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:
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,
164 if args.task in default_args:
165 for k, v in default_args[args.task].items():
166 if getattr(args, k) is None:
169 ######################################################################
172 os.mkdir(args.result_dir)
173 except FileExistsError:
174 if not args.overwrite_results:
175 print(f"result directory {args.result_dir} already exists")
178 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
181 # torch.backends.cudnn.deterministic = True
182 # torch.backends.cudnn.benchmark = False
183 # torch.use_deterministic_algorithms(True)
184 torch.manual_seed(args.seed)
185 if torch.cuda.is_available():
186 torch.cuda.manual_seed_all(args.seed)
188 ######################################################################
192 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
194 if log_file is not None:
195 log_file.write(t + s + "\n")
203 log_string(f"args.{n} {getattr(args, n)}")
205 ######################################################################
208 # ra_mask is boolean, with 1s on the values to generate
211 def masked_inplace_autoregression(
216 forbidden_tokens=None,
217 progress_bar_desc="autoregression",
218 device=torch.device("cpu"),
220 # p = logits.softmax(1)
221 # entropy[:,s]= p.xlogy(p).sum(1) / math.log(2)
222 batches = zip(input.split(batch_size), ar_mask.split(batch_size))
223 if progress_bar_desc is not None:
227 desc=progress_bar_desc,
228 total=input.size(0) // batch_size,
230 for input, ar_mask in batches:
231 i = (ar_mask.sum(0) > 0).nonzero()
234 mygpt.BracketedSequence(input, 0, i.min())
235 ) # Needed to initialize the model's cache
236 for s in range(i.min(), i.max() + 1):
237 output = model(mygpt.BracketedSequence(input, s, 1)).x
238 logits = output[:, s]
239 if forbidden_tokens is not None:
240 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
241 if args.deterministic_synthesis:
242 t_next = logits.argmax(1)
244 dist = torch.distributions.categorical.Categorical(logits=logits)
245 t_next = dist.sample()
246 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
249 ######################################################################
253 def batches(self, split="train"):
256 def vocabulary_size(self):
259 def produce_results(self, n_epoch, model):
263 ######################################################################
268 class TaskPicoCLVR(Task):
269 # Make a tensor from a list of strings
270 def tensorize(self, descr):
271 token_descr = [s.strip().split(" ") for s in descr]
272 l = max([len(s) for s in token_descr])
273 token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
274 id_descr = [[self.token2id[u] for u in s] for s in token_descr]
275 return torch.tensor(id_descr, device=self.device)
277 # Make a list of strings from a tensor
278 def detensorize(self, x):
279 return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
281 # trim all the tensors in the tuple z to remove as much token from
282 # left and right in the first tensor. If z is a tuple, all its
283 # elements are trimed according to the triming for the first
284 def trim(self, z, token="<nul>"):
285 n = self.token2id[token]
288 i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
289 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
290 return tuple([t[:, a:b] for t in z])
292 i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
293 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
296 ######################
297 # Not the cleanest part of the code
299 # Extract the last image of each sequence, from the last <img>
300 # included, and set to <nul> all the tokens from the beginning of
301 # that image to the end
302 def excise_last_image(self, input):
303 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
304 nb_img_tokens = self.height * self.width + 1
306 input = input.clone()
307 t = (input == t_img).long()
308 tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
309 i = (t * tail_masks).nonzero(as_tuple=True)
312 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
314 images = self.trim(input[j])
316 loss_masks = 1 - tail_masks
317 input, loss_masks = self.trim((input, loss_masks))
318 return input, loss_masks, images
320 def add_true_image(self, input, images, loss_masks):
321 t_nul = self.token2id["<nul>"]
322 nb_img_tokens = self.height * self.width + 1
323 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
324 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
325 t = (input == t_nul).long()
326 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
329 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
333 input, loss_masks = self.trim((input, loss_masks))
334 return input, loss_masks
336 def add_generated_image(self, input, loss_masks, model):
337 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
338 nb_img_tokens = self.height * self.width + 1
340 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
341 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
342 t = (input == t_nul).long()
343 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
350 + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
352 ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
355 torch.arange(self.vocabulary_size(), device=input.device) == t_nul
357 with torch.autograd.no_grad():
360 masked_inplace_autoregression(
366 progress_bar_desc=None,
371 input, loss_masks = self.trim((input, loss_masks))
373 return input, loss_masks
375 ######################
385 device=torch.device("cpu"),
389 def generate_descr(nb, cache_suffix, pruner):
390 return picoclvr.generate(
400 self.batch_size = batch_size
402 self.pruner_train = pruner_train
403 self.pruner_eval = pruner_eval
406 "nb_train_samples": nb_train_samples,
407 "nb_test_samples": nb_test_samples,
410 "nb_colors": nb_colors,
411 "batch_size": batch_size,
412 "rng_state": list(torch.get_rng_state()),
416 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
418 self.train_descr = generate_descr(
419 nb_train_samples, "train", pruner=self.pruner_train
421 self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
423 # Build the tokenizer
424 tokens = {"<nul>", "<img>"}
425 for d in [self.train_descr, self.test_descr]:
427 for t in s.strip().split(" "):
429 # make this set a sorted list to get the same tensors given
431 tokens = list(tokens)
433 self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
434 self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
436 # Tokenize the train and test sets
437 self.train_input = self.tensorize(self.train_descr)
438 self.test_input = self.tensorize(self.test_descr)
440 def batches(self, split="train"):
441 assert split in {"train", "test"}
442 input = self.train_input if split == "train" else self.test_input
443 for batch in tqdm.tqdm(
444 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
446 yield self.trim(batch)
448 def vocabulary_size(self):
449 return len(self.token2id)
451 def compute_missing_properties(self, n_epoch, model, pruner=None):
452 acc_nb_requested_properties = []
453 acc_nb_missing_properties = []
456 for input in tqdm.tqdm(
457 self.test_input.split(self.batch_size),
459 desc=f"test-properties",
461 tape, loss_masks, _ = self.excise_last_image(input)
462 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
463 result_descr = self.detensorize(tape)
464 np = picoclvr.nb_properties(
470 nb_requested_properties, _, nb_missing_properties = zip(*np)
471 acc_nb_requested_properties += nb_requested_properties
472 acc_nb_missing_properties += nb_missing_properties
473 acc_nb_results += len(result_descr)
475 nb_requested_properties = sum(acc_nb_requested_properties)
476 nb_missing_properties = sum(acc_nb_missing_properties)
478 prefix = "" if pruner is None else "pruned_"
479 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
481 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
484 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
487 ######################################################################
489 def produce_results(self, n_epoch, model):
490 self.compute_missing_properties(n_epoch, model)
492 if self.pruner_eval is not None:
493 self.compute_missing_properties(n_epoch, model, self.pruner_eval)
495 nb_tokens_to_generate = self.height * self.width + 3
500 for primer_descr in [
501 "red above green <sep> green top <sep> blue right of red",
502 "there is red <sep> there is yellow <sep> there is blue",
503 "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
504 "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
506 primer += [primer_descr] * nb_per_primer
508 tape = self.tensorize(primer)
509 loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
510 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
511 result_descr = self.detensorize(tape)
513 np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
515 acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
516 acc_nb_results = len(result_descr)
518 nb_requested_properties = sum(acc_nb_requested_properties)
519 nb_missing_properties = sum(acc_nb_missing_properties)
522 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
524 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
527 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
530 img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
534 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
538 torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
544 image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
545 torchvision.utils.save_image(
546 img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
548 log_string(f"wrote {image_name}")
551 ######################################################################
554 class TaskMNIST(Task):
555 def __init__(self, batch_size, device=torch.device("cpu")):
557 self.batch_size = batch_size
559 def batches(self, split="train"):
560 assert split in {"train", "test"}
561 data_set = torchvision.datasets.MNIST(
562 root="./data", train=(split == "train"), download=True
564 data_input = data_set.data.view(-1, 28 * 28).long()
565 if args.nb_train_samples is not None:
566 data_input = data_input[: args.nb_train_samples]
567 for batch in tqdm.tqdm(
568 data_input.split(self.batch_size), desc=f"epoch-{split}"
572 def vocabulary_size(self):
575 def produce_results(self, n_epoch, model):
576 results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
577 ar_mask = torch.full_like(results, 1)
578 masked_inplace_autoregression(
579 model, self.batch_size, results, ar_mask, device=self.device
581 image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
582 torchvision.utils.save_image(
583 1 - results.reshape(-1, 1, 28, 28) / 255.0,
588 log_string(f"wrote {image_name}")
591 ######################################################################
596 class TaskMaze(Task):
597 def map2seq(self, *m):
598 return torch.cat([x.flatten(1) for x in m], 1)
600 def seq2map(self, s):
601 s = s.reshape(s.size(0), -1, self.height, self.width)
602 return (s[:, k] for k in range(s.size(1)))
612 device=torch.device("cpu"),
614 self.batch_size = batch_size
619 train_mazes, train_paths, _ = maze.create_maze_data(
624 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
626 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
628 test_mazes, test_paths, _ = maze.create_maze_data(
633 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
635 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
637 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
639 def batches(self, split="train", nb_to_use=-1, desc=None):
640 assert split in {"train", "test"}
641 input = self.train_input if split == "train" else self.test_input
643 input = input[:nb_to_use]
645 desc = f"epoch-{split}"
646 for batch in tqdm.tqdm(
647 input.split(self.batch_size), dynamic_ncols=True, desc=desc
651 def vocabulary_size(self):
654 def compute_error(self, model, split="train", nb_to_use=-1):
655 nb_total, nb_correct = 0, 0
657 self.width * self.height,
658 self.width * self.height,
662 for input in tqdm.tqdm(
663 task.batches(split, nb_to_use),
667 result = input.clone()
668 ar_mask = result.new_zeros(result.size())
669 ar_mask[:, self.height * self.width :] = 1
670 result *= 1 - ar_mask
671 masked_inplace_autoregression(
676 progress_bar_desc=None,
679 mazes, paths = self.seq2map(result)
680 path_correctness = maze.path_correctness(mazes, paths)
681 nb_correct += path_correctness.long().sum()
682 nb_total += mazes.size(0)
684 optimal_path_lengths = (
685 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
687 predicted_path_lengths = (
688 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
690 optimal_path_lengths = optimal_path_lengths[path_correctness]
691 predicted_path_lengths = predicted_path_lengths[path_correctness]
692 count[optimal_path_lengths, predicted_path_lengths] += 1
698 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
701 return nb_total, nb_correct, count
703 def produce_results(self, n_epoch, model):
704 with torch.autograd.no_grad():
708 train_nb_total, train_nb_correct, count = self.compute_error(
709 model, "train", nb_to_use=1000
712 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}%"
715 test_nb_total, test_nb_correct, count = self.compute_error(
716 model, "test", nb_to_use=1000
719 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}%"
722 if count is not None:
723 proportion_optimal = count.diagonal().sum().float() / count.sum()
724 log_string(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
726 os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
728 for i in range(count.size(0)):
729 for j in range(count.size(1)):
730 eol = " " if j < count.size(1) - 1 else "\n"
731 f.write(f"{count[i,j]}{eol}")
733 input = self.test_input[:48]
734 result = input.clone()
735 ar_mask = result.new_zeros(result.size())
736 ar_mask[:, self.height * self.width :] = 1
737 result *= 1 - ar_mask
738 masked_inplace_autoregression(
739 model, self.batch_size, result, ar_mask, device=self.device
742 mazes, paths = self.seq2map(input)
743 _, predicted_paths = self.seq2map(result)
745 filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
750 predicted_paths=predicted_paths,
751 path_correct=maze.path_correctness(mazes, predicted_paths),
752 path_optimal=maze.path_optimality(paths, predicted_paths),
754 log_string(f"wrote {filename}")
759 ######################################################################
765 class TaskSnake(Task):
776 device=torch.device("cpu"),
778 self.batch_size = batch_size
782 self.prompt_length = prompt_length
784 self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
793 self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
803 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
805 def batches(self, split="train", nb_to_use=-1, desc=None):
806 assert split in {"train", "test"}
807 input = self.train_input if split == "train" else self.test_input
809 input = input[:nb_to_use]
811 desc = f"epoch-{split}"
812 for batch in tqdm.tqdm(
813 input.split(self.batch_size), dynamic_ncols=True, desc=desc
817 def vocabulary_size(self):
820 def produce_results(self, n_epoch, model):
821 with torch.autograd.no_grad():
825 def compute_nb_correct(input, prior_visits):
826 result = input.clone()
827 i = torch.arange(result.size(1), device=result.device)[None, :]
829 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
833 result *= 1 - ar_mask
835 # snake.solver(result,ar_mask)
837 masked_inplace_autoregression(
838 model, self.batch_size, result, ar_mask, device=self.device
841 nb_total = ((prior_visits > 0) * ar_mask).sum()
844 (result == input).long() * (prior_visits > 0) * ar_mask
847 # nb_total = result.size(0)
848 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
850 return nb_total, nb_correct
852 # train_nb_total, train_nb_correct = compute_nb_correct(
853 # self.train_input, self.train_prior_visits
857 # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
860 test_nb_total, test_nb_correct = compute_nb_correct(
861 self.test_input[:1000], self.test_prior_visits[:1000]
865 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}%"
871 ######################################################################
877 class TaskStack(Task):
886 fraction_values_for_train=None,
887 device=torch.device("cpu"),
889 self.batch_size = batch_size
890 self.nb_steps = nb_steps
891 self.nb_stacks = nb_stacks
892 self.nb_digits = nb_digits
895 if fraction_values_for_train is None:
896 values_for_train = None
897 values_for_test = None
899 all = torch.randperm(10**nb_digits)
900 nb_for_train = int(all.size(0) * fraction_values_for_train)
901 values_for_train = all[:nb_for_train]
902 values_for_test = all[nb_for_train:]
904 self.train_input, self.train_stack_counts = stack.generate_sequences(
913 self.test_input, self.test_stack_counts = stack.generate_sequences(
922 i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
923 counts = self.test_stack_counts.flatten()[i.flatten()]
924 counts = F.one_hot(counts).sum(0)
925 log_string(f"pop_stack_counts {counts}")
927 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
929 def batches(self, split="train", nb_to_use=-1, desc=None):
930 assert split in {"train", "test"}
931 input = self.train_input if split == "train" else self.test_input
933 input = input[:nb_to_use]
935 desc = f"epoch-{split}"
936 for batch in tqdm.tqdm(
937 input.split(self.batch_size), dynamic_ncols=True, desc=desc
941 def vocabulary_size(self):
944 def produce_results(self, n_epoch, model):
945 with torch.autograd.no_grad():
949 def compute_nb_correct(input):
950 result = input.clone()
951 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
952 ar_mask = (result != input).long()
953 masked_inplace_autoregression(
954 model, self.batch_size, result, ar_mask, device=self.device
957 errors = ((result != input).long() * ar_mask).reshape(
958 -1, 1 + self.nb_digits
960 ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
962 nb_total = ar_mask.max(1).values.sum()
963 nb_correct = nb_total - errors.max(1).values.sum()
965 return nb_total, nb_correct
967 test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
970 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}%"
973 ##############################################################
974 # Log a few generated sequences
975 input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
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 ######################################################################