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"
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: args.result_dir=f"results_{args.task}"
129 os.mkdir(args.result_dir)
130 except FileExistsError:
131 if not args.overwrite_results:
132 print(f"result directory {args.result_dir} already exists")
135 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
138 # torch.backends.cudnn.deterministic = True
139 # torch.backends.cudnn.benchmark = False
140 # torch.use_deterministic_algorithms(True)
141 torch.manual_seed(args.seed)
142 if torch.cuda.is_available():
143 torch.cuda.manual_seed_all(args.seed)
145 ######################################################################
151 "nb_train_samples": 250000,
152 "nb_test_samples": 10000,
157 "nb_train_samples": 250000,
158 "nb_test_samples": 10000,
163 "nb_train_samples": 250000,
164 "nb_test_samples": 10000,
169 "nb_train_samples": 250000,
170 "nb_test_samples": 10000,
175 "nb_train_samples": 100000,
176 "nb_test_samples": 1000,
180 if args.task in default_args:
181 for k, v in default_args[args.task].items():
182 if getattr(args, k) is None:
185 ######################################################################
189 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
191 if log_file is not None:
192 log_file.write(t + s + "\n")
200 log_string(f"args.{n} {getattr(args, n)}")
202 ######################################################################
205 # ra_mask is boolean, with 1s on the values to generate
208 def masked_inplace_autoregression(
213 forbidden_tokens=None,
214 progress_bar_desc="autoregression",
215 device=torch.device("cpu"),
217 # p = logits.softmax(1)
218 # entropy[:,s]= p.xlogy(p).sum(1) / math.log(2)
219 batches = zip(input.split(batch_size), ar_mask.split(batch_size))
220 if progress_bar_desc is not None:
224 desc=progress_bar_desc,
225 total=input.size(0) // batch_size,
227 for input, ar_mask in batches:
228 i = (ar_mask.sum(0) > 0).nonzero()
231 mygpt.BracketedSequence(input, 0, i.min())
232 ) # Needed to initialize the model's cache
233 for s in range(i.min(), i.max() + 1):
234 output = model(mygpt.BracketedSequence(input, s, 1)).x
235 logits = output[:, s]
236 if forbidden_tokens is not None:
237 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
238 if args.deterministic_synthesis:
239 t_next = logits.argmax(1)
241 dist = torch.distributions.categorical.Categorical(logits=logits)
242 t_next = dist.sample()
243 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
246 ######################################################################
250 def batches(self, split="train"):
253 def vocabulary_size(self):
256 def produce_results(self, n_epoch, model):
260 ######################################################################
265 class TaskPicoCLVR(Task):
266 # Make a tensor from a list of strings
267 def tensorize(self, descr):
268 token_descr = [s.strip().split(" ") for s in descr]
269 l = max([len(s) for s in token_descr])
270 token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
271 id_descr = [[self.token2id[u] for u in s] for s in token_descr]
272 return torch.tensor(id_descr, device=self.device)
274 # Make a list of strings from a tensor
275 def detensorize(self, x):
276 return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
278 # trim all the tensors in the tuple z to remove as much token from
279 # left and right in the first tensor. If z is a tuple, all its
280 # elements are trimed according to the triming for the first
281 def trim(self, z, token="<nul>"):
282 n = self.token2id[token]
285 i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
286 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
287 return tuple([t[:, a:b] for t in z])
289 i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
290 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
293 ######################
294 # Not the cleanest part of the code
296 # Extract the last image of each sequence, from the last <img>
297 # included, and set to <nul> all the tokens from the beginning of
298 # that image to the end
299 def excise_last_image(self, input):
300 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
301 nb_img_tokens = self.height * self.width + 1
303 input = input.clone()
304 t = (input == t_img).long()
305 tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
306 i = (t * tail_masks).nonzero(as_tuple=True)
309 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
311 images = self.trim(input[j])
313 loss_masks = 1 - tail_masks
314 input, loss_masks = self.trim((input, loss_masks))
315 return input, loss_masks, images
317 def add_true_image(self, input, images, loss_masks):
318 t_nul = self.token2id["<nul>"]
319 nb_img_tokens = self.height * self.width + 1
320 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
321 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
322 t = (input == t_nul).long()
323 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
326 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
330 input, loss_masks = self.trim((input, loss_masks))
331 return input, loss_masks
333 def add_generated_image(self, input, loss_masks, model):
334 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
335 nb_img_tokens = self.height * self.width + 1
337 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
338 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
339 t = (input == t_nul).long()
340 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
347 + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
349 ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
352 torch.arange(self.vocabulary_size(), device=input.device) == t_nul
354 with torch.autograd.no_grad():
357 masked_inplace_autoregression(
363 progress_bar_desc=None,
368 input, loss_masks = self.trim((input, loss_masks))
370 return input, loss_masks
372 ######################
382 device=torch.device("cpu"),
386 def generate_descr(nb, cache_suffix, pruner):
387 return picoclvr.generate(
397 self.batch_size = batch_size
399 self.pruner_train = pruner_train
400 self.pruner_eval = pruner_eval
403 "nb_train_samples": nb_train_samples,
404 "nb_test_samples": nb_test_samples,
407 "nb_colors": nb_colors,
408 "batch_size": batch_size,
409 "rng_state": list(torch.get_rng_state()),
413 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
415 self.train_descr = generate_descr(
416 nb_train_samples, "train", pruner=self.pruner_train
418 self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
420 # Build the tokenizer
421 tokens = {"<nul>", "<img>"}
422 for d in [self.train_descr, self.test_descr]:
424 for t in s.strip().split(" "):
426 # make this set a sorted list to get the same tensors given
428 tokens = list(tokens)
430 self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
431 self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
433 # Tokenize the train and test sets
434 self.train_input = self.tensorize(self.train_descr)
435 self.test_input = self.tensorize(self.test_descr)
437 def batches(self, split="train"):
438 assert split in {"train", "test"}
439 input = self.train_input if split == "train" else self.test_input
440 for batch in tqdm.tqdm(
441 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
443 yield self.trim(batch)
445 def vocabulary_size(self):
446 return len(self.token2id)
448 def compute_missing_properties(self, n_epoch, model, pruner=None):
449 acc_nb_requested_properties = []
450 acc_nb_missing_properties = []
453 for input in tqdm.tqdm(
454 self.test_input.split(self.batch_size),
456 desc=f"test-properties",
458 tape, loss_masks, _ = self.excise_last_image(input)
459 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
460 result_descr = self.detensorize(tape)
461 np = picoclvr.nb_properties(
467 nb_requested_properties, _, nb_missing_properties = zip(*np)
468 acc_nb_requested_properties += nb_requested_properties
469 acc_nb_missing_properties += nb_missing_properties
470 acc_nb_results += len(result_descr)
472 nb_requested_properties = sum(acc_nb_requested_properties)
473 nb_missing_properties = sum(acc_nb_missing_properties)
475 prefix = "" if pruner is None else "pruned_"
476 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
478 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
481 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
484 ######################################################################
486 def produce_results(self, n_epoch, model):
487 self.compute_missing_properties(n_epoch, model)
489 if self.pruner_eval is not None:
490 self.compute_missing_properties(n_epoch, model, self.pruner_eval)
492 nb_tokens_to_generate = self.height * self.width + 3
497 for primer_descr in [
498 "red above green <sep> green top <sep> blue right of red",
499 "there is red <sep> there is yellow <sep> there is blue",
500 "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
501 "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
503 primer += [primer_descr] * nb_per_primer
505 tape = self.tensorize(primer)
506 loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
507 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
508 result_descr = self.detensorize(tape)
510 np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
512 acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
513 acc_nb_results = len(result_descr)
515 nb_requested_properties = sum(acc_nb_requested_properties)
516 nb_missing_properties = sum(acc_nb_missing_properties)
519 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
521 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
524 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
527 img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
531 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
535 torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
541 image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
542 torchvision.utils.save_image(
543 img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
545 log_string(f"wrote {image_name}")
548 ######################################################################
551 class TaskMNIST(Task):
552 def __init__(self, batch_size, device=torch.device("cpu")):
554 self.batch_size = batch_size
556 def batches(self, split="train"):
557 assert split in {"train", "test"}
558 data_set = torchvision.datasets.MNIST(
559 root="./data", train=(split == "train"), download=True
561 data_input = data_set.data.view(-1, 28 * 28).long()
562 if args.nb_train_samples is not None:
563 data_input = data_input[: args.nb_train_samples]
564 for batch in tqdm.tqdm(
565 data_input.split(self.batch_size), desc=f"epoch-{split}"
569 def vocabulary_size(self):
572 def produce_results(self, n_epoch, model):
573 results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
574 ar_mask = torch.full_like(results, 1)
575 masked_inplace_autoregression(
576 model, self.batch_size, results, ar_mask, device=self.device
578 image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
579 torchvision.utils.save_image(
580 1 - results.reshape(-1, 1, 28, 28) / 255.0,
585 log_string(f"wrote {image_name}")
588 ######################################################################
593 class TaskMaze(Task):
594 def map2seq(self, *m):
595 return torch.cat([x.flatten(1) for x in m], 1)
597 def seq2map(self, s):
598 s = s.reshape(s.size(0), -1, self.height, self.width)
599 return (s[:, k] for k in range(s.size(1)))
609 device=torch.device("cpu"),
611 self.batch_size = batch_size
616 train_mazes, train_paths, _ = maze.create_maze_data(
621 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
623 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
625 test_mazes, test_paths, _ = maze.create_maze_data(
630 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
632 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
634 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
636 def batches(self, split="train", nb_to_use=-1, desc=None):
637 assert split in {"train", "test"}
638 input = self.train_input if split == "train" else self.test_input
640 input = input[:nb_to_use]
642 desc = f"epoch-{split}"
643 for batch in tqdm.tqdm(
644 input.split(self.batch_size), dynamic_ncols=True, desc=desc
648 def vocabulary_size(self):
651 def compute_error(self, model, split="train", nb_to_use=-1):
652 nb_total, nb_correct = 0, 0
654 self.width * self.height,
655 self.width * self.height,
659 for input in tqdm.tqdm(
660 task.batches(split, nb_to_use),
664 result = input.clone()
665 ar_mask = result.new_zeros(result.size())
666 ar_mask[:, self.height * self.width :] = 1
667 result *= 1 - ar_mask
668 masked_inplace_autoregression(
673 progress_bar_desc=None,
676 mazes, paths = self.seq2map(result)
677 path_correctness = maze.path_correctness(mazes, paths)
678 nb_correct += path_correctness.long().sum()
679 nb_total += mazes.size(0)
681 optimal_path_lengths = (
682 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
684 predicted_path_lengths = (
685 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
687 optimal_path_lengths = optimal_path_lengths[path_correctness]
688 predicted_path_lengths = predicted_path_lengths[path_correctness]
689 count[optimal_path_lengths, predicted_path_lengths] += 1
695 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
698 return nb_total, nb_correct, count
700 def produce_results(self, n_epoch, model):
701 with torch.autograd.no_grad():
705 train_nb_total, train_nb_correct, count = self.compute_error(
706 model, "train", nb_to_use=1000
709 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
712 test_nb_total, test_nb_correct, count = self.compute_error(
713 model, "test", nb_to_use=1000
716 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
719 if count is not None:
720 proportion_optimal = count.diagonal().sum().float() / count.sum()
721 log_string(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
723 os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
725 for i in range(count.size(0)):
726 for j in range(count.size(1)):
727 eol = " " if j < count.size(1) - 1 else "\n"
728 f.write(f"{count[i,j]}{eol}")
730 input = self.test_input[:48]
731 result = input.clone()
732 ar_mask = result.new_zeros(result.size())
733 ar_mask[:, self.height * self.width :] = 1
734 result *= 1 - ar_mask
735 masked_inplace_autoregression(
736 model, self.batch_size, result, ar_mask, device=self.device
739 mazes, paths = self.seq2map(input)
740 _, predicted_paths = self.seq2map(result)
742 filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
747 predicted_paths=predicted_paths,
748 path_correct=maze.path_correctness(mazes, predicted_paths),
749 path_optimal=maze.path_optimality(paths, predicted_paths),
751 log_string(f"wrote {filename}")
756 ######################################################################
762 class TaskSnake(Task):
773 device=torch.device("cpu"),
775 self.batch_size = batch_size
779 self.prompt_length = prompt_length
781 self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
790 self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
800 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
802 def batches(self, split="train", nb_to_use=-1, desc=None):
803 assert split in {"train", "test"}
804 input = self.train_input if split == "train" else self.test_input
806 input = input[:nb_to_use]
808 desc = f"epoch-{split}"
809 for batch in tqdm.tqdm(
810 input.split(self.batch_size), dynamic_ncols=True, desc=desc
814 def vocabulary_size(self):
817 def produce_results(self, n_epoch, model):
818 with torch.autograd.no_grad():
822 def compute_nb_correct(input, prior_visits):
823 result = input.clone()
824 i = torch.arange(result.size(1), device=result.device)[None, :]
826 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
830 result *= 1 - ar_mask
832 # snake.solver(result,ar_mask)
834 masked_inplace_autoregression(
835 model, self.batch_size, result, ar_mask, device=self.device
838 nb_total = ((prior_visits > 0) * ar_mask).sum()
841 (result == input).long() * (prior_visits > 0) * ar_mask
844 # nb_total = result.size(0)
845 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
847 return nb_total, nb_correct
849 # train_nb_total, train_nb_correct = compute_nb_correct(
850 # self.train_input, self.train_prior_visits
854 # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
857 test_nb_total, test_nb_correct = compute_nb_correct(
858 self.test_input[:1000], self.test_prior_visits[:1000]
862 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
868 ######################################################################
874 class TaskStack(Task):
883 fraction_values_for_train=None,
884 device=torch.device("cpu"),
886 self.batch_size = batch_size
887 self.nb_steps = nb_steps
888 self.nb_stacks = nb_stacks
889 self.nb_digits = nb_digits
892 if fraction_values_for_train is None:
893 values_for_train = None
894 values_for_test = None
896 all = torch.randperm(10**nb_digits)
897 nb_for_train = int(all.size(0) * fraction_values_for_train)
898 values_for_train = all[:nb_for_train]
899 values_for_test = all[nb_for_train:]
901 self.train_input, self.train_stack_counts = stack.generate_sequences(
910 self.test_input, self.test_stack_counts = stack.generate_sequences(
919 mask = self.test_input.clone()
920 stack.remove_popped_values(mask, self.nb_stacks, self.nb_digits)
921 mask = mask != self.test_input
922 counts = self.test_stack_counts.flatten()[mask.flatten()]
923 counts = F.one_hot(counts).sum(0)
924 log_string(f"stack_count {counts}")
926 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
928 def batches(self, split="train", nb_to_use=-1, desc=None):
929 assert split in {"train", "test"}
930 input = self.train_input if split == "train" else self.test_input
932 input = input[:nb_to_use]
934 desc = f"epoch-{split}"
935 for batch in tqdm.tqdm(
936 input.split(self.batch_size), dynamic_ncols=True, desc=desc
940 def vocabulary_size(self):
943 def produce_results(self, n_epoch, model):
944 with torch.autograd.no_grad():
948 def compute_nb_correct(input):
949 result = input.clone()
950 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
951 ar_mask = (result != input).long()
952 masked_inplace_autoregression(
953 model, self.batch_size, result, ar_mask, device=self.device
956 errors = ((result != input).long() * ar_mask).reshape(
957 -1, 1 + self.nb_digits
959 ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
961 nb_total = ar_mask.max(1).values.sum()
962 nb_correct = nb_total - errors.max(1).values.sum()
964 return nb_total, nb_correct
966 test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
969 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
972 #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
974 l = l - l % (1 + self.nb_digits)
975 input = self.test_input[:10, :l]
976 result = input.clone()
977 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
978 ar_mask = (result != input).long()
979 for n in range(result.size(0)):
981 f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
983 masked_inplace_autoregression(
984 model, self.batch_size, result, ar_mask, device=self.device
986 for n in range(result.size(0)):
988 f"test_after {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
990 #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
995 ######################################################################
998 def picoclvr_pruner_horizontal_green(p):
999 return not ("green" in p and ("left" in p or "right" in p))
1002 picoclvr_pruner_train = (
1003 picoclvr_pruner_horizontal_green
1004 if args.picocvlr_prune_properties in {"train+eval"}
1008 picoclvr_pruner_eval = (
1009 (lambda p: not picoclvr_pruner_horizontal_green(p))
1010 if args.picocvlr_prune_properties in {"train+eval", "eval"}
1014 ######################################################################
1016 if args.task == "picoclvr":
1017 task = TaskPicoCLVR(
1018 nb_train_samples=args.nb_train_samples,
1019 nb_test_samples=args.nb_test_samples,
1020 batch_size=args.batch_size,
1021 height=args.picoclvr_height,
1022 width=args.picoclvr_width,
1023 nb_colors=args.picoclvr_nb_colors,
1025 pruner_train=picoclvr_pruner_train,
1026 pruner_eval=picoclvr_pruner_eval,
1029 elif args.task == "mnist":
1031 batch_size=args.batch_size,
1035 elif args.task == "maze":
1037 nb_train_samples=args.nb_train_samples,
1038 nb_test_samples=args.nb_test_samples,
1039 batch_size=args.batch_size,
1040 height=args.maze_height,
1041 width=args.maze_width,
1042 nb_walls=args.maze_nb_walls,
1046 elif args.task == "snake":
1048 nb_train_samples=args.nb_train_samples,
1049 nb_test_samples=args.nb_test_samples,
1050 batch_size=args.batch_size,
1051 height=args.snake_height,
1052 width=args.snake_width,
1053 nb_colors=args.snake_nb_colors,
1054 length=args.snake_length,
1055 prompt_length=args.snake_length // 2,
1059 elif args.task == "stack":
1061 nb_train_samples=args.nb_train_samples,
1062 nb_test_samples=args.nb_test_samples,
1063 batch_size=args.batch_size,
1064 nb_steps=args.stack_nb_steps,
1065 nb_stacks=args.stack_nb_stacks,
1066 nb_digits=args.stack_nb_digits,
1067 fraction_values_for_train=args.stack_fraction_values_for_train,
1072 raise ValueError(f"Unknown task {args.task}")
1074 ######################################################################
1076 log_string(f"device {device}")
1078 vocabulary_size = task.vocabulary_size()
1080 log_string(f"vocabulary_size {vocabulary_size}")
1082 ##############################
1084 model = mygpt.MyGPT(
1085 vocabulary_size=vocabulary_size,
1086 dim_model=args.dim_model,
1087 dim_keys=args.dim_keys,
1088 dim_hidden=args.dim_hidden,
1089 nb_heads=args.nb_heads,
1090 nb_blocks=args.nb_blocks,
1092 dropout=args.dropout,
1097 nb_parameters = sum(p.numel() for p in model.parameters())
1098 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
1100 ######################################################################
1102 nb_epochs_finished = 0
1104 if args.no_checkpoint:
1105 log_string(f"not trying to load checkpoint.")
1109 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1110 checkpoint = torch.load(checkpoint_name)
1111 nb_epochs_finished = checkpoint["nb_epochs_finished"]
1112 model.load_state_dict(checkpoint["model_state"])
1113 torch.set_rng_state(checkpoint["rng_state"])
1114 if torch.cuda.is_available():
1115 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
1117 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
1119 except FileNotFoundError:
1120 log_string("starting from scratch.")
1123 log_string("error when loading the checkpoint.")
1126 ######################################################################
1128 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
1131 for input in task.batches(split="train"):
1132 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
1133 token_probas = token_count / token_count.sum()
1134 entropy = -torch.xlogy(token_probas, token_probas).sum()
1135 train_set_perplexity = math.exp(entropy)
1137 ##############################
1139 if args.learning_rate_schedule == "cos":
1140 learning_rate_schedule = {}
1141 for n_epoch in range(args.nb_epochs):
1142 u = n_epoch / args.nb_epochs * math.pi
1143 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
1148 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
1152 learning_rate_schedule = {}
1153 learning_rate = args.learning_rate
1154 for n_epoch in range(args.nb_epochs):
1156 learning_rate = u[n_epoch]
1157 learning_rate_schedule[n_epoch] = learning_rate
1159 log_string(f"learning_rate_schedule {learning_rate_schedule}")
1161 ##############################
1165 if nb_epochs_finished >= nb_epochs:
1166 task.produce_results(nb_epochs_finished, model)
1168 for n_epoch in range(nb_epochs_finished, nb_epochs):
1169 learning_rate = learning_rate_schedule[n_epoch]
1171 log_string(f"learning_rate {learning_rate}")
1173 if args.optim == "sgd":
1174 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
1175 elif args.optim == "adam":
1176 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
1177 elif args.optim == "adamw":
1178 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
1180 raise ValueError(f"Unknown optimizer {args.optim}.")
1184 nb_train_samples, acc_train_loss = 0, 0.0
1186 for input in task.batches(split="train"):
1187 input = input.to(device)
1188 output = model(mygpt.BracketedSequence(input)).x
1189 loss = F.cross_entropy(output.transpose(1, 2), input)
1190 acc_train_loss += loss.item() * input.size(0)
1191 nb_train_samples += input.size(0)
1192 nb_samples_seen += input.size(0)
1194 optimizer.zero_grad()
1198 with torch.autograd.no_grad():
1201 nb_test_samples, acc_test_loss = 0, 0.0
1203 for input in task.batches(split="test"):
1204 input = input.to(device)
1206 output = model(mygpt.BracketedSequence(input)).x
1207 loss = F.cross_entropy(output.transpose(1, 2), input)
1208 acc_test_loss += loss.item() * input.size(0)
1209 nb_test_samples += input.size(0)
1211 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
1212 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
1215 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1218 task.produce_results(n_epoch, model)
1221 "nb_epochs_finished": n_epoch + 1,
1222 "model_state": model.state_dict(),
1223 "rng_state": torch.get_rng_state(),
1226 if torch.cuda.is_available():
1227 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1229 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1230 torch.save(checkpoint, checkpoint_name)
1231 log_string(f"saved checkpoint {checkpoint_name}")
1233 ######################################################################