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
15 import mygpt, tasks, problems
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 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
94 ##############################
97 parser.add_argument("--sandbox_level", type=int, default=0)
99 parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
101 parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
103 parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
105 ##############################
108 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
110 parser.add_argument("--picoclvr_height", type=int, default=12)
112 parser.add_argument("--picoclvr_width", type=int, default=16)
114 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
116 ##############################
119 parser.add_argument("--maze_height", type=int, default=23)
121 parser.add_argument("--maze_width", type=int, default=39)
123 parser.add_argument("--maze_nb_walls", type=int, default=45)
125 ##############################
128 parser.add_argument("--snake_height", type=int, default=6)
130 parser.add_argument("--snake_width", type=int, default=8)
132 parser.add_argument("--snake_nb_colors", type=int, default=5)
134 parser.add_argument("--snake_length", type=int, default=200)
136 ##############################
139 parser.add_argument("--stack_nb_steps", type=int, default=100)
141 parser.add_argument("--stack_nb_stacks", type=int, default=3)
143 parser.add_argument("--stack_nb_digits", type=int, default=3)
145 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
147 ##############################
150 parser.add_argument("--expr_nb_variables", type=int, default=5)
152 parser.add_argument("--expr_sequence_length", type=int, default=40)
154 parser.add_argument("--expr_operand_max", type=int, default=9)
156 parser.add_argument("--expr_result_max", type=int, default=99)
158 parser.add_argument("--expr_input_file", type=str, default=None)
160 ##############################
163 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
165 ######################################################################
167 args = parser.parse_args()
169 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
171 if args.result_dir is None:
172 args.result_dir = f"results_{args.task}"
174 ######################################################################
176 default_task_args = {
180 "nb_train_samples": 100000,
181 "nb_test_samples": 10000,
186 "nb_train_samples": 250000,
187 "nb_test_samples": 10000,
192 "nb_train_samples": 250000,
193 "nb_test_samples": 10000,
198 "nb_train_samples": 250000,
199 "nb_test_samples": 10000,
204 "nb_train_samples": 250000,
205 "nb_test_samples": 10000,
210 "nb_train_samples": 100000,
211 "nb_test_samples": 1000,
216 "nb_train_samples": 1000000,
217 "nb_test_samples": 10000,
222 "nb_train_samples": 100000,
223 "nb_test_samples": 10000,
228 "nb_train_samples": 25000,
229 "nb_test_samples": 1000,
233 if args.task in default_task_args:
234 for k, v in default_task_args[args.task].items():
235 if getattr(args, k) is None:
238 ######################################################################
240 default_model_args = {
271 if args.model in default_model_args:
272 for k, v in default_model_args[args.model].items():
273 if getattr(args, k) is None:
276 raise ValueError(f"Unknown model {args.model}")
278 ######################################################################
281 os.mkdir(args.result_dir)
282 except FileExistsError:
283 if not args.overwrite_results:
284 print(f"result directory {args.result_dir} already exists")
287 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
290 # torch.backends.cudnn.deterministic = True
291 # torch.backends.cudnn.benchmark = False
292 # torch.use_deterministic_algorithms(True)
293 torch.manual_seed(args.seed)
294 if torch.cuda.is_available():
295 torch.cuda.manual_seed_all(args.seed)
297 ######################################################################
301 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
303 if log_file is not None:
304 log_file.write(t + s + "\n")
312 log_string(f"args.{n} {getattr(args, n)}")
315 ######################################################################
318 def picoclvr_pruner_horizontal_green(p):
319 return not ("green" in p and ("left" in p or "right" in p))
322 picoclvr_pruner_train = (
323 picoclvr_pruner_horizontal_green
324 if args.picocvlr_prune_properties in {"train+eval"}
328 picoclvr_pruner_eval = (
329 (lambda p: not picoclvr_pruner_horizontal_green(p))
330 if args.picocvlr_prune_properties in {"train+eval", "eval"}
334 ######################################################################
336 if args.task == "sandbox":
337 if args.sandbox_level == 0:
338 problem = problems.ProblemLevel0(
339 nb_sentences=args.sandbox_levels_nb_items,
340 len_prompt=args.sandbox_levels_len_source,
341 len_result=args.sandbox_levels_len_result,
343 elif args.sandbox_level == 1:
344 problem = problems.ProblemLevel1(
345 nb_operators=args.sandbox_levels_nb_items,
346 len_source=args.sandbox_levels_len_source,
347 len_result=args.sandbox_levels_len_result,
349 elif args.sandbox_level == 2:
350 problem = problems.ProblemLevel2(
351 len_source=args.sandbox_levels_len_source,
352 len_result=args.sandbox_levels_len_result,
355 raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
357 task = tasks.SandBox(
359 # problems.ProblemAddition(zero_padded=False, inverted_result=False),
360 nb_train_samples=args.nb_train_samples,
361 nb_test_samples=args.nb_test_samples,
362 batch_size=args.batch_size,
367 elif args.task == "picoclvr":
368 task = tasks.PicoCLVR(
369 nb_train_samples=args.nb_train_samples,
370 nb_test_samples=args.nb_test_samples,
371 batch_size=args.batch_size,
372 height=args.picoclvr_height,
373 width=args.picoclvr_width,
374 nb_colors=args.picoclvr_nb_colors,
377 pruner_train=picoclvr_pruner_train,
378 pruner_eval=picoclvr_pruner_eval,
381 elif args.task == "mnist":
383 nb_train_samples=args.nb_train_samples,
384 nb_test_samples=args.nb_test_samples,
385 batch_size=args.batch_size,
389 elif args.task == "maze":
391 nb_train_samples=args.nb_train_samples,
392 nb_test_samples=args.nb_test_samples,
393 batch_size=args.batch_size,
394 height=args.maze_height,
395 width=args.maze_width,
396 nb_walls=args.maze_nb_walls,
400 elif args.task == "snake":
402 nb_train_samples=args.nb_train_samples,
403 nb_test_samples=args.nb_test_samples,
404 batch_size=args.batch_size,
405 height=args.snake_height,
406 width=args.snake_width,
407 nb_colors=args.snake_nb_colors,
408 length=args.snake_length,
409 prompt_length=args.snake_length // 2,
413 elif args.task == "stack":
415 nb_train_samples=args.nb_train_samples,
416 nb_test_samples=args.nb_test_samples,
417 batch_size=args.batch_size,
419 nb_steps=args.stack_nb_steps,
420 nb_stacks=args.stack_nb_stacks,
421 nb_digits=args.stack_nb_digits,
422 fraction_values_for_train=args.stack_fraction_values_for_train,
426 elif args.task == "expr":
428 nb_train_samples=args.nb_train_samples,
429 nb_test_samples=args.nb_test_samples,
430 nb_variables=args.expr_nb_variables,
431 sequence_length=args.expr_sequence_length,
432 operand_max=args.expr_operand_max,
433 result_max=args.expr_result_max,
434 batch_size=args.batch_size,
438 elif args.task == "rpl":
440 nb_train_samples=args.nb_train_samples,
441 nb_test_samples=args.nb_test_samples,
442 batch_size=args.batch_size,
443 nb_starting_values=args.rpl_nb_starting_values,
444 max_input=args.rpl_max_input,
445 prog_len=args.rpl_prog_len,
446 nb_runs=args.rpl_nb_runs,
447 no_prog=args.rpl_no_prog,
452 elif args.task == "world":
454 nb_train_samples=args.nb_train_samples,
455 nb_test_samples=args.nb_test_samples,
456 batch_size=args.batch_size,
457 vqae_nb_epochs=args.world_vqae_nb_epochs,
463 raise ValueError(f"Unknown task {args.task}")
465 ######################################################################
467 log_string(f"device {device}")
469 vocabulary_size = task.vocabulary_size()
471 log_string(f"vocabulary_size {vocabulary_size}")
473 ##############################
476 vocabulary_size=vocabulary_size,
477 dim_model=args.dim_model,
478 dim_keys=args.dim_keys,
479 dim_hidden=args.dim_hidden,
480 nb_heads=args.nb_heads,
481 nb_blocks=args.nb_blocks,
483 dropout=args.dropout,
488 nb_parameters = sum(p.numel() for p in model.parameters())
489 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
491 ######################################################################
493 nb_epochs_finished = 0
495 if args.no_checkpoint:
496 log_string(f"not trying to load checkpoint.")
500 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
501 checkpoint = torch.load(checkpoint_name)
502 nb_epochs_finished = checkpoint["nb_epochs_finished"]
503 model.load_state_dict(checkpoint["model_state"])
504 torch.set_rng_state(checkpoint["rng_state"])
505 if torch.cuda.is_available():
506 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
508 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
510 except FileNotFoundError:
511 log_string("starting from scratch.")
514 log_string("error when loading the checkpoint.")
517 ######################################################################
519 if args.task == "expr" and args.expr_input_file is not None:
520 task.produce_results(
521 n_epoch=nb_epochs_finished,
523 result_dir=args.result_dir,
525 deterministic_synthesis=args.deterministic_synthesis,
526 input_file=args.expr_input_file,
531 ######################################################################
533 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
535 # Compute the entropy of the training tokens
538 for input in task.batches(split="train"):
539 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
540 token_probas = token_count / token_count.sum()
541 entropy = -torch.xlogy(token_probas, token_probas).sum()
542 train_set_perplexity = math.exp(entropy)
544 ##############################
546 # A bit of paranoia never hurts
551 for input in task.batches(split="train"):
552 assert input.dim() == 2 and input.dtype == torch.int64
554 train_examples[x.sum().item()] = x
556 nb_total, nb_collisions = 0, 0
557 for input in task.batches(split="test"):
558 assert input.dim() == 2 and input.dtype == torch.int64
561 y = train_examples.get(x.sum().item())
563 if x.size() == y.size() and (x - y).abs().sum() == 0:
569 f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
572 ##############################
574 if args.learning_rate_schedule == "cos":
575 learning_rate_schedule = {}
576 for n_epoch in range(args.nb_epochs):
577 u = n_epoch / args.nb_epochs * math.pi
578 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
583 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
587 learning_rate_schedule = {}
588 learning_rate = args.learning_rate
589 for n_epoch in range(args.nb_epochs):
591 learning_rate = u[n_epoch]
592 learning_rate_schedule[n_epoch] = learning_rate
594 log_string(f"learning_rate_schedule {learning_rate_schedule}")
596 ##############################
600 if nb_epochs_finished >= nb_epochs:
601 task.produce_results(
602 n_epoch=nb_epochs_finished,
604 result_dir=args.result_dir,
606 deterministic_synthesis=args.deterministic_synthesis,
609 for n_epoch in range(nb_epochs_finished, nb_epochs):
610 learning_rate = learning_rate_schedule[n_epoch]
612 log_string(f"learning_rate {learning_rate}")
614 if args.optim == "sgd":
615 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
616 elif args.optim == "adam":
617 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
618 elif args.optim == "adamw":
619 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
621 raise ValueError(f"Unknown optimizer {args.optim}.")
625 nb_train_samples, acc_train_loss = 0, 0.0
627 for input in task.batches(split="train"):
628 input = input.to(device)
629 output = model(mygpt.BracketedSequence(input)).x
630 loss = F.cross_entropy(output.transpose(1, 2), input)
631 acc_train_loss += loss.item() * input.size(0)
632 nb_train_samples += input.size(0)
633 nb_samples_seen += input.size(0)
635 optimizer.zero_grad()
639 with torch.autograd.no_grad():
642 nb_test_samples, acc_test_loss = 0, 0.0
644 for input in task.batches(split="test"):
645 input = input.to(device)
647 output = model(mygpt.BracketedSequence(input)).x
648 loss = F.cross_entropy(output.transpose(1, 2), input)
649 acc_test_loss += loss.item() * input.size(0)
650 nb_test_samples += input.size(0)
652 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
653 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
656 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
659 task.produce_results(
662 result_dir=args.result_dir,
664 deterministic_synthesis=args.deterministic_synthesis,
668 "nb_epochs_finished": n_epoch + 1,
669 "model_state": model.state_dict(),
670 "rng_state": torch.get_rng_state(),
673 if torch.cuda.is_available():
674 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
676 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
677 torch.save(checkpoint, checkpoint_name)
678 log_string(f"saved checkpoint {checkpoint_name}")
680 ######################################################################