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
17 import mygpt, tensorstack
19 ######################################################################
21 if torch.cuda.is_available():
22 device = torch.device("cuda")
23 torch.backends.cuda.matmul.allow_tf32 = True
25 device = torch.device("cpu")
27 ######################################################################
29 parser = argparse.ArgumentParser(
30 description="An implementation of GPT with cache.",
31 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
35 "--task", type=str, default="picoclvr", help="picoclvr, mnist, maze, snake, stack, expr"
38 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
40 parser.add_argument("--result_dir", type=str, default=None)
42 parser.add_argument("--seed", type=int, default=0)
44 parser.add_argument("--nb_epochs", type=int, default=None)
46 parser.add_argument("--batch_size", type=int, default=None)
48 parser.add_argument("--nb_train_samples", type=int, default=None)
50 parser.add_argument("--nb_test_samples", type=int, default=None)
52 parser.add_argument("--optim", type=str, default="adam")
54 parser.add_argument("--learning_rate", type=float, default=1e-4)
56 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
58 parser.add_argument("--dim_model", type=int, default=512)
60 parser.add_argument("--dim_keys", type=int, default=64)
62 parser.add_argument("--dim_hidden", type=int, default=2048)
64 parser.add_argument("--nb_heads", type=int, default=8)
66 parser.add_argument("--nb_blocks", type=int, default=12)
68 parser.add_argument("--dropout", type=float, default=0.1)
70 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
72 parser.add_argument("--no_checkpoint", action="store_true", default=False)
74 parser.add_argument("--overwrite_results", action="store_true", default=False)
76 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
78 ##############################
81 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
83 parser.add_argument("--picoclvr_height", type=int, default=12)
85 parser.add_argument("--picoclvr_width", type=int, default=16)
87 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
89 ##############################
92 parser.add_argument("--maze_height", type=int, default=13)
94 parser.add_argument("--maze_width", type=int, default=21)
96 parser.add_argument("--maze_nb_walls", type=int, default=15)
98 ##############################
101 parser.add_argument("--snake_height", type=int, default=6)
103 parser.add_argument("--snake_width", type=int, default=8)
105 parser.add_argument("--snake_nb_colors", type=int, default=5)
107 parser.add_argument("--snake_length", type=int, default=200)
109 ##############################
112 parser.add_argument("--stack_nb_steps", type=int, default=100)
114 parser.add_argument("--stack_nb_stacks", type=int, default=1)
116 parser.add_argument("--stack_nb_digits", type=int, default=3)
118 parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
120 ######################################################################
122 args = parser.parse_args()
124 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
126 if args.result_dir is None:
127 args.result_dir = f"results_{args.task}"
129 ######################################################################
135 "nb_train_samples": 250000,
136 "nb_test_samples": 10000,
141 "nb_train_samples": 250000,
142 "nb_test_samples": 10000,
147 "nb_train_samples": 250000,
148 "nb_test_samples": 10000,
153 "nb_train_samples": 250000,
154 "nb_test_samples": 10000,
159 "nb_train_samples": 100000,
160 "nb_test_samples": 1000,
165 "nb_train_samples": 100000,
166 "nb_test_samples": 1000,
170 if args.task in default_args:
171 for k, v in default_args[args.task].items():
172 if getattr(args, k) is None:
175 ######################################################################
178 os.mkdir(args.result_dir)
179 except FileExistsError:
180 if not args.overwrite_results:
181 print(f"result directory {args.result_dir} already exists")
184 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
187 # torch.backends.cudnn.deterministic = True
188 # torch.backends.cudnn.benchmark = False
189 # torch.use_deterministic_algorithms(True)
190 torch.manual_seed(args.seed)
191 if torch.cuda.is_available():
192 torch.cuda.manual_seed_all(args.seed)
194 ######################################################################
198 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
200 if log_file is not None:
201 log_file.write(t + s + "\n")
209 log_string(f"args.{n} {getattr(args, n)}")
211 ######################################################################
214 # ra_mask is boolean, with 1s on the values to generate
217 def masked_inplace_autoregression(
222 forbidden_tokens=None,
223 progress_bar_desc="autoregression",
224 device=torch.device("cpu"),
227 batches = zip(input.split(batch_size), ar_mask.split(batch_size))
229 if progress_bar_desc is not None:
233 desc=progress_bar_desc,
234 total=input.size(0) // batch_size,
237 for input, ar_mask in batches:
238 i = (ar_mask.sum(0) > 0).nonzero()
241 mygpt.BracketedSequence(input, 0, i.min())
242 ) # Needed to initialize the model's cache
243 for s in range(i.min(), i.max() + 1):
244 output = model(mygpt.BracketedSequence(input, s, 1)).x
245 logits = output[:, s]
246 if forbidden_tokens is not None:
247 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
248 if args.deterministic_synthesis:
249 t_next = logits.argmax(1)
251 dist = torch.distributions.categorical.Categorical(logits=logits)
252 t_next = dist.sample()
253 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
256 ######################################################################
260 def batches(self, split="train"):
263 def vocabulary_size(self):
266 def produce_results(self, n_epoch, model):
270 ######################################################################
275 class TaskPicoCLVR(Task):
276 # Make a tensor from a list of strings
277 def tensorize(self, descr):
278 token_descr = [s.strip().split(" ") for s in descr]
279 l = max([len(s) for s in token_descr])
280 token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
281 id_descr = [[self.token2id[u] for u in s] for s in token_descr]
282 return torch.tensor(id_descr, device=self.device)
284 # Make a list of strings from a tensor
285 def detensorize(self, x):
286 return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
288 # trim all the tensors in the tuple z to remove as much token from
289 # left and right in the first tensor. If z is a tuple, all its
290 # elements are trimed according to the triming for the first
291 def trim(self, z, token="<nul>"):
292 n = self.token2id[token]
295 i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
296 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
297 return tuple([t[:, a:b] for t in z])
299 i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
300 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
303 ######################
304 # Not the cleanest part of the code
306 # Extract the last image of each sequence, from the last <img>
307 # included, and set to <nul> all the tokens from the beginning of
308 # that image to the end
309 def excise_last_image(self, input):
310 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
311 nb_img_tokens = self.height * self.width + 1
313 input = input.clone()
314 t = (input == t_img).long()
315 tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
316 i = (t * tail_masks).nonzero(as_tuple=True)
319 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
321 images = self.trim(input[j])
323 loss_masks = 1 - tail_masks
324 input, loss_masks = self.trim((input, loss_masks))
325 return input, loss_masks, images
327 def add_true_image(self, input, images, loss_masks):
328 t_nul = self.token2id["<nul>"]
329 nb_img_tokens = self.height * self.width + 1
330 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
331 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
332 t = (input == t_nul).long()
333 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
336 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
340 input, loss_masks = self.trim((input, loss_masks))
341 return input, loss_masks
343 def add_generated_image(self, input, loss_masks, model):
344 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
345 nb_img_tokens = self.height * self.width + 1
347 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
348 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
349 t = (input == t_nul).long()
350 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
357 + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
359 ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
362 torch.arange(self.vocabulary_size(), device=input.device) == t_nul
364 with torch.autograd.no_grad():
367 masked_inplace_autoregression(
373 progress_bar_desc=None,
378 input, loss_masks = self.trim((input, loss_masks))
380 return input, loss_masks
382 ######################
392 device=torch.device("cpu"),
396 def generate_descr(nb, cache_suffix, pruner):
397 return picoclvr.generate(
407 self.batch_size = batch_size
409 self.pruner_train = pruner_train
410 self.pruner_eval = pruner_eval
413 "nb_train_samples": nb_train_samples,
414 "nb_test_samples": nb_test_samples,
417 "nb_colors": nb_colors,
418 "batch_size": batch_size,
419 "rng_state": list(torch.get_rng_state()),
423 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
425 self.train_descr = generate_descr(
426 nb_train_samples, "train", pruner=self.pruner_train
428 self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
430 # Build the tokenizer
431 tokens = {"<nul>", "<img>"}
432 for d in [self.train_descr, self.test_descr]:
434 for t in s.strip().split(" "):
436 # make this set a sorted list to get the same tensors given
438 tokens = list(tokens)
440 self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
441 self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
443 # Tokenize the train and test sets
444 self.train_input = self.tensorize(self.train_descr)
445 self.test_input = self.tensorize(self.test_descr)
447 def batches(self, split="train"):
448 assert split in {"train", "test"}
449 input = self.train_input if split == "train" else self.test_input
450 for batch in tqdm.tqdm(
451 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
453 yield self.trim(batch)
455 def vocabulary_size(self):
456 return len(self.token2id)
458 def compute_missing_properties(self, n_epoch, model, pruner=None):
459 acc_nb_requested_properties = []
460 acc_nb_missing_properties = []
463 for input in tqdm.tqdm(
464 self.test_input.split(self.batch_size),
466 desc=f"test-properties",
468 tape, loss_masks, _ = self.excise_last_image(input)
469 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
470 result_descr = self.detensorize(tape)
471 np = picoclvr.nb_properties(
477 nb_requested_properties, _, nb_missing_properties = zip(*np)
478 acc_nb_requested_properties += nb_requested_properties
479 acc_nb_missing_properties += nb_missing_properties
480 acc_nb_results += len(result_descr)
482 nb_requested_properties = sum(acc_nb_requested_properties)
483 nb_missing_properties = sum(acc_nb_missing_properties)
485 prefix = "" if pruner is None else "pruned_"
486 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
488 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
491 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
494 ######################################################################
496 def produce_results(self, n_epoch, model):
497 self.compute_missing_properties(n_epoch, model)
499 if self.pruner_eval is not None:
500 self.compute_missing_properties(n_epoch, model, self.pruner_eval)
502 nb_tokens_to_generate = self.height * self.width + 3
507 for primer_descr in [
508 "red above green <sep> green top <sep> blue right of red",
509 "there is red <sep> there is yellow <sep> there is blue",
510 "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
511 "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
513 primer += [primer_descr] * nb_per_primer
515 tape = self.tensorize(primer)
516 loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
517 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
518 result_descr = self.detensorize(tape)
520 np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
522 acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
523 acc_nb_results = len(result_descr)
525 nb_requested_properties = sum(acc_nb_requested_properties)
526 nb_missing_properties = sum(acc_nb_missing_properties)
529 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
531 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
534 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
537 img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
541 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
545 torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
551 image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
552 torchvision.utils.save_image(
553 img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
555 log_string(f"wrote {image_name}")
558 ######################################################################
561 class TaskMNIST(Task):
562 def __init__(self, batch_size, device=torch.device("cpu")):
564 self.batch_size = batch_size
566 def batches(self, split="train"):
567 assert split in {"train", "test"}
568 data_set = torchvision.datasets.MNIST(
569 root="./data", train=(split == "train"), download=True
571 data_input = data_set.data.view(-1, 28 * 28).long()
572 if args.nb_train_samples is not None:
573 data_input = data_input[: args.nb_train_samples]
574 for batch in tqdm.tqdm(
575 data_input.split(self.batch_size), desc=f"epoch-{split}"
579 def vocabulary_size(self):
582 def produce_results(self, n_epoch, model):
583 results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
584 ar_mask = torch.full_like(results, 1)
585 masked_inplace_autoregression(
586 model, self.batch_size, results, ar_mask, device=self.device
588 image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
589 torchvision.utils.save_image(
590 1 - results.reshape(-1, 1, 28, 28) / 255.0,
595 log_string(f"wrote {image_name}")
598 ######################################################################
603 class TaskMaze(Task):
604 def map2seq(self, *m):
605 return torch.cat([x.flatten(1) for x in m], 1)
607 def seq2map(self, s):
608 s = s.reshape(s.size(0), -1, self.height, self.width)
609 return (s[:, k] for k in range(s.size(1)))
619 device=torch.device("cpu"),
621 self.batch_size = batch_size
626 train_mazes, train_paths, _ = maze.create_maze_data(
631 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
633 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
635 test_mazes, test_paths, _ = maze.create_maze_data(
640 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
642 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
644 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
646 def batches(self, split="train", nb_to_use=-1, desc=None):
647 assert split in {"train", "test"}
648 input = self.train_input if split == "train" else self.test_input
650 input = input[:nb_to_use]
652 desc = f"epoch-{split}"
653 for batch in tqdm.tqdm(
654 input.split(self.batch_size), dynamic_ncols=True, desc=desc
658 def vocabulary_size(self):
661 def compute_error(self, model, split="train", nb_to_use=-1):
662 nb_total, nb_correct = 0, 0
664 self.width * self.height,
665 self.width * self.height,
669 for input in tqdm.tqdm(
670 task.batches(split, nb_to_use),
674 result = input.clone()
675 ar_mask = result.new_zeros(result.size())
676 ar_mask[:, self.height * self.width :] = 1
677 result *= 1 - ar_mask
678 masked_inplace_autoregression(
683 progress_bar_desc=None,
686 mazes, paths = self.seq2map(result)
687 path_correctness = maze.path_correctness(mazes, paths)
688 nb_correct += path_correctness.long().sum()
689 nb_total += mazes.size(0)
691 optimal_path_lengths = (
692 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
694 predicted_path_lengths = (
695 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
697 optimal_path_lengths = optimal_path_lengths[path_correctness]
698 predicted_path_lengths = predicted_path_lengths[path_correctness]
699 count[optimal_path_lengths, predicted_path_lengths] += 1
705 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
708 return nb_total, nb_correct, count
710 def produce_results(self, n_epoch, model):
711 with torch.autograd.no_grad():
715 train_nb_total, train_nb_correct, count = self.compute_error(
716 model, "train", nb_to_use=1000
719 f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
722 test_nb_total, test_nb_correct, count = self.compute_error(
723 model, "test", nb_to_use=1000
726 f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
729 if count is not None:
730 proportion_optimal = count.diagonal().sum().float() / count.sum()
731 log_string(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
733 os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
735 for i in range(count.size(0)):
736 for j in range(count.size(1)):
737 eol = " " if j < count.size(1) - 1 else "\n"
738 f.write(f"{count[i,j]}{eol}")
740 input = self.test_input[:48]
741 result = input.clone()
742 ar_mask = result.new_zeros(result.size())
743 ar_mask[:, self.height * self.width :] = 1
744 result *= 1 - ar_mask
745 masked_inplace_autoregression(
746 model, self.batch_size, result, ar_mask, device=self.device
749 mazes, paths = self.seq2map(input)
750 _, predicted_paths = self.seq2map(result)
752 filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
757 predicted_paths=predicted_paths,
758 path_correct=maze.path_correctness(mazes, predicted_paths),
759 path_optimal=maze.path_optimality(paths, predicted_paths),
761 log_string(f"wrote {filename}")
766 ######################################################################
772 class TaskSnake(Task):
783 device=torch.device("cpu"),
785 self.batch_size = batch_size
789 self.prompt_length = prompt_length
791 self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
800 self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
810 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
812 def batches(self, split="train", nb_to_use=-1, desc=None):
813 assert split in {"train", "test"}
814 input = self.train_input if split == "train" else self.test_input
816 input = input[:nb_to_use]
818 desc = f"epoch-{split}"
819 for batch in tqdm.tqdm(
820 input.split(self.batch_size), dynamic_ncols=True, desc=desc
824 def vocabulary_size(self):
827 def produce_results(self, n_epoch, model):
828 with torch.autograd.no_grad():
832 def compute_nb_correct(input, prior_visits):
833 result = input.clone()
834 i = torch.arange(result.size(1), device=result.device)[None, :]
836 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
840 result *= 1 - ar_mask
842 # snake.solver(result,ar_mask)
844 masked_inplace_autoregression(
845 model, self.batch_size, result, ar_mask, device=self.device
848 nb_total = ((prior_visits > 0) * ar_mask).sum()
851 (result == input).long() * (prior_visits > 0) * ar_mask
854 # nb_total = result.size(0)
855 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
857 return nb_total, nb_correct
859 # train_nb_total, train_nb_correct = compute_nb_correct(
860 # self.train_input, self.train_prior_visits
864 # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
867 test_nb_total, test_nb_correct = compute_nb_correct(
868 self.test_input[:1000], self.test_prior_visits[:1000]
872 f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
878 ######################################################################
884 class TaskStack(Task):
893 fraction_values_for_train=None,
894 device=torch.device("cpu"),
896 self.batch_size = batch_size
897 self.nb_steps = nb_steps
898 self.nb_stacks = nb_stacks
899 self.nb_digits = nb_digits
902 if fraction_values_for_train is None:
903 values_for_train = None
904 values_for_test = None
906 all = torch.randperm(10**nb_digits)
907 nb_for_train = int(all.size(0) * fraction_values_for_train)
908 values_for_train = all[:nb_for_train]
909 values_for_test = all[nb_for_train:]
911 self.train_input, self.train_stack_counts = stack.generate_sequences(
920 self.test_input, self.test_stack_counts = stack.generate_sequences(
929 i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
930 counts = self.test_stack_counts.flatten()[i.flatten()]
931 counts = F.one_hot(counts).sum(0)
932 log_string(f"test_pop_stack_counts {counts}")
934 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
936 def batches(self, split="train", nb_to_use=-1, desc=None):
937 assert split in {"train", "test"}
938 input = self.train_input if split == "train" else self.test_input
940 input = input[:nb_to_use]
942 desc = f"epoch-{split}"
943 for batch in tqdm.tqdm(
944 input.split(self.batch_size), dynamic_ncols=True, desc=desc
948 def vocabulary_size(self):
951 def produce_results(self, n_epoch, model):
952 with torch.autograd.no_grad():
956 def compute_nb_correct(input):
957 result = input.clone()
958 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
959 ar_mask = (result != input).long()
960 masked_inplace_autoregression(
961 model, self.batch_size, result, ar_mask, device=self.device
964 errors = ((result != input).long() * ar_mask).reshape(
965 -1, 1 + self.nb_digits
967 ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
969 nb_total = ar_mask.max(1).values.sum()
970 nb_correct = nb_total - errors.max(1).values.sum()
972 return nb_total, nb_correct
974 test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
977 f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
980 ##############################################################
981 # Log a few generated sequences
982 input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
983 result = input.clone()
984 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
985 ar_mask = (result != input).long()
986 for n in range(result.size(0)):
988 f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
990 masked_inplace_autoregression(
991 model, self.batch_size, result, ar_mask, device=self.device
993 for n in range(result.size(0)):
995 f"test_after {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
997 ##############################################################
1002 ######################################################################
1008 class TaskExpr(Task):
1014 device=torch.device("cpu"),
1016 self.batch_size = batch_size
1017 self.device = device
1019 train_sequences = expr.generate_sequences(nb_train_samples)
1020 test_sequences = expr.generate_sequences(nb_test_samples)
1021 self.char2id = dict([ (c,n) for n,c in enumerate(set("".join(train_sequences + test_sequences))) ])
1022 self.id2char = dict([ (n,c) for n,c in self.char2id.items() ])
1023 len_max = max([len(x) for x in train_sequences + test_sequences])
1024 self.train_input = torch.cat([torch.tensor([char2id(c) for c in s + " "*(len_max-len(s))] for s in train_sequences)], 0)
1025 self.test_input = torch.cat([torch.tensor([char2id(c) for c in s + " "*(len_max-len(s))] for s in test_sequences)], 0)
1026 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1028 def batches(self, split="train", nb_to_use=-1, desc=None):
1029 assert split in {"train", "test"}
1030 input = self.train_input if split == "train" else self.test_input
1032 input = input[:nb_to_use]
1034 desc = f"epoch-{split}"
1035 for batch in tqdm.tqdm(
1036 input.split(self.batch_size), dynamic_ncols=True, desc=desc
1040 def vocabulary_size(self):
1041 return self.nb_codes
1043 def produce_results(self, n_epoch, model):
1044 # !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
1045 with torch.autograd.no_grad():
1049 def compute_nb_correct(input):
1050 result = input.clone()
1051 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
1052 ar_mask = (result != input).long()
1053 masked_inplace_autoregression(
1054 model, self.batch_size, result, ar_mask, device=self.device
1057 errors = ((result != input).long() * ar_mask).reshape(
1058 -1, 1 + self.nb_digits
1060 ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
1062 nb_total = ar_mask.max(1).values.sum()
1063 nb_correct = nb_total - errors.max(1).values.sum()
1065 return nb_total, nb_correct
1067 test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
1070 f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
1073 ##############################################################
1074 # Log a few generated sequences
1075 input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
1076 result = input.clone()
1077 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
1078 ar_mask = (result != input).long()
1079 for n in range(result.size(0)):
1081 f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
1083 masked_inplace_autoregression(
1084 model, self.batch_size, result, ar_mask, device=self.device
1086 for n in range(result.size(0)):
1088 f"test_after {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
1090 ##############################################################
1095 ######################################################################
1098 def picoclvr_pruner_horizontal_green(p):
1099 return not ("green" in p and ("left" in p or "right" in p))
1102 picoclvr_pruner_train = (
1103 picoclvr_pruner_horizontal_green
1104 if args.picocvlr_prune_properties in {"train+eval"}
1108 picoclvr_pruner_eval = (
1109 (lambda p: not picoclvr_pruner_horizontal_green(p))
1110 if args.picocvlr_prune_properties in {"train+eval", "eval"}
1114 ######################################################################
1116 if args.task == "picoclvr":
1117 task = TaskPicoCLVR(
1118 nb_train_samples=args.nb_train_samples,
1119 nb_test_samples=args.nb_test_samples,
1120 batch_size=args.batch_size,
1121 height=args.picoclvr_height,
1122 width=args.picoclvr_width,
1123 nb_colors=args.picoclvr_nb_colors,
1125 pruner_train=picoclvr_pruner_train,
1126 pruner_eval=picoclvr_pruner_eval,
1129 elif args.task == "mnist":
1131 batch_size=args.batch_size,
1135 elif args.task == "maze":
1137 nb_train_samples=args.nb_train_samples,
1138 nb_test_samples=args.nb_test_samples,
1139 batch_size=args.batch_size,
1140 height=args.maze_height,
1141 width=args.maze_width,
1142 nb_walls=args.maze_nb_walls,
1146 elif args.task == "snake":
1148 nb_train_samples=args.nb_train_samples,
1149 nb_test_samples=args.nb_test_samples,
1150 batch_size=args.batch_size,
1151 height=args.snake_height,
1152 width=args.snake_width,
1153 nb_colors=args.snake_nb_colors,
1154 length=args.snake_length,
1155 prompt_length=args.snake_length // 2,
1159 elif args.task == "stack":
1161 nb_train_samples=args.nb_train_samples,
1162 nb_test_samples=args.nb_test_samples,
1163 batch_size=args.batch_size,
1164 nb_steps=args.stack_nb_steps,
1165 nb_stacks=args.stack_nb_stacks,
1166 nb_digits=args.stack_nb_digits,
1167 fraction_values_for_train=args.stack_fraction_values_for_train,
1171 elif args.task == "expr":
1173 nb_train_samples=args.nb_train_samples,
1174 nb_test_samples=args.nb_test_samples,
1175 batch_size=args.batch_size,
1180 raise ValueError(f"Unknown task {args.task}")
1182 ######################################################################
1184 log_string(f"device {device}")
1186 vocabulary_size = task.vocabulary_size()
1188 log_string(f"vocabulary_size {vocabulary_size}")
1190 ##############################
1192 model = mygpt.MyGPT(
1193 vocabulary_size=vocabulary_size,
1194 dim_model=args.dim_model,
1195 dim_keys=args.dim_keys,
1196 dim_hidden=args.dim_hidden,
1197 nb_heads=args.nb_heads,
1198 nb_blocks=args.nb_blocks,
1200 dropout=args.dropout,
1205 nb_parameters = sum(p.numel() for p in model.parameters())
1206 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
1208 ######################################################################
1210 nb_epochs_finished = 0
1212 if args.no_checkpoint:
1213 log_string(f"not trying to load checkpoint.")
1217 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1218 checkpoint = torch.load(checkpoint_name)
1219 nb_epochs_finished = checkpoint["nb_epochs_finished"]
1220 model.load_state_dict(checkpoint["model_state"])
1221 torch.set_rng_state(checkpoint["rng_state"])
1222 if torch.cuda.is_available():
1223 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
1225 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
1227 except FileNotFoundError:
1228 log_string("starting from scratch.")
1231 log_string("error when loading the checkpoint.")
1234 ######################################################################
1236 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
1239 for input in task.batches(split="train"):
1240 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
1241 token_probas = token_count / token_count.sum()
1242 entropy = -torch.xlogy(token_probas, token_probas).sum()
1243 train_set_perplexity = math.exp(entropy)
1245 ##############################
1247 if args.learning_rate_schedule == "cos":
1248 learning_rate_schedule = {}
1249 for n_epoch in range(args.nb_epochs):
1250 u = n_epoch / args.nb_epochs * math.pi
1251 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
1256 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
1260 learning_rate_schedule = {}
1261 learning_rate = args.learning_rate
1262 for n_epoch in range(args.nb_epochs):
1264 learning_rate = u[n_epoch]
1265 learning_rate_schedule[n_epoch] = learning_rate
1267 log_string(f"learning_rate_schedule {learning_rate_schedule}")
1269 ##############################
1273 if nb_epochs_finished >= nb_epochs:
1274 task.produce_results(nb_epochs_finished, model)
1276 for n_epoch in range(nb_epochs_finished, nb_epochs):
1277 learning_rate = learning_rate_schedule[n_epoch]
1279 log_string(f"learning_rate {learning_rate}")
1281 if args.optim == "sgd":
1282 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
1283 elif args.optim == "adam":
1284 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
1285 elif args.optim == "adamw":
1286 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
1288 raise ValueError(f"Unknown optimizer {args.optim}.")
1292 nb_train_samples, acc_train_loss = 0, 0.0
1294 for input in task.batches(split="train"):
1295 input = input.to(device)
1296 output = model(mygpt.BracketedSequence(input)).x
1297 loss = F.cross_entropy(output.transpose(1, 2), input)
1298 acc_train_loss += loss.item() * input.size(0)
1299 nb_train_samples += input.size(0)
1300 nb_samples_seen += input.size(0)
1302 optimizer.zero_grad()
1306 with torch.autograd.no_grad():
1309 nb_test_samples, acc_test_loss = 0, 0.0
1311 for input in task.batches(split="test"):
1312 input = input.to(device)
1314 output = model(mygpt.BracketedSequence(input)).x
1315 loss = F.cross_entropy(output.transpose(1, 2), input)
1316 acc_test_loss += loss.item() * input.size(0)
1317 nb_test_samples += input.size(0)
1319 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
1320 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
1323 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1326 task.produce_results(n_epoch, model)
1329 "nb_epochs_finished": n_epoch + 1,
1330 "model_state": model.state_dict(),
1331 "rng_state": torch.get_rng_state(),
1334 if torch.cuda.is_available():
1335 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1337 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1338 torch.save(checkpoint, checkpoint_name)
1339 log_string(f"saved checkpoint {checkpoint_name}")
1341 ######################################################################