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,
38 help="picoclvr, mnist, maze, snake, stack, expr",
41 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
43 parser.add_argument("--result_dir", type=str, default=None)
45 parser.add_argument("--seed", type=int, default=0)
47 parser.add_argument("--nb_epochs", type=int, default=None)
49 parser.add_argument("--batch_size", type=int, default=None)
51 parser.add_argument("--nb_train_samples", type=int, default=None)
53 parser.add_argument("--nb_test_samples", type=int, default=None)
55 parser.add_argument("--optim", type=str, default="adam")
57 parser.add_argument("--learning_rate", type=float, default=1e-4)
59 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
61 parser.add_argument("--dim_model", type=int, default=512)
63 parser.add_argument("--dim_keys", type=int, default=64)
65 parser.add_argument("--dim_hidden", type=int, default=2048)
67 parser.add_argument("--nb_heads", type=int, default=8)
69 parser.add_argument("--nb_blocks", type=int, default=12)
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("--picoclvr_nb_colors", type=int, default=5)
86 parser.add_argument("--picoclvr_height", type=int, default=12)
88 parser.add_argument("--picoclvr_width", type=int, default=16)
90 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
92 ##############################
95 parser.add_argument("--maze_height", type=int, default=13)
97 parser.add_argument("--maze_width", type=int, default=21)
99 parser.add_argument("--maze_nb_walls", type=int, default=15)
101 ##############################
104 parser.add_argument("--snake_height", type=int, default=6)
106 parser.add_argument("--snake_width", type=int, default=8)
108 parser.add_argument("--snake_nb_colors", type=int, default=5)
110 parser.add_argument("--snake_length", type=int, default=200)
112 ##############################
115 parser.add_argument("--stack_nb_steps", type=int, default=100)
117 parser.add_argument("--stack_nb_stacks", type=int, default=1)
119 parser.add_argument("--stack_nb_digits", type=int, default=3)
121 parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
123 ######################################################################
125 args = parser.parse_args()
127 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
129 if args.result_dir is None:
130 args.result_dir = f"results_{args.task}"
132 ######################################################################
138 "nb_train_samples": 250000,
139 "nb_test_samples": 10000,
144 "nb_train_samples": 250000,
145 "nb_test_samples": 10000,
150 "nb_train_samples": 250000,
151 "nb_test_samples": 10000,
156 "nb_train_samples": 250000,
157 "nb_test_samples": 10000,
162 "nb_train_samples": 100000,
163 "nb_test_samples": 1000,
168 "nb_train_samples": 100000,
169 "nb_test_samples": 1000,
173 if args.task in default_args:
174 for k, v in default_args[args.task].items():
175 if getattr(args, k) is None:
178 ######################################################################
181 os.mkdir(args.result_dir)
182 except FileExistsError:
183 if not args.overwrite_results:
184 print(f"result directory {args.result_dir} already exists")
187 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
190 # torch.backends.cudnn.deterministic = True
191 # torch.backends.cudnn.benchmark = False
192 # torch.use_deterministic_algorithms(True)
193 torch.manual_seed(args.seed)
194 if torch.cuda.is_available():
195 torch.cuda.manual_seed_all(args.seed)
197 ######################################################################
201 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
203 if log_file is not None:
204 log_file.write(t + s + "\n")
212 log_string(f"args.{n} {getattr(args, n)}")
214 ######################################################################
217 # ra_mask is boolean, with 1s on the values to generate
220 def masked_inplace_autoregression(
225 forbidden_tokens=None,
226 progress_bar_desc="autoregression",
227 device=torch.device("cpu"),
229 batches = zip(input.split(batch_size), ar_mask.split(batch_size))
231 if progress_bar_desc is not None:
235 desc=progress_bar_desc,
236 total=input.size(0) // batch_size,
239 for input, ar_mask in batches:
240 i = (ar_mask.sum(0) > 0).nonzero()
243 mygpt.BracketedSequence(input, 0, i.min())
244 ) # Needed to initialize the model's cache
245 for s in range(i.min(), i.max() + 1):
246 output = model(mygpt.BracketedSequence(input, s, 1)).x
247 logits = output[:, s]
248 if forbidden_tokens is not None:
249 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
250 if args.deterministic_synthesis:
251 t_next = logits.argmax(1)
253 dist = torch.distributions.categorical.Categorical(logits=logits)
254 t_next = dist.sample()
255 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
258 ######################################################################
262 def batches(self, split="train"):
265 def vocabulary_size(self):
268 def produce_results(self, n_epoch, model):
272 ######################################################################
277 class TaskPicoCLVR(Task):
278 # Make a tensor from a list of strings
279 def tensorize(self, descr):
280 token_descr = [s.strip().split(" ") for s in descr]
281 l = max([len(s) for s in token_descr])
282 token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
283 id_descr = [[self.token2id[u] for u in s] for s in token_descr]
284 return torch.tensor(id_descr, device=self.device)
286 # Make a list of strings from a tensor
287 def detensorize(self, x):
288 return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
290 # trim all the tensors in the tuple z to remove as much token from
291 # left and right in the first tensor. If z is a tuple, all its
292 # elements are trimed according to the triming for the first
293 def trim(self, z, token="<nul>"):
294 n = self.token2id[token]
297 i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
298 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
299 return tuple([t[:, a:b] for t in z])
301 i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
302 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
305 ######################
306 # Not the cleanest part of the code
308 # Extract the last image of each sequence, from the last <img>
309 # included, and set to <nul> all the tokens from the beginning of
310 # that image to the end
311 def excise_last_image(self, input):
312 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
313 nb_img_tokens = self.height * self.width + 1
315 input = input.clone()
316 t = (input == t_img).long()
317 tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
318 i = (t * tail_masks).nonzero(as_tuple=True)
321 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
323 images = self.trim(input[j])
325 loss_masks = 1 - tail_masks
326 input, loss_masks = self.trim((input, loss_masks))
327 return input, loss_masks, images
329 def add_true_image(self, input, images, loss_masks):
330 t_nul = self.token2id["<nul>"]
331 nb_img_tokens = self.height * self.width + 1
332 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
333 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
334 t = (input == t_nul).long()
335 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
338 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
342 input, loss_masks = self.trim((input, loss_masks))
343 return input, loss_masks
345 def add_generated_image(self, input, loss_masks, model):
346 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
347 nb_img_tokens = self.height * self.width + 1
349 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
350 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
351 t = (input == t_nul).long()
352 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
359 + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
361 ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
364 torch.arange(self.vocabulary_size(), device=input.device) == t_nul
366 with torch.autograd.no_grad():
369 masked_inplace_autoregression(
375 progress_bar_desc=None,
380 input, loss_masks = self.trim((input, loss_masks))
382 return input, loss_masks
384 ######################
394 device=torch.device("cpu"),
398 def generate_descr(nb, cache_suffix, pruner):
399 return picoclvr.generate(
409 self.batch_size = batch_size
411 self.pruner_train = pruner_train
412 self.pruner_eval = pruner_eval
415 "nb_train_samples": nb_train_samples,
416 "nb_test_samples": nb_test_samples,
419 "nb_colors": nb_colors,
420 "batch_size": batch_size,
421 "rng_state": list(torch.get_rng_state()),
425 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
427 self.train_descr = generate_descr(
428 nb_train_samples, "train", pruner=self.pruner_train
430 self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
432 # Build the tokenizer
433 tokens = {"<nul>", "<img>"}
434 for d in [self.train_descr, self.test_descr]:
436 for t in s.strip().split(" "):
438 # make this set a sorted list to get the same tensors given
440 tokens = list(tokens)
442 self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
443 self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
445 # Tokenize the train and test sets
446 self.train_input = self.tensorize(self.train_descr)
447 self.test_input = self.tensorize(self.test_descr)
449 def batches(self, split="train"):
450 assert split in {"train", "test"}
451 input = self.train_input if split == "train" else self.test_input
452 for batch in tqdm.tqdm(
453 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
455 yield self.trim(batch)
457 def vocabulary_size(self):
458 return len(self.token2id)
460 def compute_missing_properties(self, n_epoch, model, pruner=None):
461 acc_nb_requested_properties = []
462 acc_nb_missing_properties = []
465 for input in tqdm.tqdm(
466 self.test_input.split(self.batch_size),
468 desc=f"test-properties",
470 tape, loss_masks, _ = self.excise_last_image(input)
471 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
472 result_descr = self.detensorize(tape)
473 np = picoclvr.nb_properties(
479 nb_requested_properties, _, nb_missing_properties = zip(*np)
480 acc_nb_requested_properties += nb_requested_properties
481 acc_nb_missing_properties += nb_missing_properties
482 acc_nb_results += len(result_descr)
484 nb_requested_properties = sum(acc_nb_requested_properties)
485 nb_missing_properties = sum(acc_nb_missing_properties)
487 prefix = "" if pruner is None else "pruned_"
488 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
490 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
493 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
496 ######################################################################
498 def produce_results(self, n_epoch, model):
499 self.compute_missing_properties(n_epoch, model)
501 if self.pruner_eval is not None:
502 self.compute_missing_properties(n_epoch, model, self.pruner_eval)
504 nb_tokens_to_generate = self.height * self.width + 3
509 for primer_descr in [
510 "red above green <sep> green top <sep> blue right of red",
511 "there is red <sep> there is yellow <sep> there is blue",
512 "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
513 "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
515 primer += [primer_descr] * nb_per_primer
517 tape = self.tensorize(primer)
518 loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
519 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
520 result_descr = self.detensorize(tape)
522 np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
524 acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
525 acc_nb_results = len(result_descr)
527 nb_requested_properties = sum(acc_nb_requested_properties)
528 nb_missing_properties = sum(acc_nb_missing_properties)
531 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
533 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
536 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
539 img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
543 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
547 torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
553 image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
554 torchvision.utils.save_image(
555 img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
557 log_string(f"wrote {image_name}")
560 ######################################################################
563 class TaskMNIST(Task):
564 def __init__(self, batch_size, device=torch.device("cpu")):
566 self.batch_size = batch_size
568 def batches(self, split="train"):
569 assert split in {"train", "test"}
570 data_set = torchvision.datasets.MNIST(
571 root="./data", train=(split == "train"), download=True
573 data_input = data_set.data.view(-1, 28 * 28).long()
574 if args.nb_train_samples is not None:
575 data_input = data_input[: args.nb_train_samples]
576 for batch in tqdm.tqdm(
577 data_input.split(self.batch_size), desc=f"epoch-{split}"
581 def vocabulary_size(self):
584 def produce_results(self, n_epoch, model):
585 results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
586 ar_mask = torch.full_like(results, 1)
587 masked_inplace_autoregression(
588 model, self.batch_size, results, ar_mask, device=self.device
590 image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
591 torchvision.utils.save_image(
592 1 - results.reshape(-1, 1, 28, 28) / 255.0,
597 log_string(f"wrote {image_name}")
600 ######################################################################
605 class TaskMaze(Task):
606 def map2seq(self, *m):
607 return torch.cat([x.flatten(1) for x in m], 1)
609 def seq2map(self, s):
610 s = s.reshape(s.size(0), -1, self.height, self.width)
611 return (s[:, k] for k in range(s.size(1)))
621 device=torch.device("cpu"),
623 self.batch_size = batch_size
628 train_mazes, train_paths, _ = maze.create_maze_data(
633 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
635 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
637 test_mazes, test_paths, _ = maze.create_maze_data(
642 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
644 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
646 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
648 def batches(self, split="train", nb_to_use=-1, desc=None):
649 assert split in {"train", "test"}
650 input = self.train_input if split == "train" else self.test_input
652 input = input[:nb_to_use]
654 desc = f"epoch-{split}"
655 for batch in tqdm.tqdm(
656 input.split(self.batch_size), dynamic_ncols=True, desc=desc
660 def vocabulary_size(self):
663 def compute_error(self, model, split="train", nb_to_use=-1):
664 nb_total, nb_correct = 0, 0
666 self.width * self.height,
667 self.width * self.height,
671 for input in tqdm.tqdm(
672 task.batches(split, nb_to_use),
676 result = input.clone()
677 ar_mask = result.new_zeros(result.size())
678 ar_mask[:, self.height * self.width :] = 1
679 result *= 1 - ar_mask
680 masked_inplace_autoregression(
685 progress_bar_desc=None,
688 mazes, paths = self.seq2map(result)
689 path_correctness = maze.path_correctness(mazes, paths)
690 nb_correct += path_correctness.long().sum()
691 nb_total += mazes.size(0)
693 optimal_path_lengths = (
694 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
696 predicted_path_lengths = (
697 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
699 optimal_path_lengths = optimal_path_lengths[path_correctness]
700 predicted_path_lengths = predicted_path_lengths[path_correctness]
701 count[optimal_path_lengths, predicted_path_lengths] += 1
707 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
710 return nb_total, nb_correct, count
712 def produce_results(self, n_epoch, model):
713 with torch.autograd.no_grad():
717 train_nb_total, train_nb_correct, count = self.compute_error(
718 model, "train", nb_to_use=1000
721 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}%"
724 test_nb_total, test_nb_correct, count = self.compute_error(
725 model, "test", nb_to_use=1000
728 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}%"
731 if count is not None:
732 proportion_optimal = count.diagonal().sum().float() / count.sum()
733 log_string(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
735 os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
737 for i in range(count.size(0)):
738 for j in range(count.size(1)):
739 eol = " " if j < count.size(1) - 1 else "\n"
740 f.write(f"{count[i,j]}{eol}")
742 input = self.test_input[:48]
743 result = input.clone()
744 ar_mask = result.new_zeros(result.size())
745 ar_mask[:, self.height * self.width :] = 1
746 result *= 1 - ar_mask
747 masked_inplace_autoregression(
748 model, self.batch_size, result, ar_mask, device=self.device
751 mazes, paths = self.seq2map(input)
752 _, predicted_paths = self.seq2map(result)
754 filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
759 predicted_paths=predicted_paths,
760 path_correct=maze.path_correctness(mazes, predicted_paths),
761 path_optimal=maze.path_optimality(paths, predicted_paths),
763 log_string(f"wrote {filename}")
768 ######################################################################
774 class TaskSnake(Task):
785 device=torch.device("cpu"),
787 self.batch_size = batch_size
791 self.prompt_length = prompt_length
793 self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
802 self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
812 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
814 def batches(self, split="train", nb_to_use=-1, desc=None):
815 assert split in {"train", "test"}
816 input = self.train_input if split == "train" else self.test_input
818 input = input[:nb_to_use]
820 desc = f"epoch-{split}"
821 for batch in tqdm.tqdm(
822 input.split(self.batch_size), dynamic_ncols=True, desc=desc
826 def vocabulary_size(self):
829 def produce_results(self, n_epoch, model):
830 with torch.autograd.no_grad():
834 def compute_nb_correct(input, prior_visits):
835 result = input.clone()
836 i = torch.arange(result.size(1), device=result.device)[None, :]
838 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
842 result *= 1 - ar_mask
844 # snake.solver(result,ar_mask)
846 masked_inplace_autoregression(
847 model, self.batch_size, result, ar_mask, device=self.device
850 nb_total = ((prior_visits > 0) * ar_mask).sum()
853 (result == input).long() * (prior_visits > 0) * ar_mask
856 # nb_total = result.size(0)
857 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
859 return nb_total, nb_correct
861 # train_nb_total, train_nb_correct = compute_nb_correct(
862 # self.train_input, self.train_prior_visits
866 # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
869 test_nb_total, test_nb_correct = compute_nb_correct(
870 self.test_input[:1000], self.test_prior_visits[:1000]
874 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}%"
880 ######################################################################
886 class TaskStack(Task):
895 fraction_values_for_train=None,
896 device=torch.device("cpu"),
898 self.batch_size = batch_size
899 self.nb_steps = nb_steps
900 self.nb_stacks = nb_stacks
901 self.nb_digits = nb_digits
904 if fraction_values_for_train is None:
905 values_for_train = None
906 values_for_test = None
908 all = torch.randperm(10**nb_digits)
909 nb_for_train = int(all.size(0) * fraction_values_for_train)
910 values_for_train = all[:nb_for_train]
911 values_for_test = all[nb_for_train:]
913 self.train_input, self.train_stack_counts = stack.generate_sequences(
922 self.test_input, self.test_stack_counts = stack.generate_sequences(
931 i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
932 counts = self.test_stack_counts.flatten()[i.flatten()]
933 counts = F.one_hot(counts).sum(0)
934 log_string(f"test_pop_stack_counts {counts}")
936 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
938 def batches(self, split="train", nb_to_use=-1, desc=None):
939 assert split in {"train", "test"}
940 input = self.train_input if split == "train" else self.test_input
942 input = input[:nb_to_use]
944 desc = f"epoch-{split}"
945 for batch in tqdm.tqdm(
946 input.split(self.batch_size), dynamic_ncols=True, desc=desc
950 def vocabulary_size(self):
953 def produce_results(self, n_epoch, model):
954 with torch.autograd.no_grad():
958 def compute_nb_correct(input):
959 result = input.clone()
960 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
961 ar_mask = (result != input).long()
962 masked_inplace_autoregression(
963 model, self.batch_size, result, ar_mask, device=self.device
966 errors = ((result != input).long() * ar_mask).reshape(
967 -1, 1 + self.nb_digits
969 ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
971 nb_total = ar_mask.max(1).values.sum()
972 nb_correct = nb_total - errors.max(1).values.sum()
974 return nb_total, nb_correct
976 test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
979 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}%"
982 ##############################################################
983 # Log a few generated sequences
984 input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
985 result = input.clone()
986 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
987 ar_mask = (result != input).long()
988 for n in range(result.size(0)):
990 f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
992 masked_inplace_autoregression(
993 model, self.batch_size, result, ar_mask, device=self.device
995 for n in range(result.size(0)):
997 f"test_after {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
999 ##############################################################
1004 ######################################################################
1010 class TaskExpr(Task):
1016 device=torch.device("cpu"),
1018 self.batch_size = batch_size
1019 self.device = device
1021 train_sequences = expr.generate_sequences(nb_train_samples)
1022 test_sequences = expr.generate_sequences(nb_test_samples)
1023 self.char2id = dict(
1026 for n, c in enumerate(set("#"+"".join(train_sequences + test_sequences)))
1029 self.id2char = dict([(n, c) for c, n in self.char2id.items()])
1030 len_max = max([len(x) for x in train_sequences + test_sequences])
1031 self.train_input = torch.cat(
1035 [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1036 for s in train_sequences
1042 self.test_input = torch.cat(
1046 [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1047 for s in test_sequences
1053 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1055 def batches(self, split="train", nb_to_use=-1, desc=None):
1056 assert split in {"train", "test"}
1057 input = self.train_input if split == "train" else self.test_input
1059 input = input[:nb_to_use]
1061 desc = f"epoch-{split}"
1062 for batch in tqdm.tqdm(
1063 input.split(self.batch_size), dynamic_ncols=True, desc=desc
1067 def vocabulary_size(self):
1068 return self.nb_codes
1070 def produce_results(self, n_epoch, model):
1071 with torch.autograd.no_grad():
1075 def compute_nb_correct(input):
1076 result = input.clone()
1077 space = self.char2id["#"]
1078 ar_mask = (result == space).long().cumsum(dim=1).clamp(max=1)
1079 result = (1 - ar_mask) * result + space * ar_mask
1080 masked_inplace_autoregression(
1081 model, self.batch_size, result, ar_mask, device=self.device
1084 nb_total = ar_mask.sum()
1085 nb_correct = ((input == result).long() * ar_mask).sum()
1087 return nb_total, nb_correct
1089 test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
1092 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}%"
1095 ##############################################################
1096 # Log a few generated sequences
1097 input = self.test_input[:10]
1098 result = input.clone()
1099 space = self.char2id["#"]
1100 ar_mask = (result == space).long().cumsum(dim=1).clamp(max=1)
1101 result = (1 - ar_mask) * result + space * ar_mask
1102 for n in range(result.size(0)):
1103 s = "".join([self.id2char[k.item()] for k in result[n]])
1104 log_string(f"test_before {s}")
1105 masked_inplace_autoregression(
1106 model, self.batch_size, result, ar_mask, device=self.device
1108 for n in range(result.size(0)):
1109 s = "".join([self.id2char[k.item()] for k in result[n]])
1110 log_string(f"test_after {s}")
1111 ##############################################################
1116 ######################################################################
1119 def picoclvr_pruner_horizontal_green(p):
1120 return not ("green" in p and ("left" in p or "right" in p))
1123 picoclvr_pruner_train = (
1124 picoclvr_pruner_horizontal_green
1125 if args.picocvlr_prune_properties in {"train+eval"}
1129 picoclvr_pruner_eval = (
1130 (lambda p: not picoclvr_pruner_horizontal_green(p))
1131 if args.picocvlr_prune_properties in {"train+eval", "eval"}
1135 ######################################################################
1137 if args.task == "picoclvr":
1138 task = TaskPicoCLVR(
1139 nb_train_samples=args.nb_train_samples,
1140 nb_test_samples=args.nb_test_samples,
1141 batch_size=args.batch_size,
1142 height=args.picoclvr_height,
1143 width=args.picoclvr_width,
1144 nb_colors=args.picoclvr_nb_colors,
1146 pruner_train=picoclvr_pruner_train,
1147 pruner_eval=picoclvr_pruner_eval,
1150 elif args.task == "mnist":
1152 batch_size=args.batch_size,
1156 elif args.task == "maze":
1158 nb_train_samples=args.nb_train_samples,
1159 nb_test_samples=args.nb_test_samples,
1160 batch_size=args.batch_size,
1161 height=args.maze_height,
1162 width=args.maze_width,
1163 nb_walls=args.maze_nb_walls,
1167 elif args.task == "snake":
1169 nb_train_samples=args.nb_train_samples,
1170 nb_test_samples=args.nb_test_samples,
1171 batch_size=args.batch_size,
1172 height=args.snake_height,
1173 width=args.snake_width,
1174 nb_colors=args.snake_nb_colors,
1175 length=args.snake_length,
1176 prompt_length=args.snake_length // 2,
1180 elif args.task == "stack":
1182 nb_train_samples=args.nb_train_samples,
1183 nb_test_samples=args.nb_test_samples,
1184 batch_size=args.batch_size,
1185 nb_steps=args.stack_nb_steps,
1186 nb_stacks=args.stack_nb_stacks,
1187 nb_digits=args.stack_nb_digits,
1188 fraction_values_for_train=args.stack_fraction_values_for_train,
1192 elif args.task == "expr":
1194 nb_train_samples=args.nb_train_samples,
1195 nb_test_samples=args.nb_test_samples,
1196 batch_size=args.batch_size,
1201 raise ValueError(f"Unknown task {args.task}")
1203 ######################################################################
1205 log_string(f"device {device}")
1207 vocabulary_size = task.vocabulary_size()
1209 log_string(f"vocabulary_size {vocabulary_size}")
1211 ##############################
1213 model = mygpt.MyGPT(
1214 vocabulary_size=vocabulary_size,
1215 dim_model=args.dim_model,
1216 dim_keys=args.dim_keys,
1217 dim_hidden=args.dim_hidden,
1218 nb_heads=args.nb_heads,
1219 nb_blocks=args.nb_blocks,
1221 dropout=args.dropout,
1226 nb_parameters = sum(p.numel() for p in model.parameters())
1227 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
1229 ######################################################################
1231 nb_epochs_finished = 0
1233 if args.no_checkpoint:
1234 log_string(f"not trying to load checkpoint.")
1238 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1239 checkpoint = torch.load(checkpoint_name)
1240 nb_epochs_finished = checkpoint["nb_epochs_finished"]
1241 model.load_state_dict(checkpoint["model_state"])
1242 torch.set_rng_state(checkpoint["rng_state"])
1243 if torch.cuda.is_available():
1244 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
1246 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
1248 except FileNotFoundError:
1249 log_string("starting from scratch.")
1252 log_string("error when loading the checkpoint.")
1255 ######################################################################
1257 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
1260 for input in task.batches(split="train"):
1261 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
1262 token_probas = token_count / token_count.sum()
1263 entropy = -torch.xlogy(token_probas, token_probas).sum()
1264 train_set_perplexity = math.exp(entropy)
1266 ##############################
1268 if args.learning_rate_schedule == "cos":
1269 learning_rate_schedule = {}
1270 for n_epoch in range(args.nb_epochs):
1271 u = n_epoch / args.nb_epochs * math.pi
1272 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
1277 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
1281 learning_rate_schedule = {}
1282 learning_rate = args.learning_rate
1283 for n_epoch in range(args.nb_epochs):
1285 learning_rate = u[n_epoch]
1286 learning_rate_schedule[n_epoch] = learning_rate
1288 log_string(f"learning_rate_schedule {learning_rate_schedule}")
1290 ##############################
1294 if nb_epochs_finished >= nb_epochs:
1295 task.produce_results(nb_epochs_finished, model)
1297 for n_epoch in range(nb_epochs_finished, nb_epochs):
1298 learning_rate = learning_rate_schedule[n_epoch]
1300 log_string(f"learning_rate {learning_rate}")
1302 if args.optim == "sgd":
1303 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
1304 elif args.optim == "adam":
1305 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
1306 elif args.optim == "adamw":
1307 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
1309 raise ValueError(f"Unknown optimizer {args.optim}.")
1313 nb_train_samples, acc_train_loss = 0, 0.0
1315 for input in task.batches(split="train"):
1316 input = input.to(device)
1317 output = model(mygpt.BracketedSequence(input)).x
1318 loss = F.cross_entropy(output.transpose(1, 2), input)
1319 acc_train_loss += loss.item() * input.size(0)
1320 nb_train_samples += input.size(0)
1321 nb_samples_seen += input.size(0)
1323 optimizer.zero_grad()
1327 with torch.autograd.no_grad():
1330 nb_test_samples, acc_test_loss = 0, 0.0
1332 for input in task.batches(split="test"):
1333 input = input.to(device)
1335 output = model(mygpt.BracketedSequence(input)).x
1336 loss = F.cross_entropy(output.transpose(1, 2), input)
1337 acc_test_loss += loss.item() * input.size(0)
1338 nb_test_samples += input.size(0)
1340 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
1341 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
1344 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1347 task.produce_results(n_epoch, model)
1350 "nb_epochs_finished": n_epoch + 1,
1351 "model_state": model.state_dict(),
1352 "rng_state": torch.get_rng_state(),
1355 if torch.cuda.is_available():
1356 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1358 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1359 torch.save(checkpoint, checkpoint_name)
1360 log_string(f"saved checkpoint {checkpoint_name}")
1362 ######################################################################