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 import math, sys, argparse, time, tqdm, os
10 import torch, torchvision
12 from torch.nn import functional as F
17 ######################################################################
19 if torch.cuda.is_available():
20 device = torch.device("cuda")
21 torch.backends.cuda.matmul.allow_tf32 = True
23 device = torch.device("cpu")
25 ######################################################################
27 parser = argparse.ArgumentParser(
28 description="An implementation of GPT with cache.",
29 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
36 help="sandbox, picoclvr, mnist, maze, snake, stack, expr, rpl, world",
39 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
41 parser.add_argument("--result_dir", type=str, default=None)
43 parser.add_argument("--seed", type=int, default=0)
45 parser.add_argument("--nb_epochs", type=int, default=None)
47 parser.add_argument("--batch_size", type=int, default=None)
49 parser.add_argument("--nb_train_samples", type=int, default=None)
51 parser.add_argument("--nb_test_samples", type=int, default=None)
53 parser.add_argument("--optim", type=str, default="adam")
55 parser.add_argument("--learning_rate", type=float, default=1e-4)
57 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
59 parser.add_argument("--model", type=str, default="37M")
61 parser.add_argument("--dim_model", type=int, default=None)
63 parser.add_argument("--dim_keys", type=int, default=None)
65 parser.add_argument("--dim_hidden", type=int, default=None)
67 parser.add_argument("--nb_heads", type=int, default=None)
69 parser.add_argument("--nb_blocks", type=int, default=None)
71 parser.add_argument("--dropout", type=float, default=0.1)
73 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
75 parser.add_argument("--no_checkpoint", action="store_true", default=False)
77 parser.add_argument("--overwrite_results", action="store_true", default=False)
79 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
81 ##############################
84 parser.add_argument("--sandbox_level", type=int, default=0)
86 parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
88 parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
90 parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
92 ##############################
95 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
97 parser.add_argument("--picoclvr_height", type=int, default=12)
99 parser.add_argument("--picoclvr_width", type=int, default=16)
101 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
103 ##############################
106 parser.add_argument("--maze_height", type=int, default=23)
108 parser.add_argument("--maze_width", type=int, default=39)
110 parser.add_argument("--maze_nb_walls", type=int, default=45)
112 ##############################
115 parser.add_argument("--snake_height", type=int, default=6)
117 parser.add_argument("--snake_width", type=int, default=8)
119 parser.add_argument("--snake_nb_colors", type=int, default=5)
121 parser.add_argument("--snake_length", type=int, default=200)
123 ##############################
126 parser.add_argument("--stack_nb_steps", type=int, default=100)
128 parser.add_argument("--stack_nb_stacks", type=int, default=3)
130 parser.add_argument("--stack_nb_digits", type=int, default=3)
132 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
134 ##############################
137 parser.add_argument("--expr_nb_variables", type=int, default=5)
139 parser.add_argument("--expr_sequence_length", type=int, default=40)
141 parser.add_argument("--expr_operand_max", type=int, default=9)
143 parser.add_argument("--expr_result_max", type=int, default=99)
145 parser.add_argument("--expr_input_file", type=str, default=None)
147 ##############################
150 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
152 ######################################################################
154 args = parser.parse_args()
156 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
158 if args.result_dir is None:
159 args.result_dir = f"results_{args.task}"
161 ######################################################################
163 default_task_args = {
167 "nb_train_samples": 100000,
168 "nb_test_samples": 10000,
173 "nb_train_samples": 250000,
174 "nb_test_samples": 10000,
179 "nb_train_samples": 250000,
180 "nb_test_samples": 10000,
185 "nb_train_samples": 250000,
186 "nb_test_samples": 10000,
191 "nb_train_samples": 250000,
192 "nb_test_samples": 10000,
197 "nb_train_samples": 100000,
198 "nb_test_samples": 1000,
203 "nb_train_samples": 1000000,
204 "nb_test_samples": 10000,
209 "nb_train_samples": 100000,
210 "nb_test_samples": 10000,
215 "nb_train_samples": 25000,
216 "nb_test_samples": 1000,
220 if args.task in default_task_args:
221 for k, v in default_task_args[args.task].items():
222 if getattr(args, k) is None:
225 ######################################################################
227 default_model_args = {
258 if args.model in default_model_args:
259 for k, v in default_model_args[args.model].items():
260 if getattr(args, k) is None:
263 raise ValueError(f"Unknown model {args.model}")
265 ######################################################################
268 os.mkdir(args.result_dir)
269 except FileExistsError:
270 if not args.overwrite_results:
271 print(f"result directory {args.result_dir} already exists")
274 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
277 # torch.backends.cudnn.deterministic = True
278 # torch.backends.cudnn.benchmark = False
279 # torch.use_deterministic_algorithms(True)
280 torch.manual_seed(args.seed)
281 if torch.cuda.is_available():
282 torch.cuda.manual_seed_all(args.seed)
284 ######################################################################
288 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
290 if log_file is not None:
291 log_file.write(t + s + "\n")
299 log_string(f"args.{n} {getattr(args, n)}")
302 ######################################################################
305 def picoclvr_pruner_horizontal_green(p):
306 return not ("green" in p and ("left" in p or "right" in p))
309 picoclvr_pruner_train = (
310 picoclvr_pruner_horizontal_green
311 if args.picocvlr_prune_properties in {"train+eval"}
315 picoclvr_pruner_eval = (
316 (lambda p: not picoclvr_pruner_horizontal_green(p))
317 if args.picocvlr_prune_properties in {"train+eval", "eval"}
321 ######################################################################
323 if args.task == "sandbox":
324 if args.sandbox_level == 0:
325 problem = tasks.ProblemLevel0(
326 nb_sentences=args.sandbox_levels_nb_items,
327 len_prompt=args.sandbox_levels_len_source,
328 len_result=args.sandbox_levels_len_result,
330 elif args.sandbox_level == 1:
331 problem = tasks.ProblemLevel1(
332 nb_operators=args.sandbox_levels_nb_items,
333 len_source=args.sandbox_levels_len_source,
334 len_result=args.sandbox_levels_len_result,
336 elif args.sandbox_level == 2:
337 problem = tasks.ProblemLevel2(
338 len_source=args.sandbox_levels_len_source,
339 len_result=args.sandbox_levels_len_result,
342 raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
344 task = tasks.SandBox(
346 # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
347 nb_train_samples=args.nb_train_samples,
348 nb_test_samples=args.nb_test_samples,
349 batch_size=args.batch_size,
354 elif args.task == "picoclvr":
355 task = tasks.PicoCLVR(
356 nb_train_samples=args.nb_train_samples,
357 nb_test_samples=args.nb_test_samples,
358 batch_size=args.batch_size,
359 height=args.picoclvr_height,
360 width=args.picoclvr_width,
361 nb_colors=args.picoclvr_nb_colors,
364 pruner_train=picoclvr_pruner_train,
365 pruner_eval=picoclvr_pruner_eval,
368 elif args.task == "mnist":
370 nb_train_samples=args.nb_train_samples,
371 nb_test_samples=args.nb_test_samples,
372 batch_size=args.batch_size,
376 elif args.task == "maze":
378 nb_train_samples=args.nb_train_samples,
379 nb_test_samples=args.nb_test_samples,
380 batch_size=args.batch_size,
381 height=args.maze_height,
382 width=args.maze_width,
383 nb_walls=args.maze_nb_walls,
387 elif args.task == "snake":
389 nb_train_samples=args.nb_train_samples,
390 nb_test_samples=args.nb_test_samples,
391 batch_size=args.batch_size,
392 height=args.snake_height,
393 width=args.snake_width,
394 nb_colors=args.snake_nb_colors,
395 length=args.snake_length,
396 prompt_length=args.snake_length // 2,
400 elif args.task == "stack":
402 nb_train_samples=args.nb_train_samples,
403 nb_test_samples=args.nb_test_samples,
404 batch_size=args.batch_size,
406 nb_steps=args.stack_nb_steps,
407 nb_stacks=args.stack_nb_stacks,
408 nb_digits=args.stack_nb_digits,
409 fraction_values_for_train=args.stack_fraction_values_for_train,
413 elif args.task == "expr":
415 nb_train_samples=args.nb_train_samples,
416 nb_test_samples=args.nb_test_samples,
417 nb_variables=args.expr_nb_variables,
418 sequence_length=args.expr_sequence_length,
419 operand_max=args.expr_operand_max,
420 result_max=args.expr_result_max,
421 batch_size=args.batch_size,
425 elif args.task == "rpl":
427 nb_train_samples=args.nb_train_samples,
428 nb_test_samples=args.nb_test_samples,
429 batch_size=args.batch_size,
434 elif args.task == "world":
436 nb_train_samples=args.nb_train_samples,
437 nb_test_samples=args.nb_test_samples,
438 batch_size=args.batch_size,
439 vqae_nb_epochs=args.world_vqae_nb_epochs,
445 raise ValueError(f"Unknown task {args.task}")
447 ######################################################################
449 log_string(f"device {device}")
451 vocabulary_size = task.vocabulary_size()
453 log_string(f"vocabulary_size {vocabulary_size}")
455 ##############################
458 vocabulary_size=vocabulary_size,
459 dim_model=args.dim_model,
460 dim_keys=args.dim_keys,
461 dim_hidden=args.dim_hidden,
462 nb_heads=args.nb_heads,
463 nb_blocks=args.nb_blocks,
465 dropout=args.dropout,
470 nb_parameters = sum(p.numel() for p in model.parameters())
471 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
473 ######################################################################
475 nb_epochs_finished = 0
477 if args.no_checkpoint:
478 log_string(f"not trying to load checkpoint.")
482 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
483 checkpoint = torch.load(checkpoint_name)
484 nb_epochs_finished = checkpoint["nb_epochs_finished"]
485 model.load_state_dict(checkpoint["model_state"])
486 torch.set_rng_state(checkpoint["rng_state"])
487 if torch.cuda.is_available():
488 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
490 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
492 except FileNotFoundError:
493 log_string("starting from scratch.")
496 log_string("error when loading the checkpoint.")
499 ######################################################################
501 if args.task == "expr" and args.expr_input_file is not None:
502 task.produce_results(
507 args.deterministic_synthesis,
508 args.expr_input_file,
513 ######################################################################
515 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
517 # Compute the entropy of the training tokens
520 for input in task.batches(split="train"):
521 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
522 token_probas = token_count / token_count.sum()
523 entropy = -torch.xlogy(token_probas, token_probas).sum()
524 train_set_perplexity = math.exp(entropy)
526 ##############################
528 # A bit of paranoia never hurts
533 for input in task.batches(split="train"):
534 assert input.dim() == 2 and input.dtype == torch.int64
536 train_examples[x.sum().item()] = x
538 nb_total, nb_collisions = 0, 0
539 for input in task.batches(split="test"):
540 assert input.dim() == 2 and input.dtype == torch.int64
543 y = train_examples.get(x.sum().item())
545 if x.size() == y.size() and (x - y).abs().sum() == 0:
551 f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
554 ##############################
556 if args.learning_rate_schedule == "cos":
557 learning_rate_schedule = {}
558 for n_epoch in range(args.nb_epochs):
559 u = n_epoch / args.nb_epochs * math.pi
560 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
565 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
569 learning_rate_schedule = {}
570 learning_rate = args.learning_rate
571 for n_epoch in range(args.nb_epochs):
573 learning_rate = u[n_epoch]
574 learning_rate_schedule[n_epoch] = learning_rate
576 log_string(f"learning_rate_schedule {learning_rate_schedule}")
578 ##############################
582 if nb_epochs_finished >= nb_epochs:
583 task.produce_results(
588 args.deterministic_synthesis,
591 for n_epoch in range(nb_epochs_finished, nb_epochs):
592 learning_rate = learning_rate_schedule[n_epoch]
594 log_string(f"learning_rate {learning_rate}")
596 if args.optim == "sgd":
597 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
598 elif args.optim == "adam":
599 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
600 elif args.optim == "adamw":
601 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
603 raise ValueError(f"Unknown optimizer {args.optim}.")
607 nb_train_samples, acc_train_loss = 0, 0.0
609 for input in task.batches(split="train"):
610 input = input.to(device)
611 output = model(mygpt.BracketedSequence(input)).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)
615 nb_samples_seen += input.size(0)
617 optimizer.zero_grad()
621 with torch.autograd.no_grad():
624 nb_test_samples, acc_test_loss = 0, 0.0
626 for input in task.batches(split="test"):
627 input = input.to(device)
629 output = model(mygpt.BracketedSequence(input)).x
630 loss = F.cross_entropy(output.transpose(1, 2), input)
631 acc_test_loss += loss.item() * input.size(0)
632 nb_test_samples += input.size(0)
634 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
635 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
638 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
641 task.produce_results(
642 n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
646 "nb_epochs_finished": n_epoch + 1,
647 "model_state": model.state_dict(),
648 "rng_state": torch.get_rng_state(),
651 if torch.cuda.is_available():
652 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
654 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
655 torch.save(checkpoint, checkpoint_name)
656 log_string(f"saved checkpoint {checkpoint_name}")
658 ######################################################################