = 0, 0.0
597 for input in task.batches(split="train"):
598 input = input.to(device)
599 output = model(mygpt.BracketedSequence(input)).x
600 loss = F.cross_entropy(output.transpose(1, 2), input)
601 acc_train_loss += loss.item() * input.size(0)
602 nb_train_samples += input.size(0)
603 nb_samples_seen += input.size(0)
605 optimizer.zero_grad()
609 with torch.autograd.no_grad():
612 nb_test_samples, acc_test_loss = 0, 0.0
614 for input in task.batches(split="test"):
615 input = input.to(device)
617 output = model(mygpt.BracketedSequence(input)).x
618 loss = F.cross_entropy(output.transpose(1, 2), input)
619 acc_test_loss += loss.item() * input.size(0)
620 nb_test_samples += input.size(0)
622 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
623 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
626 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
629 task.produce_results(
630 n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
634 "nb_epochs_finished": n_epoch + 1,
635 "model_state": model.state_dict(),
636 "rng_state": torch.get_rng_state(),
639 if torch.cuda.is_available():
640 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
642 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
643 torch.save(checkpoint, checkpoint_name)
644 log_string(f"saved checkpoint {checkpoint_name}")
646 ######################################################################