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("--rpl-nb_starting_values", type=int, default=5)
86 parser.add_argument("--rpl-max_input", type=int, default=9)
88 parser.add_argument("--rpl-prog_len", type=int, default=10)
90 parser.add_argument("--rpl-nb_runs", type=int, default=8)
92 ##############################
95 parser.add_argument("--sandbox_level", type=int, default=0)
97 parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
99 parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
101 parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
103 ##############################
106 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
108 parser.add_argument("--picoclvr_height", type=int, default=12)
110 parser.add_argument("--picoclvr_width", type=int, default=16)
112 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
114 ##############################
117 parser.add_argument("--maze_height", type=int, default=23)
119 parser.add_argument("--maze_width", type=int, default=39)
121 parser.add_argument("--maze_nb_walls", type=int, default=45)
123 ##############################
126 parser.add_argument("--snake_height", type=int, default=6)
128 parser.add_argument("--snake_width", type=int, default=8)
130 parser.add_argument("--snake_nb_colors", type=int, default=5)
132 parser.add_argument("--snake_length", type=int, default=200)
134 ##############################
137 parser.add_argument("--stack_nb_steps", type=int, default=100)
139 parser.add_argument("--stack_nb_stacks", type=int, default=3)
141 parser.add_argument("--stack_nb_digits", type=int, default=3)
143 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
145 ##############################
148 parser.add_argument("--expr_nb_variables", type=int, default=5)
150 parser.add_argument("--expr_sequence_length", type=int, default=40)
152 parser.add_argument("--expr_operand_max", type=int, default=9)
154 parser.add_argument("--expr_result_max", type=int, default=99)
156 parser.add_argument("--expr_input_file", type=str, default=None)
158 ##############################
161 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
163 ######################################################################
165 args = parser.parse_args()
167 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
169 if args.result_dir is None:
170 args.result_dir = f"results_{args.task}"
172 ######################################################################
174 default_task_args = {
178 "nb_train_samples": 100000,
179 "nb_test_samples": 10000,
184 "nb_train_samples": 250000,
185 "nb_test_samples": 10000,
190 "nb_train_samples": 250000,
191 "nb_test_samples": 10000,
196 "nb_train_samples": 250000,
197 "nb_test_samples": 10000,
202 "nb_train_samples": 250000,
203 "nb_test_samples": 10000,
208 "nb_train_samples": 100000,
209 "nb_test_samples": 1000,
214 "nb_train_samples": 1000000,
215 "nb_test_samples": 10000,
220 "nb_train_samples": 100000,
221 "nb_test_samples": 10000,
226 "nb_train_samples": 25000,
227 "nb_test_samples": 1000,
231 if args.task in default_task_args:
232 for k, v in default_task_args[args.task].items():
233 if getattr(args, k) is None:
236 ######################################################################
238 default_model_args = {
269 if args.model in default_model_args:
270 for k, v in default_model_args[args.model].items():
271 if getattr(args, k) is None:
274 raise ValueError(f"Unknown model {args.model}")
276 ######################################################################
279 os.mkdir(args.result_dir)
280 except FileExistsError:
281 if not args.overwrite_results:
282 print(f"result directory {args.result_dir} already exists")
285 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
288 # torch.backends.cudnn.deterministic = True
289 # torch.backends.cudnn.benchmark = False
290 # torch.use_deterministic_algorithms(True)
291 torch.manual_seed(args.seed)
292 if torch.cuda.is_available():
293 torch.cuda.manual_seed_all(args.seed)
295 ######################################################################
299 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
301 if log_file is not None:
302 log_file.write(t + s + "\n")
310 log_string(f"args.{n} {getattr(args, n)}")
313 ######################################################################
316 def picoclvr_pruner_horizontal_green(p):
317 return not ("green" in p and ("left" in p or "right" in p))
320 picoclvr_pruner_train = (
321 picoclvr_pruner_horizontal_green
322 if args.picocvlr_prune_properties in {"train+eval"}
326 picoclvr_pruner_eval = (
327 (lambda p: not picoclvr_pruner_horizontal_green(p))
328 if args.picocvlr_prune_properties in {"train+eval", "eval"}
332 ######################################################################
334 if args.task == "sandbox":
335 if args.sandbox_level == 0:
336 problem = tasks.ProblemLevel0(
337 nb_sentences=args.sandbox_levels_nb_items,
338 len_prompt=args.sandbox_levels_len_source,
339 len_result=args.sandbox_levels_len_result,
341 elif args.sandbox_level == 1:
342 problem = tasks.ProblemLevel1(
343 nb_operators=args.sandbox_levels_nb_items,
344 len_source=args.sandbox_levels_len_source,
345 len_result=args.sandbox_levels_len_result,
347 elif args.sandbox_level == 2:
348 problem = tasks.ProblemLevel2(
349 len_source=args.sandbox_levels_len_source,
350 len_result=args.sandbox_levels_len_result,
353 raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
355 task = tasks.SandBox(
357 # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
358 nb_train_samples=args.nb_train_samples,
359 nb_test_samples=args.nb_test_samples,
360 batch_size=args.batch_size,
365 elif args.task == "picoclvr":
366 task = tasks.PicoCLVR(
367 nb_train_samples=args.nb_train_samples,
368 nb_test_samples=args.nb_test_samples,
369 batch_size=args.batch_size,
370 height=args.picoclvr_height,
371 width=args.picoclvr_width,
372 nb_colors=args.picoclvr_nb_colors,
375 pruner_train=picoclvr_pruner_train,
376 pruner_eval=picoclvr_pruner_eval,
379 elif args.task == "mnist":
381 nb_train_samples=args.nb_train_samples,
382 nb_test_samples=args.nb_test_samples,
383 batch_size=args.batch_size,
387 elif args.task == "maze":
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.maze_height,
393 width=args.maze_width,
394 nb_walls=args.maze_nb_walls,
398 elif args.task == "snake":
400 nb_train_samples=args.nb_train_samples,
401 nb_test_samples=args.nb_test_samples,
402 batch_size=args.batch_size,
403 height=args.snake_height,
404 width=args.snake_width,
405 nb_colors=args.snake_nb_colors,
406 length=args.snake_length,
407 prompt_length=args.snake_length // 2,
411 elif args.task == "stack":
413 nb_train_samples=args.nb_train_samples,
414 nb_test_samples=args.nb_test_samples,
415 batch_size=args.batch_size,
417 nb_steps=args.stack_nb_steps,
418 nb_stacks=args.stack_nb_stacks,
419 nb_digits=args.stack_nb_digits,
420 fraction_values_for_train=args.stack_fraction_values_for_train,
424 elif args.task == "expr":
426 nb_train_samples=args.nb_train_samples,
427 nb_test_samples=args.nb_test_samples,
428 nb_variables=args.expr_nb_variables,
429 sequence_length=args.expr_sequence_length,
430 operand_max=args.expr_operand_max,
431 result_max=args.expr_result_max,
432 batch_size=args.batch_size,
436 elif args.task == "rpl":
438 nb_train_samples=args.nb_train_samples,
439 nb_test_samples=args.nb_test_samples,
440 batch_size=args.batch_size,
441 nb_starting_values=args.rpl_nb_starting_values,
442 max_input=args.rpl_max_input,
443 prog_len=args.rpl_prog_len,
444 nb_runs=args.rpl_nb_runs,
449 elif args.task == "world":
451 nb_train_samples=args.nb_train_samples,
452 nb_test_samples=args.nb_test_samples,
453 batch_size=args.batch_size,
454 vqae_nb_epochs=args.world_vqae_nb_epochs,
460 raise ValueError(f"Unknown task {args.task}")
462 ######################################################################
464 log_string(f"device {device}")
466 vocabulary_size = task.vocabulary_size()
468 log_string(f"vocabulary_size {vocabulary_size}")
470 ##############################
473 vocabulary_size=vocabulary_size,
474 dim_model=args.dim_model,
475 dim_keys=args.dim_keys,
476 dim_hidden=args.dim_hidden,
477 nb_heads=args.nb_heads,
478 nb_blocks=args.nb_blocks,
480 dropout=args.dropout,
485 nb_parameters = sum(p.numel() for p in model.parameters())
486 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
488 ######################################################################
490 nb_epochs_finished = 0
492 if args.no_checkpoint:
493 log_string(f"not trying to load checkpoint.")
497 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
498 checkpoint = torch.load(checkpoint_name)
499 nb_epochs_finished = checkpoint["nb_epochs_finished"]
500 model.load_state_dict(checkpoint["model_state"])
501 torch.set_rng_state(checkpoint["rng_state"])
502 if torch.cuda.is_available():
503 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
505 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
507 except FileNotFoundError:
508 log_string("starting from scratch.")
511 log_string("error when loading the checkpoint.")
514 ######################################################################
516 if args.task == "expr" and args.expr_input_file is not None:
517 task.produce_results(
522 args.deterministic_synthesis,
523 args.expr_input_file,
528 ######################################################################
530 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
532 # Compute the entropy of the training tokens
535 for input in task.batches(split="train"):
536 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
537 token_probas = token_count / token_count.sum()
538 entropy = -torch.xlogy(token_probas, token_probas).sum()
539 train_set_perplexity = math.exp(entropy)
541 ##############################
543 # A bit of paranoia never hurts
548 for input in task.batches(split="train"):
549 assert input.dim() == 2 and input.dtype == torch.int64
551 train_examples[x.sum().item()] = x
553 nb_total, nb_collisions = 0, 0
554 for input in task.batches(split="test"):
555 assert input.dim() == 2 and input.dtype == torch.int64
558 y = train_examples.get(x.sum().item())
560 if x.size() == y.size() and (x - y).abs().sum() == 0:
566 f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
569 ##############################
571 if args.learning_rate_schedule == "cos":
572 learning_rate_schedule = {}
573 for n_epoch in range(args.nb_epochs):
574 u = n_epoch / args.nb_epochs * math.pi
575 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
580 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
584 learning_rate_schedule = {}
585 learning_rate = args.learning_rate
586 for n_epoch in range(args.nb_epochs):
588 learning_rate = u[n_epoch]
589 learning_rate_schedule[n_epoch] = learning_rate
591 log_string(f"learning_rate_schedule {learning_rate_schedule}")
593 ##############################
597 if nb_epochs_finished >= nb_epochs:
598 task.produce_results(
603 args.deterministic_synthesis,
606 for n_epoch in range(nb_epochs_finished, nb_epochs):
607 learning_rate = learning_rate_schedule[n_epoch]
609 log_string(f"learning_rate {learning_rate}")
611 if args.optim == "sgd":
612 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
613 elif args.optim == "adam":
614 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
615 elif args.optim == "adamw":
616 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
618 raise ValueError(f"Unknown optimizer {args.optim}.")
622 nb_train_samples, acc_train_loss = 0, 0.0
624 for input in task.batches(split="train"):
625 input = input.to(device)
626 output = model(mygpt.BracketedSequence(input)).x
627 loss = F.cross_entropy(output.transpose(1, 2), input)
628 acc_train_loss += loss.item() * input.size(0)
629 nb_train_samples += input.size(0)
630 nb_samples_seen += input.size(0)
632 optimizer.zero_grad()
636 with torch.autograd.no_grad():
639 nb_test_samples, acc_test_loss = 0, 0.0
641 for input in task.batches(split="test"):
642 input = input.to(device)
644 output = model(mygpt.BracketedSequence(input)).x
645 loss = F.cross_entropy(output.transpose(1, 2), input)
646 acc_test_loss += loss.item() * input.size(0)
647 nb_test_samples += input.size(0)
649 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
650 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
653 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
656 task.produce_results(
657 n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
661 "nb_epochs_finished": n_epoch + 1,
662 "model_state": model.state_dict(),
663 "rng_state": torch.get_rng_state(),
666 if torch.cuda.is_available():
667 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
669 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
670 torch.save(checkpoint, checkpoint_name)
671 log_string(f"saved checkpoint {checkpoint_name}")
673 ######################################################################