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"
38 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
40 parser.add_argument("--result_dir", type=str, default="results_default")
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=250000)
50 parser.add_argument("--nb_test_samples", type=int, default=10000)
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 ######################################################################
111 args = parser.parse_args()
113 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
116 os.mkdir(args.result_dir)
117 except FileExistsError:
118 if not args.overwrite_results:
119 print(f"result directory {args.result_dir} already exists")
122 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
125 # torch.backends.cudnn.deterministic = True
126 # torch.backends.cudnn.benchmark = False
127 # torch.use_deterministic_algorithms(True)
128 torch.manual_seed(args.seed)
129 if torch.cuda.is_available():
130 torch.cuda.manual_seed_all(args.seed)
132 ######################################################################
153 if args.task in default_args:
154 for k, v in default_args[args.task].items():
155 if getattr(args, k) is None:
158 ######################################################################
162 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
164 if log_file is not None:
165 log_file.write(t + s + "\n")
173 log_string(f"args.{n} {getattr(args, n)}")
175 ######################################################################
178 # ra_mask is boolean, with 1s on the values to generate
181 def masked_inplace_autoregression(
186 forbidden_tokens=None,
187 progress_bar_desc="autoregression",
188 device=torch.device("cpu"),
190 # p = logits.softmax(1)
191 # entropy[:,s]= p.xlogy(p).sum(1) / math.log(2)
192 batches = zip(input.split(batch_size), ar_mask.split(batch_size))
193 if progress_bar_desc is not None:
197 desc=progress_bar_desc,
198 total=input.size(0) // batch_size,
200 for input, ar_mask in batches:
201 i = (ar_mask.sum(0) > 0).nonzero()
204 mygpt.BracketedSequence(input, 0, i.min())
205 ) # Needed to initialize the model's cache
206 for s in range(i.min(), i.max() + 1):
207 output = model(mygpt.BracketedSequence(input, s, 1)).x
208 logits = output[:, s]
209 if forbidden_tokens is not None:
210 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
211 if args.deterministic_synthesis:
212 t_next = logits.argmax(1)
214 dist = torch.distributions.categorical.Categorical(logits=logits)
215 t_next = dist.sample()
216 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
219 ######################################################################
223 def batches(self, split="train"):
226 def vocabulary_size(self):
229 def produce_results(self, n_epoch, model):
233 ######################################################################
238 class TaskPicoCLVR(Task):
239 # Make a tensor from a list of strings
240 def tensorize(self, descr):
241 token_descr = [s.strip().split(" ") for s in descr]
242 l = max([len(s) for s in token_descr])
243 token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
244 id_descr = [[self.token2id[u] for u in s] for s in token_descr]
245 return torch.tensor(id_descr, device=self.device)
247 # Make a list of strings from a tensor
248 def detensorize(self, x):
249 return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
251 # trim all the tensors in the tuple z to remove as much token from
252 # left and right in the first tensor. If z is a tuple, all its
253 # elements are trimed according to the triming for the first
254 def trim(self, z, token="<nul>"):
255 n = self.token2id[token]
258 i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
259 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
260 return tuple([t[:, a:b] for t in z])
262 i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
263 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
266 ######################
267 # Not the cleanest part of the code
269 # Extract the last image of each sequence, from the last <img>
270 # included, and set to <nul> all the tokens from the beginning of
271 # that image to the end
272 def excise_last_image(self, input):
273 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
274 nb_img_tokens = self.height * self.width + 1
276 input = input.clone()
277 t = (input == t_img).long()
278 tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
279 i = (t * tail_masks).nonzero(as_tuple=True)
282 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
284 images = self.trim(input[j])
286 loss_masks = 1 - tail_masks
287 input, loss_masks = self.trim((input, loss_masks))
288 return input, loss_masks, images
290 def add_true_image(self, input, images, loss_masks):
291 t_nul = self.token2id["<nul>"]
292 nb_img_tokens = self.height * self.width + 1
293 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
294 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
295 t = (input == t_nul).long()
296 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
299 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
303 input, loss_masks = self.trim((input, loss_masks))
304 return input, loss_masks
306 def add_generated_image(self, input, loss_masks, model):
307 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
308 nb_img_tokens = self.height * self.width + 1
310 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
311 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
312 t = (input == t_nul).long()
313 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
320 + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
322 ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
325 torch.arange(self.vocabulary_size(), device=input.device) == t_nul
327 with torch.autograd.no_grad():
330 masked_inplace_autoregression(
336 progress_bar_desc=None,
341 input, loss_masks = self.trim((input, loss_masks))
343 return input, loss_masks
345 ######################
355 device=torch.device("cpu"),
359 def generate_descr(nb, cache_suffix, pruner):
360 return picoclvr.generate(
370 self.batch_size = batch_size
372 self.pruner_train = pruner_train
373 self.pruner_eval = pruner_eval
376 "nb_train_samples": nb_train_samples,
377 "nb_test_samples": nb_test_samples,
380 "nb_colors": nb_colors,
381 "batch_size": batch_size,
382 "rng_state": list(torch.get_rng_state()),
386 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
388 self.train_descr = generate_descr(
389 nb_train_samples, "train", pruner=self.pruner_train
391 self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
393 # Build the tokenizer
394 tokens = {"<nul>", "<img>"}
395 for d in [self.train_descr, self.test_descr]:
397 for t in s.strip().split(" "):
399 # make this set a sorted list to get the same tensors given
401 tokens = list(tokens)
403 self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
404 self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
406 # Tokenize the train and test sets
407 self.train_input = self.tensorize(self.train_descr)
408 self.test_input = self.tensorize(self.test_descr)
410 def batches(self, split="train"):
411 assert split in {"train", "test"}
412 input = self.train_input if split == "train" else self.test_input
413 for batch in tqdm.tqdm(
414 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
416 yield self.trim(batch)
418 def vocabulary_size(self):
419 return len(self.token2id)
421 def compute_missing_properties(self, n_epoch, model, pruner=None):
422 acc_nb_requested_properties = []
423 acc_nb_missing_properties = []
426 for input in tqdm.tqdm(
427 self.test_input.split(self.batch_size),
429 desc=f"test-properties",
431 tape, loss_masks, _ = self.excise_last_image(input)
432 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
433 result_descr = self.detensorize(tape)
434 np = picoclvr.nb_properties(
440 nb_requested_properties, _, nb_missing_properties = zip(*np)
441 acc_nb_requested_properties += nb_requested_properties
442 acc_nb_missing_properties += nb_missing_properties
443 acc_nb_results += len(result_descr)
445 nb_requested_properties = sum(acc_nb_requested_properties)
446 nb_missing_properties = sum(acc_nb_missing_properties)
448 prefix = "" if pruner is None else "pruned_"
449 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
451 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
454 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
457 ######################################################################
459 def produce_results(self, n_epoch, model):
460 self.compute_missing_properties(n_epoch, model)
462 if self.pruner_eval is not None:
463 self.compute_missing_properties(n_epoch, model, self.pruner_eval)
465 nb_tokens_to_generate = self.height * self.width + 3
470 for primer_descr in [
471 "red above green <sep> green top <sep> blue right of red",
472 "there is red <sep> there is yellow <sep> there is blue",
473 "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
474 "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
476 primer += [primer_descr] * nb_per_primer
478 tape = self.tensorize(primer)
479 loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
480 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
481 result_descr = self.detensorize(tape)
483 np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
485 acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
486 acc_nb_results = len(result_descr)
488 nb_requested_properties = sum(acc_nb_requested_properties)
489 nb_missing_properties = sum(acc_nb_missing_properties)
492 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
494 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
497 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
500 img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
504 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
508 torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
514 image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
515 torchvision.utils.save_image(
516 img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
518 log_string(f"wrote {image_name}")
521 ######################################################################
524 class TaskMNIST(Task):
525 def __init__(self, batch_size, device=torch.device("cpu")):
527 self.batch_size = batch_size
529 def batches(self, split="train"):
530 assert split in {"train", "test"}
531 data_set = torchvision.datasets.MNIST(
532 root="./data", train=(split == "train"), download=True
534 data_input = data_set.data.view(-1, 28 * 28).long()
535 if args.nb_train_samples is not None:
536 data_input = data_input[: args.nb_train_samples]
537 for batch in tqdm.tqdm(
538 data_input.split(self.batch_size), desc=f"epoch-{split}"
542 def vocabulary_size(self):
545 def produce_results(self, n_epoch, model):
546 results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
547 ar_mask = torch.full_like(results, 1)
548 masked_inplace_autoregression(
549 model, self.batch_size, results, ar_mask, device=self.device
551 image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
552 torchvision.utils.save_image(
553 1 - results.reshape(-1, 1, 28, 28) / 255.0,
558 log_string(f"wrote {image_name}")
561 ######################################################################
566 class TaskMaze(Task):
567 def map2seq(self, *m):
568 return torch.cat([x.flatten(1) for x in m], 1)
570 def seq2map(self, s):
571 s = s.reshape(s.size(0), -1, self.height, self.width)
572 return (s[:, k] for k in range(s.size(1)))
582 device=torch.device("cpu"),
584 self.batch_size = batch_size
589 train_mazes, train_paths, _ = maze.create_maze_data(
594 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
596 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
598 test_mazes, test_paths, _ = maze.create_maze_data(
603 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
605 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
607 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
609 def batches(self, split="train", nb_to_use=-1, desc=None):
610 assert split in {"train", "test"}
611 input = self.train_input if split == "train" else self.test_input
613 input = input[:nb_to_use]
615 desc = f"epoch-{split}"
616 for batch in tqdm.tqdm(
617 input.split(self.batch_size), dynamic_ncols=True, desc=desc
621 def vocabulary_size(self):
624 def compute_error(self, model, split="train", nb_to_use=-1):
625 nb_total, nb_correct = 0, 0
627 self.width * self.height,
628 self.width * self.height,
632 for input in tqdm.tqdm(
633 task.batches(split, nb_to_use),
637 result = input.clone()
638 ar_mask = result.new_zeros(result.size())
639 ar_mask[:, self.height * self.width :] = 1
640 result *= 1 - ar_mask
641 masked_inplace_autoregression(
646 progress_bar_desc=None,
649 mazes, paths = self.seq2map(result)
650 path_correctness = maze.path_correctness(mazes, paths)
651 nb_correct += path_correctness.long().sum()
652 nb_total += mazes.size(0)
654 optimal_path_lengths = (
655 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
657 predicted_path_lengths = (
658 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
660 optimal_path_lengths = optimal_path_lengths[path_correctness]
661 predicted_path_lengths = predicted_path_lengths[path_correctness]
662 count[optimal_path_lengths, predicted_path_lengths] += 1
668 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
671 return nb_total, nb_correct, count
673 def produce_results(self, n_epoch, model):
674 with torch.autograd.no_grad():
678 train_nb_total, train_nb_correct, count = self.compute_error(
679 model, "train", nb_to_use=1000
682 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
685 test_nb_total, test_nb_correct, count = self.compute_error(
686 model, "test", nb_to_use=1000
689 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
692 if count is not None:
693 proportion_optimal = count.diagonal().sum().float() / count.sum()
694 log_string(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
696 os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
698 for i in range(count.size(0)):
699 for j in range(count.size(1)):
700 eol = " " if j < count.size(1) - 1 else "\n"
701 f.write(f"{count[i,j]}{eol}")
703 input = self.test_input[:48]
704 result = input.clone()
705 ar_mask = result.new_zeros(result.size())
706 ar_mask[:, self.height * self.width :] = 1
707 result *= 1 - ar_mask
708 masked_inplace_autoregression(
709 model, self.batch_size, result, ar_mask, device=self.device
712 mazes, paths = self.seq2map(input)
713 _, predicted_paths = self.seq2map(result)
715 filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
720 predicted_paths=predicted_paths,
721 path_correct=maze.path_correctness(mazes, predicted_paths),
722 path_optimal=maze.path_optimality(paths, predicted_paths),
724 log_string(f"wrote {filename}")
729 ######################################################################
735 class TaskSnake(Task):
746 device=torch.device("cpu"),
748 self.batch_size = batch_size
752 self.prompt_length = prompt_length
754 self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
763 self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
773 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
775 def batches(self, split="train", nb_to_use=-1, desc=None):
776 assert split in {"train", "test"}
777 input = self.train_input if split == "train" else self.test_input
779 input = input[:nb_to_use]
781 desc = f"epoch-{split}"
782 for batch in tqdm.tqdm(
783 input.split(self.batch_size), dynamic_ncols=True, desc=desc
787 def vocabulary_size(self):
790 def produce_results(self, n_epoch, model):
791 with torch.autograd.no_grad():
795 def compute_nb_correct(input, prior_visits):
796 result = input.clone()
797 i = torch.arange(result.size(1), device=result.device)[None, :]
799 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
803 result *= 1 - ar_mask
805 # snake.solver(result,ar_mask)
807 masked_inplace_autoregression(
808 model, self.batch_size, result, ar_mask, device=self.device
811 nb_total = ((prior_visits > 0) * ar_mask).sum()
814 (result == input).long() * (prior_visits > 0) * ar_mask
817 # nb_total = result.size(0)
818 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
820 return nb_total, nb_correct
822 # train_nb_total, train_nb_correct = compute_nb_correct(
823 # self.train_input, self.train_prior_visits
827 # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
830 test_nb_total, test_nb_correct = compute_nb_correct(
831 self.test_input[:1000], self.test_prior_visits[:1000]
835 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
841 ######################################################################
844 def picoclvr_pruner_horizontal_green(p):
845 return not ("green" in p and ("left" in p or "right" in p))
848 picoclvr_pruner_train = (
849 picoclvr_pruner_horizontal_green
850 if args.picocvlr_prune_properties in {"train+eval"}
854 picoclvr_pruner_eval = (
855 (lambda p: not picoclvr_pruner_horizontal_green(p))
856 if args.picocvlr_prune_properties in {"train+eval", "eval"}
860 ######################################################################
862 if args.task == "picoclvr":
864 nb_train_samples=args.nb_train_samples,
865 nb_test_samples=args.nb_test_samples,
866 batch_size=args.batch_size,
867 height=args.picoclvr_height,
868 width=args.picoclvr_width,
869 nb_colors=args.picoclvr_nb_colors,
871 pruner_train=picoclvr_pruner_train,
872 pruner_eval=picoclvr_pruner_eval,
875 elif args.task == "mnist":
877 batch_size=args.batch_size,
881 elif args.task == "maze":
883 nb_train_samples=args.nb_train_samples,
884 nb_test_samples=args.nb_test_samples,
885 batch_size=args.batch_size,
886 height=args.maze_height,
887 width=args.maze_width,
888 nb_walls=args.maze_nb_walls,
892 elif args.task == "snake":
894 nb_train_samples=args.nb_train_samples,
895 nb_test_samples=args.nb_test_samples,
896 batch_size=args.batch_size,
897 height=args.snake_height,
898 width=args.snake_width,
899 nb_colors=args.snake_nb_colors,
900 length=args.snake_length,
901 prompt_length=args.snake_length // 2,
906 raise ValueError(f"Unknown task {args.task}")
908 ######################################################################
910 log_string(f"device {device}")
912 vocabulary_size = task.vocabulary_size()
914 log_string(f"vocabulary_size {vocabulary_size}")
916 ##############################
919 vocabulary_size=vocabulary_size,
920 dim_model=args.dim_model,
921 dim_keys=args.dim_keys,
922 dim_hidden=args.dim_hidden,
923 nb_heads=args.nb_heads,
924 nb_blocks=args.nb_blocks,
926 dropout=args.dropout,
931 nb_parameters = sum(p.numel() for p in model.parameters())
932 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
934 ######################################################################
936 nb_epochs_finished = 0
938 if args.no_checkpoint:
939 log_string(f"not trying to load checkpoint.")
943 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
944 checkpoint = torch.load(checkpoint_name)
945 nb_epochs_finished = checkpoint["nb_epochs_finished"]
946 model.load_state_dict(checkpoint["model_state"])
947 torch.set_rng_state(checkpoint["rng_state"])
948 if torch.cuda.is_available():
949 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
951 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
953 except FileNotFoundError:
954 log_string("starting from scratch.")
957 log_string("error when loading the checkpoint.")
960 ######################################################################
962 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
965 for input in task.batches(split="train"):
966 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
967 token_probas = token_count / token_count.sum()
968 entropy = -torch.xlogy(token_probas, token_probas).sum()
969 train_set_perplexity = math.exp(entropy)
971 ##############################
973 if args.learning_rate_schedule == "cos":
974 learning_rate_schedule = {}
975 for n_epoch in range(args.nb_epochs):
976 u = n_epoch / args.nb_epochs * math.pi
977 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
982 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
986 learning_rate_schedule = {}
987 learning_rate = args.learning_rate
988 for n_epoch in range(args.nb_epochs):
990 learning_rate = u[n_epoch]
991 learning_rate_schedule[n_epoch] = learning_rate
993 log_string(f"learning_rate_schedule {learning_rate_schedule}")
995 ##############################
999 if nb_epochs_finished >= nb_epochs:
1000 task.produce_results(nb_epochs_finished, model)
1002 for n_epoch in range(nb_epochs_finished, nb_epochs):
1003 learning_rate = learning_rate_schedule[n_epoch]
1005 log_string(f"learning_rate {learning_rate}")
1007 if args.optim == "sgd":
1008 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
1009 elif args.optim == "adam":
1010 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
1011 elif args.optim == "adamw":
1012 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
1014 raise ValueError(f"Unknown optimizer {args.optim}.")
1018 nb_train_samples, acc_train_loss = 0, 0.0
1020 for input in task.batches(split="train"):
1021 input = input.to(device)
1022 output = model(mygpt.BracketedSequence(input)).x
1023 loss = F.cross_entropy(output.transpose(1, 2), input)
1024 acc_train_loss += loss.item() * input.size(0)
1025 nb_train_samples += input.size(0)
1026 nb_samples_seen += input.size(0)
1028 optimizer.zero_grad()
1032 with torch.autograd.no_grad():
1035 nb_test_samples, acc_test_loss = 0, 0.0
1037 for input in task.batches(split="test"):
1038 input = input.to(device)
1040 output = model(mygpt.BracketedSequence(input)).x
1041 loss = F.cross_entropy(output.transpose(1, 2), input)
1042 acc_test_loss += loss.item() * input.size(0)
1043 nb_test_samples += input.size(0)
1045 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
1046 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
1049 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1052 task.produce_results(n_epoch, model)
1055 "nb_epochs_finished": n_epoch + 1,
1056 "model_state": model.state_dict(),
1057 "rng_state": torch.get_rng_state(),
1060 if torch.cuda.is_available():
1061 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1063 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1064 torch.save(checkpoint, checkpoint_name)
1065 log_string(f"saved checkpoint {checkpoint_name}")
1067 ######################################################################