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,
34 parser.add_argument("--task", type=str, default="picoclvr")
36 parser.add_argument("--log_filename", type=str, default="train.log")
38 parser.add_argument("--result_dir", type=str, default="results_default")
40 parser.add_argument("--seed", type=int, default=0)
42 parser.add_argument("--nb_epochs", type=int, default=None)
44 parser.add_argument("--batch_size", type=int, default=None)
46 parser.add_argument("--nb_train_samples", type=int, default=250000)
48 parser.add_argument("--nb_test_samples", type=int, default=10000)
50 parser.add_argument("--optim", type=str, default="adam")
52 parser.add_argument("--learning_rate", type=float, default=1e-4)
54 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
56 parser.add_argument("--dim_model", type=int, default=512)
58 parser.add_argument("--dim_keys", type=int, default=64)
60 parser.add_argument("--dim_hidden", type=int, default=2048)
62 parser.add_argument("--nb_heads", type=int, default=8)
64 parser.add_argument("--nb_blocks", type=int, default=12)
66 parser.add_argument("--dropout", type=float, default=0.1)
68 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
70 parser.add_argument("--no_checkpoint", action="store_true", default=False)
72 parser.add_argument("--overwrite_results", action="store_true", default=False)
74 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
76 ##############################
79 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
81 parser.add_argument("--picoclvr_height", type=int, default=12)
83 parser.add_argument("--picoclvr_width", type=int, default=16)
85 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
87 ##############################
90 parser.add_argument("--maze_height", type=int, default=13)
92 parser.add_argument("--maze_width", type=int, default=21)
94 parser.add_argument("--maze_nb_walls", type=int, default=15)
96 ##############################
99 parser.add_argument("--snake_height", type=int, default=6)
101 parser.add_argument("--snake_width", type=int, default=8)
103 parser.add_argument("--snake_nb_colors", type=int, default=5)
105 parser.add_argument("--snake_length", type=int, default=200)
107 ######################################################################
109 args = parser.parse_args()
111 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
114 os.mkdir(args.result_dir)
115 except FileExistsError:
116 if not args.overwrite_results:
117 print(f"result directory {args.result_dir} already exists")
120 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
123 # torch.backends.cudnn.deterministic = True
124 # torch.backends.cudnn.benchmark = False
125 # torch.use_deterministic_algorithms(True)
126 torch.manual_seed(args.seed)
127 if torch.cuda.is_available():
128 torch.cuda.manual_seed_all(args.seed)
130 ######################################################################
151 if args.task in default_args:
152 for k, v in default_args[args.task].items():
153 if getattr(args, k) is None:
156 ######################################################################
160 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
162 if log_file is not None:
163 log_file.write(t + s + "\n")
171 log_string(f"args.{n} {getattr(args, n)}")
173 ######################################################################
176 def masked_inplace_autoregression(
177 model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu")
179 for input, ar_mask in tqdm.tqdm(
180 zip(input.split(batch_size), ar_mask.split(batch_size)),
182 desc="autoregression",
183 total=input.size(0) // batch_size,
185 i = (ar_mask.sum(0) > 0).nonzero()
188 mygpt.BracketedSequence(input, 0, i.min())
189 ) # Needed to initialize the model's cache
190 for s in range(i.min(), i.max() + 1):
191 output = model(mygpt.BracketedSequence(input, s, 1)).x
192 logits = output[:, s]
193 if forbidden_tokens is not None:
194 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
195 if args.deterministic_synthesis:
196 t_next = logits.argmax(1)
198 dist = torch.distributions.categorical.Categorical(logits=logits)
199 t_next = dist.sample()
200 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
203 ######################################################################
207 def batches(self, split="train"):
210 def vocabulary_size(self):
213 def produce_results(self, n_epoch, model):
217 ######################################################################
222 class TaskPicoCLVR(Task):
223 # Make a tensor from a list of strings
224 def tensorize(self, descr):
225 token_descr = [s.strip().split(" ") for s in descr]
226 l = max([len(s) for s in token_descr])
227 token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
228 id_descr = [[self.token2id[u] for u in s] for s in token_descr]
229 return torch.tensor(id_descr, device=self.device)
231 # Make a list of strings from a tensor
232 def detensorize(self, x):
233 return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
235 # trim all the tensors in the tuple z to remove as much token from
236 # left and right in the first tensor. If z is a tuple, all its
237 # elements are trimed according to the triming for the first
238 def trim(self, z, token="<nul>"):
239 n = self.token2id[token]
242 i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
243 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
244 return tuple([t[:, a:b] for t in z])
246 i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
247 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
250 ######################
251 # Not the cleanest part of the code
253 # Extract the last image of each sequence, from the last <img>
254 # included, and set to <nul> all the tokens from the beginning of
255 # that image to the end
256 def excise_last_image(self, input):
257 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
258 nb_img_tokens = self.height * self.width + 1
260 input = input.clone()
261 t = (input == t_img).long()
262 tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
263 i = (t * tail_masks).nonzero(as_tuple=True)
266 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
268 images = self.trim(input[j])
270 loss_masks = 1 - tail_masks
271 input, loss_masks = self.trim((input, loss_masks))
272 return input, loss_masks, images
274 def add_true_image(self, input, images, loss_masks):
275 t_nul = self.token2id["<nul>"]
276 nb_img_tokens = self.height * self.width + 1
277 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
278 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
279 t = (input == t_nul).long()
280 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
283 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
287 input, loss_masks = self.trim((input, loss_masks))
288 return input, loss_masks
290 def add_generated_image(self, input, loss_masks, model):
291 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
292 nb_img_tokens = self.height * self.width + 1
294 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
295 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
296 t = (input == t_nul).long()
297 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
304 + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
306 ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
309 torch.arange(self.vocabulary_size(), device=input.device) == t_nul
311 with torch.autograd.no_grad():
314 masked_inplace_autoregression(
324 input, loss_masks = self.trim((input, loss_masks))
326 return input, loss_masks
328 ######################
338 device=torch.device("cpu"),
342 def generate_descr(nb, cache_suffix, pruner):
343 return picoclvr.generate(
353 self.batch_size = batch_size
355 self.pruner_train = pruner_train
356 self.pruner_eval = pruner_eval
359 "nb_train_samples": nb_train_samples,
360 "nb_test_samples": nb_test_samples,
363 "nb_colors": nb_colors,
364 "batch_size": batch_size,
365 "rng_state": list(torch.get_rng_state()),
369 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
371 self.train_descr = generate_descr(
372 nb_train_samples, "train", pruner=self.pruner_train
374 self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
376 # Build the tokenizer
377 tokens = {"<nul>", "<img>"}
378 for d in [self.train_descr, self.test_descr]:
380 for t in s.strip().split(" "):
382 # make this set a sorted list to get the same tensors given
384 tokens = list(tokens)
386 self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
387 self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
389 # Tokenize the train and test sets
390 self.train_input = self.tensorize(self.train_descr)
391 self.test_input = self.tensorize(self.test_descr)
393 def batches(self, split="train"):
394 assert split in {"train", "test"}
395 input = self.train_input if split == "train" else self.test_input
396 for batch in tqdm.tqdm(
397 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
399 yield self.trim(batch)
401 def vocabulary_size(self):
402 return len(self.token2id)
404 def compute_missing_properties(self, n_epoch, model, pruner=None):
405 acc_nb_requested_properties = []
406 acc_nb_missing_properties = []
409 for input in tqdm.tqdm(
410 self.test_input.split(self.batch_size),
412 desc=f"test-properties",
414 tape, loss_masks, _ = self.excise_last_image(input)
415 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
416 result_descr = self.detensorize(tape)
417 np = picoclvr.nb_properties(
423 nb_requested_properties, _, nb_missing_properties = zip(*np)
424 acc_nb_requested_properties += nb_requested_properties
425 acc_nb_missing_properties += nb_missing_properties
426 acc_nb_results += len(result_descr)
428 nb_requested_properties = sum(acc_nb_requested_properties)
429 nb_missing_properties = sum(acc_nb_missing_properties)
431 prefix = "" if pruner is None else "pruned_"
432 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
434 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
437 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
440 ######################################################################
442 def produce_results(self, n_epoch, model):
443 self.compute_missing_properties(n_epoch, model)
445 if self.pruner_eval is not None:
446 self.compute_missing_properties(n_epoch, model, self.pruner_eval)
448 nb_tokens_to_generate = self.height * self.width + 3
453 for primer_descr in [
454 "red above green <sep> green top <sep> blue right of red",
455 "there is red <sep> there is yellow <sep> there is blue",
456 "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
457 "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
459 primer += [primer_descr] * nb_per_primer
461 tape = self.tensorize(primer)
462 loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
463 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
464 result_descr = self.detensorize(tape)
466 np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
468 acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
469 acc_nb_results = len(result_descr)
471 nb_requested_properties = sum(acc_nb_requested_properties)
472 nb_missing_properties = sum(acc_nb_missing_properties)
475 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
477 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
480 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
483 img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
487 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
491 torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
497 image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
498 torchvision.utils.save_image(
499 img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
501 log_string(f"wrote {image_name}")
504 ######################################################################
507 class TaskMNIST(Task):
508 def __init__(self, batch_size, device=torch.device("cpu")):
510 self.batch_size = batch_size
512 def batches(self, split="train"):
513 assert split in {"train", "test"}
514 data_set = torchvision.datasets.MNIST(
515 root="./data", train=(split == "train"), download=True
517 data_input = data_set.data.view(-1, 28 * 28).long()
518 if args.nb_train_samples is not None:
519 data_input = data_input[: args.nb_train_samples]
520 for batch in tqdm.tqdm(
521 data_input.split(self.batch_size), desc=f"epoch-{split}"
525 def vocabulary_size(self):
528 def produce_results(self, n_epoch, model):
529 results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
530 ar_mask = torch.full_like(results, 1)
531 masked_inplace_autoregression(
532 model, self.batch_size, results, ar_mask, device=self.device
534 image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
535 torchvision.utils.save_image(
536 1 - results.reshape(-1, 1, 28, 28) / 255.0,
541 log_string(f"wrote {image_name}")
544 ######################################################################
549 class TaskMaze(Task):
550 def map2seq(self, *m):
551 return torch.cat([x.flatten(1) for x in m], 1)
553 def seq2map(self, s):
554 s = s.reshape(s.size(0), -1, self.height, self.width)
555 return (s[:, k] for k in range(s.size(1)))
565 device=torch.device("cpu"),
567 self.batch_size = batch_size
572 train_mazes, train_paths, _ = maze.create_maze_data(
577 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
579 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
581 test_mazes, test_paths, _ = maze.create_maze_data(
586 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
588 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
590 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
592 def batches(self, split="train", nb_to_use=-1, desc=None):
593 assert split in {"train", "test"}
594 input = self.train_input if split == "train" else self.test_input
596 input = input[:nb_to_use]
598 desc = f"epoch-{split}"
599 for batch in tqdm.tqdm(
600 input.split(self.batch_size), dynamic_ncols=True, desc=desc
604 def vocabulary_size(self):
607 def compute_error(self, model, split="train", nb_to_use=-1):
608 nb_total, nb_correct = 0, 0
609 for input in task.batches(split, nb_to_use):
610 result = input.clone()
611 ar_mask = result.new_zeros(result.size())
612 ar_mask[:, self.height * self.width :] = 1
613 result *= 1 - ar_mask
614 masked_inplace_autoregression(
615 model, self.batch_size, result, ar_mask, device=self.device
617 mazes, paths = self.seq2map(result)
618 nb_correct += maze.path_correctness(mazes, paths).long().sum()
619 nb_total += mazes.size(0)
621 return nb_total, nb_correct
623 def produce_results(self, n_epoch, model):
624 with torch.autograd.no_grad():
628 train_nb_total, train_nb_correct = self.compute_error(
629 model, "train", nb_to_use=1000
632 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
635 test_nb_total, test_nb_correct = self.compute_error(
636 model, "test", nb_to_use=1000
639 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
642 input = self.test_input[:48]
643 result = input.clone()
644 ar_mask = result.new_zeros(result.size())
645 ar_mask[:, self.height * self.width :] = 1
646 result *= 1 - ar_mask
647 masked_inplace_autoregression(
648 model, self.batch_size, result, ar_mask, device=self.device
651 mazes, paths = self.seq2map(input)
652 _, predicted_paths = self.seq2map(result)
654 filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
659 predicted_paths=predicted_paths,
660 path_correct=maze.path_correctness(mazes, predicted_paths),
662 log_string(f"wrote {filename}")
667 ######################################################################
673 class TaskSnake(Task):
684 device=torch.device("cpu"),
686 self.batch_size = batch_size
690 self.prompt_length = prompt_length
692 self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
701 self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
711 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
713 def batches(self, split="train", nb_to_use=-1, desc=None):
714 assert split in {"train", "test"}
715 input = self.train_input if split == "train" else self.test_input
717 input = input[:nb_to_use]
719 desc = f"epoch-{split}"
720 for batch in tqdm.tqdm(
721 input.split(self.batch_size), dynamic_ncols=True, desc=desc
725 def vocabulary_size(self):
728 def produce_results(self, n_epoch, model):
729 with torch.autograd.no_grad():
733 def compute_nb_correct(input, prior_visits):
734 result = input.clone()
735 i = torch.arange(result.size(1), device=result.device)[None, :]
737 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
741 result *= 1 - ar_mask
743 # snake.solver(result,ar_mask)
745 masked_inplace_autoregression(
746 model, self.batch_size, result, ar_mask, device=self.device
749 nb_total = ((prior_visits > 0) * ar_mask).sum()
752 (result == input).long() * (prior_visits > 0) * ar_mask
755 # nb_total = result.size(0)
756 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
758 return nb_total, nb_correct
760 # train_nb_total, train_nb_correct = compute_nb_correct(
761 # self.train_input, self.train_prior_visits
765 # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
768 test_nb_total, test_nb_correct = compute_nb_correct(
769 self.test_input[:1000], self.test_prior_visits[:1000]
773 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
779 ######################################################################
782 def picoclvr_pruner_horizontal_green(p):
783 return not ("green" in p and ("left" in p or "right" in p))
786 picoclvr_pruner_train = (
787 picoclvr_pruner_horizontal_green
788 if args.picocvlr_prune_properties in {"train+eval"}
792 picoclvr_pruner_eval = (
793 (lambda p: not picoclvr_pruner_horizontal_green(p))
794 if args.picocvlr_prune_properties in {"train+eval", "eval"}
798 ######################################################################
800 if args.task == "picoclvr":
802 nb_train_samples=args.nb_train_samples,
803 nb_test_samples=args.nb_test_samples,
804 batch_size=args.batch_size,
805 height=args.picoclvr_height,
806 width=args.picoclvr_width,
807 nb_colors=args.picoclvr_nb_colors,
809 pruner_train=picoclvr_pruner_train,
810 pruner_eval=picoclvr_pruner_eval,
813 elif args.task == "mnist":
815 batch_size=args.batch_size,
819 elif args.task == "maze":
821 nb_train_samples=args.nb_train_samples,
822 nb_test_samples=args.nb_test_samples,
823 batch_size=args.batch_size,
824 height=args.maze_height,
825 width=args.maze_width,
826 nb_walls=args.maze_nb_walls,
830 elif args.task == "snake":
832 nb_train_samples=args.nb_train_samples,
833 nb_test_samples=args.nb_test_samples,
834 batch_size=args.batch_size,
835 height=args.snake_height,
836 width=args.snake_width,
837 nb_colors=args.snake_nb_colors,
838 length=args.snake_length,
839 prompt_length=args.snake_length // 2,
844 raise ValueError(f"Unknown task {args.task}")
846 ######################################################################
848 log_string(f"device {device}")
850 vocabulary_size = task.vocabulary_size()
852 log_string(f"vocabulary_size {vocabulary_size}")
854 ##############################
857 vocabulary_size=vocabulary_size,
858 dim_model=args.dim_model,
859 dim_keys=args.dim_keys,
860 dim_hidden=args.dim_hidden,
861 nb_heads=args.nb_heads,
862 nb_blocks=args.nb_blocks,
864 dropout=args.dropout,
869 nb_parameters = sum(p.numel() for p in model.parameters())
870 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
872 ######################################################################
874 nb_epochs_finished = 0
876 if args.no_checkpoint:
877 log_string(f"not trying to load checkpoint.")
881 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
882 checkpoint = torch.load(checkpoint_name)
883 nb_epochs_finished = checkpoint["nb_epochs_finished"]
884 model.load_state_dict(checkpoint["model_state"])
885 torch.set_rng_state(checkpoint["rng_state"])
886 if torch.cuda.is_available():
887 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
889 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
891 except FileNotFoundError:
892 log_string("starting from scratch.")
895 log_string("error when loading the checkpoint.")
898 ######################################################################
900 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
903 for input in task.batches(split="train"):
904 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
905 token_probas = token_count / token_count.sum()
906 entropy = -torch.xlogy(token_probas, token_probas).sum()
907 train_set_perplexity = math.exp(entropy)
909 ##############################
911 if args.learning_rate_schedule == "cos":
912 learning_rate_schedule = {}
913 for n_epoch in range(args.nb_epochs):
914 u = n_epoch / args.nb_epochs * math.pi
915 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
920 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
924 learning_rate_schedule = {}
925 learning_rate = args.learning_rate
926 for n_epoch in range(args.nb_epochs):
928 learning_rate = u[n_epoch]
929 learning_rate_schedule[n_epoch] = learning_rate
931 log_string(f"learning_rate_schedule {learning_rate_schedule}")
933 ##############################
937 if nb_epochs_finished >= nb_epochs:
938 task.produce_results(nb_epochs_finished, model)
940 for n_epoch in range(nb_epochs_finished, nb_epochs):
941 learning_rate = learning_rate_schedule[n_epoch]
943 log_string(f"learning_rate {learning_rate}")
945 if args.optim == "sgd":
946 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
947 elif args.optim == "adam":
948 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
949 elif args.optim == "adamw":
950 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
952 raise ValueError(f"Unknown optimizer {args.optim}.")
956 nb_train_samples, acc_train_loss = 0, 0.0
958 for input in task.batches(split="train"):
959 input = input.to(device)
960 output = model(mygpt.BracketedSequence(input)).x
961 loss = F.cross_entropy(output.transpose(1, 2), input)
962 acc_train_loss += loss.item() * input.size(0)
963 nb_train_samples += input.size(0)
964 nb_samples_seen += input.size(0)
966 optimizer.zero_grad()
970 with torch.autograd.no_grad():
973 nb_test_samples, acc_test_loss = 0, 0.0
975 for input in task.batches(split="test"):
976 input = input.to(device)
978 # input, loss_masks, true_images = task.excise_last_image(input)
979 # input, loss_masks = task.add_true_image(input, true_images, loss_masks)
981 output = model(mygpt.BracketedSequence(input)).x
982 loss = F.cross_entropy(output.transpose(1, 2), input)
983 acc_test_loss += loss.item() * input.size(0)
984 nb_test_samples += input.size(0)
986 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
987 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
990 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
993 task.produce_results(n_epoch, model)
996 "nb_epochs_finished": n_epoch + 1,
997 "model_state": model.state_dict(),
998 "rng_state": torch.get_rng_state(),
1001 if torch.cuda.is_available():
1002 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1004 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1005 torch.save(checkpoint, checkpoint_name)
1006 log_string(f"saved checkpoint {checkpoint_name}")
1008 ######################################################################