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, rpl, 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": 100000,
213 "nb_test_samples": 10000,
218 "nb_train_samples": 25000,
219 "nb_test_samples": 1000,
223 if args.task in default_task_args:
224 for k, v in default_task_args[args.task].items():
225 if getattr(args, k) is None:
228 ######################################################################
230 default_model_args = {
261 if args.model in default_model_args:
262 for k, v in default_model_args[args.model].items():
263 if getattr(args, k) is None:
266 raise ValueError(f"Unknown model {args.model}")
268 ######################################################################
271 os.mkdir(args.result_dir)
272 except FileExistsError:
273 if not args.overwrite_results:
274 print(f"result directory {args.result_dir} already exists")
277 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
280 # torch.backends.cudnn.deterministic = True
281 # torch.backends.cudnn.benchmark = False
282 # torch.use_deterministic_algorithms(True)
283 torch.manual_seed(args.seed)
284 if torch.cuda.is_available():
285 torch.cuda.manual_seed_all(args.seed)
287 ######################################################################
291 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
293 if log_file is not None:
294 log_file.write(t + s + "\n")
302 log_string(f"args.{n} {getattr(args, n)}")
305 ######################################################################
308 def picoclvr_pruner_horizontal_green(p):
309 return not ("green" in p and ("left" in p or "right" in p))
312 picoclvr_pruner_train = (
313 picoclvr_pruner_horizontal_green
314 if args.picocvlr_prune_properties in {"train+eval"}
318 picoclvr_pruner_eval = (
319 (lambda p: not picoclvr_pruner_horizontal_green(p))
320 if args.picocvlr_prune_properties in {"train+eval", "eval"}
324 ######################################################################
326 if args.task == "sandbox":
327 if args.sandbox_level == 0:
328 problem = tasks.ProblemLevel0(
329 nb_sentences=args.sandbox_levels_nb_items,
330 len_prompt=args.sandbox_levels_len_source,
331 len_result=args.sandbox_levels_len_result,
333 elif args.sandbox_level == 1:
334 problem = tasks.ProblemLevel1(
335 nb_operators=args.sandbox_levels_nb_items,
336 len_source=args.sandbox_levels_len_source,
337 len_result=args.sandbox_levels_len_result,
339 elif args.sandbox_level == 2:
340 problem = tasks.ProblemLevel2(
341 len_source=args.sandbox_levels_len_source,
342 len_result=args.sandbox_levels_len_result,
345 raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
347 task = tasks.SandBox(
349 # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
350 nb_train_samples=args.nb_train_samples,
351 nb_test_samples=args.nb_test_samples,
352 batch_size=args.batch_size,
357 elif args.task == "picoclvr":
358 task = tasks.PicoCLVR(
359 nb_train_samples=args.nb_train_samples,
360 nb_test_samples=args.nb_test_samples,
361 batch_size=args.batch_size,
362 height=args.picoclvr_height,
363 width=args.picoclvr_width,
364 nb_colors=args.picoclvr_nb_colors,
367 pruner_train=picoclvr_pruner_train,
368 pruner_eval=picoclvr_pruner_eval,
371 elif args.task == "mnist":
373 nb_train_samples=args.nb_train_samples,
374 nb_test_samples=args.nb_test_samples,
375 batch_size=args.batch_size,
379 elif args.task == "maze":
381 nb_train_samples=args.nb_train_samples,
382 nb_test_samples=args.nb_test_samples,
383 batch_size=args.batch_size,
384 height=args.maze_height,
385 width=args.maze_width,
386 nb_walls=args.maze_nb_walls,
390 elif args.task == "snake":
392 nb_train_samples=args.nb_train_samples,
393 nb_test_samples=args.nb_test_samples,
394 batch_size=args.batch_size,
395 height=args.snake_height,
396 width=args.snake_width,
397 nb_colors=args.snake_nb_colors,
398 length=args.snake_length,
399 prompt_length=args.snake_length // 2,
403 elif args.task == "stack":
405 nb_train_samples=args.nb_train_samples,
406 nb_test_samples=args.nb_test_samples,
407 batch_size=args.batch_size,
409 nb_steps=args.stack_nb_steps,
410 nb_stacks=args.stack_nb_stacks,
411 nb_digits=args.stack_nb_digits,
412 fraction_values_for_train=args.stack_fraction_values_for_train,
416 elif args.task == "expr":
418 nb_train_samples=args.nb_train_samples,
419 nb_test_samples=args.nb_test_samples,
420 nb_variables=args.expr_nb_variables,
421 sequence_length=args.expr_sequence_length,
422 operand_max=args.expr_operand_max,
423 result_max=args.expr_result_max,
424 batch_size=args.batch_size,
428 elif args.task == "rpl":
430 nb_train_samples=args.nb_train_samples,
431 nb_test_samples=args.nb_test_samples,
432 batch_size=args.batch_size,
437 elif args.task == "world":
439 nb_train_samples=args.nb_train_samples,
440 nb_test_samples=args.nb_test_samples,
441 batch_size=args.batch_size,
442 vqae_nb_epochs=args.world_vqae_nb_epochs,
448 raise ValueError(f"Unknown task {args.task}")
450 ######################################################################
452 log_string(f"device {device}")
454 vocabulary_size = task.vocabulary_size()
456 log_string(f"vocabulary_size {vocabulary_size}")
458 ##############################
461 vocabulary_size=vocabulary_size,
462 dim_model=args.dim_model,
463 dim_keys=args.dim_keys,
464 dim_hidden=args.dim_hidden,
465 nb_heads=args.nb_heads,
466 nb_blocks=args.nb_blocks,
468 dropout=args.dropout,
473 nb_parameters = sum(p.numel() for p in model.parameters())
474 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
476 ######################################################################
478 nb_epochs_finished = 0
480 if args.no_checkpoint:
481 log_string(f"not trying to load checkpoint.")
485 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
486 checkpoint = torch.load(checkpoint_name)
487 nb_epochs_finished = checkpoint["nb_epochs_finished"]
488 model.load_state_dict(checkpoint["model_state"])
489 torch.set_rng_state(checkpoint["rng_state"])
490 if torch.cuda.is_available():
491 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
493 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
495 except FileNotFoundError:
496 log_string("starting from scratch.")
499 log_string("error when loading the checkpoint.")
502 ######################################################################
504 if args.task == "expr" and args.expr_input_file is not None:
505 task.produce_results(
510 args.deterministic_synthesis,
511 args.expr_input_file,
516 ######################################################################
518 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
520 # Compute the entropy of the training tokens
523 for input in task.batches(split="train"):
524 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
525 token_probas = token_count / token_count.sum()
526 entropy = -torch.xlogy(token_probas, token_probas).sum()
527 train_set_perplexity = math.exp(entropy)
529 ##############################
531 # A bit of paranoia never hurts
536 for input in task.batches(split="train"):
537 assert input.dim() == 2 and input.dtype == torch.int64
539 train_examples[x.sum().item()] = x
541 nb_total, nb_collisions = 0, 0
542 for input in task.batches(split="test"):
543 assert input.dim() == 2 and input.dtype == torch.int64
546 y = train_examples.get(x.sum().item())
548 if x.size() == y.size() and (x - y).abs().sum() == 0:
554 f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
557 ##############################
559 if args.learning_rate_schedule == "cos":
560 learning_rate_schedule = {}
561 for n_epoch in range(args.nb_epochs):
562 u = n_epoch / args.nb_epochs * math.pi
563 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
568 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
572 learning_rate_schedule = {}
573 learning_rate = args.learning_rate
574 for n_epoch in range(args.nb_epochs):
576 learning_rate = u[n_epoch]
577 learning_rate_schedule[n_epoch] = learning_rate
579 log_string(f"learning_rate_schedule {learning_rate_schedule}")
581 ##############################
585 if nb_epochs_finished >= nb_epochs:
586 task.produce_results(
591 args.deterministic_synthesis,
594 for n_epoch in range(nb_epochs_finished, nb_epochs):
595 learning_rate = learning_rate_schedule[n_epoch]
597 log_string(f"learning_rate {learning_rate}")
599 if args.optim == "sgd":
600 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
601 elif args.optim == "adam":
602 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
603 elif args.optim == "adamw":
604 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
606 raise ValueError(f"Unknown optimizer {args.optim}.")
610 nb_train_samples, acc_train_loss = 0, 0.0
612 for input in task.batches(split="train"):
613 input = input.to(device)
614 output = model(mygpt.BracketedSequence(input)).x
615 loss = F.cross_entropy(output.transpose(1, 2), input)
616 acc_train_loss += loss.item() * input.size(0)
617 nb_train_samples += input.size(0)
618 nb_samples_seen += input.size(0)
620 optimizer.zero_grad()
624 with torch.autograd.no_grad():
627 nb_test_samples, acc_test_loss = 0, 0.0
629 for input in task.batches(split="test"):
630 input = input.to(device)
632 output = model(mygpt.BracketedSequence(input)).x
633 loss = F.cross_entropy(output.transpose(1, 2), input)
634 acc_test_loss += loss.item() * input.size(0)
635 nb_test_samples += input.size(0)
637 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
638 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
641 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
644 task.produce_results(
645 n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
649 "nb_epochs_finished": n_epoch + 1,
650 "model_state": model.state_dict(),
651 "rng_state": torch.get_rng_state(),
654 if torch.cuda.is_available():
655 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
657 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
658 torch.save(checkpoint, checkpoint_name)
659 log_string(f"saved checkpoint {checkpoint_name}")
661 ######################################################################