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 # ra_mask is boolean, with 1s on the values to generate
179 def masked_inplace_autoregression(
184 forbidden_tokens=None,
185 progress_bar_desc="autoregression",
186 device=torch.device("cpu"),
188 batches = zip(input.split(batch_size), ar_mask.split(batch_size))
189 if progress_bar_desc is not None:
193 desc=progress_bar_desc,
194 total=input.size(0) // batch_size,
196 for input, ar_mask in batches:
197 i = (ar_mask.sum(0) > 0).nonzero()
200 mygpt.BracketedSequence(input, 0, i.min())
201 ) # Needed to initialize the model's cache
202 for s in range(i.min(), i.max() + 1):
203 output = model(mygpt.BracketedSequence(input, s, 1)).x
204 logits = output[:, s]
205 if forbidden_tokens is not None:
206 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
207 if args.deterministic_synthesis:
208 t_next = logits.argmax(1)
210 dist = torch.distributions.categorical.Categorical(logits=logits)
211 t_next = dist.sample()
212 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
215 ######################################################################
219 def batches(self, split="train"):
222 def vocabulary_size(self):
225 def produce_results(self, n_epoch, model):
229 ######################################################################
234 class TaskPicoCLVR(Task):
235 # Make a tensor from a list of strings
236 def tensorize(self, descr):
237 token_descr = [s.strip().split(" ") for s in descr]
238 l = max([len(s) for s in token_descr])
239 token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
240 id_descr = [[self.token2id[u] for u in s] for s in token_descr]
241 return torch.tensor(id_descr, device=self.device)
243 # Make a list of strings from a tensor
244 def detensorize(self, x):
245 return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
247 # trim all the tensors in the tuple z to remove as much token from
248 # left and right in the first tensor. If z is a tuple, all its
249 # elements are trimed according to the triming for the first
250 def trim(self, z, token="<nul>"):
251 n = self.token2id[token]
254 i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
255 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
256 return tuple([t[:, a:b] for t in z])
258 i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
259 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
262 ######################
263 # Not the cleanest part of the code
265 # Extract the last image of each sequence, from the last <img>
266 # included, and set to <nul> all the tokens from the beginning of
267 # that image to the end
268 def excise_last_image(self, input):
269 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
270 nb_img_tokens = self.height * self.width + 1
272 input = input.clone()
273 t = (input == t_img).long()
274 tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
275 i = (t * tail_masks).nonzero(as_tuple=True)
278 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
280 images = self.trim(input[j])
282 loss_masks = 1 - tail_masks
283 input, loss_masks = self.trim((input, loss_masks))
284 return input, loss_masks, images
286 def add_true_image(self, input, images, loss_masks):
287 t_nul = self.token2id["<nul>"]
288 nb_img_tokens = self.height * self.width + 1
289 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
290 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
291 t = (input == t_nul).long()
292 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
295 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
299 input, loss_masks = self.trim((input, loss_masks))
300 return input, loss_masks
302 def add_generated_image(self, input, loss_masks, model):
303 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
304 nb_img_tokens = self.height * self.width + 1
306 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
307 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
308 t = (input == t_nul).long()
309 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
316 + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
318 ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
321 torch.arange(self.vocabulary_size(), device=input.device) == t_nul
323 with torch.autograd.no_grad():
326 masked_inplace_autoregression(
332 progress_bar_desc=None,
337 input, loss_masks = self.trim((input, loss_masks))
339 return input, loss_masks
341 ######################
351 device=torch.device("cpu"),
355 def generate_descr(nb, cache_suffix, pruner):
356 return picoclvr.generate(
366 self.batch_size = batch_size
368 self.pruner_train = pruner_train
369 self.pruner_eval = pruner_eval
372 "nb_train_samples": nb_train_samples,
373 "nb_test_samples": nb_test_samples,
376 "nb_colors": nb_colors,
377 "batch_size": batch_size,
378 "rng_state": list(torch.get_rng_state()),
382 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
384 self.train_descr = generate_descr(
385 nb_train_samples, "train", pruner=self.pruner_train
387 self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
389 # Build the tokenizer
390 tokens = {"<nul>", "<img>"}
391 for d in [self.train_descr, self.test_descr]:
393 for t in s.strip().split(" "):
395 # make this set a sorted list to get the same tensors given
397 tokens = list(tokens)
399 self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
400 self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
402 # Tokenize the train and test sets
403 self.train_input = self.tensorize(self.train_descr)
404 self.test_input = self.tensorize(self.test_descr)
406 def batches(self, split="train"):
407 assert split in {"train", "test"}
408 input = self.train_input if split == "train" else self.test_input
409 for batch in tqdm.tqdm(
410 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
412 yield self.trim(batch)
414 def vocabulary_size(self):
415 return len(self.token2id)
417 def compute_missing_properties(self, n_epoch, model, pruner=None):
418 acc_nb_requested_properties = []
419 acc_nb_missing_properties = []
422 for input in tqdm.tqdm(
423 self.test_input.split(self.batch_size),
425 desc=f"test-properties",
427 tape, loss_masks, _ = self.excise_last_image(input)
428 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
429 result_descr = self.detensorize(tape)
430 np = picoclvr.nb_properties(
436 nb_requested_properties, _, nb_missing_properties = zip(*np)
437 acc_nb_requested_properties += nb_requested_properties
438 acc_nb_missing_properties += nb_missing_properties
439 acc_nb_results += len(result_descr)
441 nb_requested_properties = sum(acc_nb_requested_properties)
442 nb_missing_properties = sum(acc_nb_missing_properties)
444 prefix = "" if pruner is None else "pruned_"
445 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
447 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
450 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
453 ######################################################################
455 def produce_results(self, n_epoch, model):
456 self.compute_missing_properties(n_epoch, model)
458 if self.pruner_eval is not None:
459 self.compute_missing_properties(n_epoch, model, self.pruner_eval)
461 nb_tokens_to_generate = self.height * self.width + 3
466 for primer_descr in [
467 "red above green <sep> green top <sep> blue right of red",
468 "there is red <sep> there is yellow <sep> there is blue",
469 "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
470 "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
472 primer += [primer_descr] * nb_per_primer
474 tape = self.tensorize(primer)
475 loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
476 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
477 result_descr = self.detensorize(tape)
479 np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
481 acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
482 acc_nb_results = len(result_descr)
484 nb_requested_properties = sum(acc_nb_requested_properties)
485 nb_missing_properties = sum(acc_nb_missing_properties)
488 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
490 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
493 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
496 img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
500 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
504 torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
510 image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
511 torchvision.utils.save_image(
512 img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
514 log_string(f"wrote {image_name}")
517 ######################################################################
520 class TaskMNIST(Task):
521 def __init__(self, batch_size, device=torch.device("cpu")):
523 self.batch_size = batch_size
525 def batches(self, split="train"):
526 assert split in {"train", "test"}
527 data_set = torchvision.datasets.MNIST(
528 root="./data", train=(split == "train"), download=True
530 data_input = data_set.data.view(-1, 28 * 28).long()
531 if args.nb_train_samples is not None:
532 data_input = data_input[: args.nb_train_samples]
533 for batch in tqdm.tqdm(
534 data_input.split(self.batch_size), desc=f"epoch-{split}"
538 def vocabulary_size(self):
541 def produce_results(self, n_epoch, model):
542 results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
543 ar_mask = torch.full_like(results, 1)
544 masked_inplace_autoregression(
545 model, self.batch_size, results, ar_mask, device=self.device
547 image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
548 torchvision.utils.save_image(
549 1 - results.reshape(-1, 1, 28, 28) / 255.0,
554 log_string(f"wrote {image_name}")
557 ######################################################################
562 class TaskMaze(Task):
563 def map2seq(self, *m):
564 return torch.cat([x.flatten(1) for x in m], 1)
566 def seq2map(self, s):
567 s = s.reshape(s.size(0), -1, self.height, self.width)
568 return (s[:, k] for k in range(s.size(1)))
578 device=torch.device("cpu"),
580 self.batch_size = batch_size
585 train_mazes, train_paths, _ = maze.create_maze_data(
590 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
592 self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
594 test_mazes, test_paths, _ = maze.create_maze_data(
599 progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
601 self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
603 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
605 def batches(self, split="train", nb_to_use=-1, desc=None):
606 assert split in {"train", "test"}
607 input = self.train_input if split == "train" else self.test_input
609 input = input[:nb_to_use]
611 desc = f"epoch-{split}"
612 for batch in tqdm.tqdm(
613 input.split(self.batch_size), dynamic_ncols=True, desc=desc
617 def vocabulary_size(self):
620 def compute_error(self, model, split="train", nb_to_use=-1):
621 nb_total, nb_correct = 0, 0
622 for input in task.batches(split, nb_to_use):
623 result = input.clone()
624 ar_mask = result.new_zeros(result.size())
625 ar_mask[:, self.height * self.width :] = 1
626 result *= 1 - ar_mask
627 masked_inplace_autoregression(
628 model, self.batch_size, result, ar_mask, device=self.device
630 mazes, paths = self.seq2map(result)
631 nb_correct += maze.path_correctness(mazes, paths).long().sum()
632 nb_total += mazes.size(0)
634 return nb_total, nb_correct
636 def produce_results(self, n_epoch, model):
637 with torch.autograd.no_grad():
641 train_nb_total, train_nb_correct = self.compute_error(
642 model, "train", nb_to_use=1000
645 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
648 test_nb_total, test_nb_correct = self.compute_error(
649 model, "test", nb_to_use=1000
652 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
655 input = self.test_input[:48]
656 result = input.clone()
657 ar_mask = result.new_zeros(result.size())
658 ar_mask[:, self.height * self.width :] = 1
659 result *= 1 - ar_mask
660 masked_inplace_autoregression(
661 model, self.batch_size, result, ar_mask, device=self.device
664 mazes, paths = self.seq2map(input)
665 _, predicted_paths = self.seq2map(result)
667 filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
672 predicted_paths=predicted_paths,
673 path_correct=maze.path_correctness(mazes, predicted_paths),
675 log_string(f"wrote {filename}")
680 ######################################################################
686 class TaskSnake(Task):
697 device=torch.device("cpu"),
699 self.batch_size = batch_size
703 self.prompt_length = prompt_length
705 self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
714 self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
724 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
726 def batches(self, split="train", nb_to_use=-1, desc=None):
727 assert split in {"train", "test"}
728 input = self.train_input if split == "train" else self.test_input
730 input = input[:nb_to_use]
732 desc = f"epoch-{split}"
733 for batch in tqdm.tqdm(
734 input.split(self.batch_size), dynamic_ncols=True, desc=desc
738 def vocabulary_size(self):
741 def produce_results(self, n_epoch, model):
742 with torch.autograd.no_grad():
746 def compute_nb_correct(input, prior_visits):
747 result = input.clone()
748 i = torch.arange(result.size(1), device=result.device)[None, :]
750 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
754 result *= 1 - ar_mask
756 # snake.solver(result,ar_mask)
758 masked_inplace_autoregression(
759 model, self.batch_size, result, ar_mask, device=self.device
762 nb_total = ((prior_visits > 0) * ar_mask).sum()
765 (result == input).long() * (prior_visits > 0) * ar_mask
768 # nb_total = result.size(0)
769 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
771 return nb_total, nb_correct
773 # train_nb_total, train_nb_correct = compute_nb_correct(
774 # self.train_input, self.train_prior_visits
778 # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
781 test_nb_total, test_nb_correct = compute_nb_correct(
782 self.test_input[:1000], self.test_prior_visits[:1000]
786 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
792 ######################################################################
795 def picoclvr_pruner_horizontal_green(p):
796 return not ("green" in p and ("left" in p or "right" in p))
799 picoclvr_pruner_train = (
800 picoclvr_pruner_horizontal_green
801 if args.picocvlr_prune_properties in {"train+eval"}
805 picoclvr_pruner_eval = (
806 (lambda p: not picoclvr_pruner_horizontal_green(p))
807 if args.picocvlr_prune_properties in {"train+eval", "eval"}
811 ######################################################################
813 if args.task == "picoclvr":
815 nb_train_samples=args.nb_train_samples,
816 nb_test_samples=args.nb_test_samples,
817 batch_size=args.batch_size,
818 height=args.picoclvr_height,
819 width=args.picoclvr_width,
820 nb_colors=args.picoclvr_nb_colors,
822 pruner_train=picoclvr_pruner_train,
823 pruner_eval=picoclvr_pruner_eval,
826 elif args.task == "mnist":
828 batch_size=args.batch_size,
832 elif args.task == "maze":
834 nb_train_samples=args.nb_train_samples,
835 nb_test_samples=args.nb_test_samples,
836 batch_size=args.batch_size,
837 height=args.maze_height,
838 width=args.maze_width,
839 nb_walls=args.maze_nb_walls,
843 elif args.task == "snake":
845 nb_train_samples=args.nb_train_samples,
846 nb_test_samples=args.nb_test_samples,
847 batch_size=args.batch_size,
848 height=args.snake_height,
849 width=args.snake_width,
850 nb_colors=args.snake_nb_colors,
851 length=args.snake_length,
852 prompt_length=args.snake_length // 2,
857 raise ValueError(f"Unknown task {args.task}")
859 ######################################################################
861 log_string(f"device {device}")
863 vocabulary_size = task.vocabulary_size()
865 log_string(f"vocabulary_size {vocabulary_size}")
867 ##############################
870 vocabulary_size=vocabulary_size,
871 dim_model=args.dim_model,
872 dim_keys=args.dim_keys,
873 dim_hidden=args.dim_hidden,
874 nb_heads=args.nb_heads,
875 nb_blocks=args.nb_blocks,
877 dropout=args.dropout,
882 nb_parameters = sum(p.numel() for p in model.parameters())
883 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
885 ######################################################################
887 nb_epochs_finished = 0
889 if args.no_checkpoint:
890 log_string(f"not trying to load checkpoint.")
894 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
895 checkpoint = torch.load(checkpoint_name)
896 nb_epochs_finished = checkpoint["nb_epochs_finished"]
897 model.load_state_dict(checkpoint["model_state"])
898 torch.set_rng_state(checkpoint["rng_state"])
899 if torch.cuda.is_available():
900 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
902 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
904 except FileNotFoundError:
905 log_string("starting from scratch.")
908 log_string("error when loading the checkpoint.")
911 ######################################################################
913 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
916 for input in task.batches(split="train"):
917 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
918 token_probas = token_count / token_count.sum()
919 entropy = -torch.xlogy(token_probas, token_probas).sum()
920 train_set_perplexity = math.exp(entropy)
922 ##############################
924 if args.learning_rate_schedule == "cos":
925 learning_rate_schedule = {}
926 for n_epoch in range(args.nb_epochs):
927 u = n_epoch / args.nb_epochs * math.pi
928 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
933 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
937 learning_rate_schedule = {}
938 learning_rate = args.learning_rate
939 for n_epoch in range(args.nb_epochs):
941 learning_rate = u[n_epoch]
942 learning_rate_schedule[n_epoch] = learning_rate
944 log_string(f"learning_rate_schedule {learning_rate_schedule}")
946 ##############################
950 if nb_epochs_finished >= nb_epochs:
951 task.produce_results(nb_epochs_finished, model)
953 for n_epoch in range(nb_epochs_finished, nb_epochs):
954 learning_rate = learning_rate_schedule[n_epoch]
956 log_string(f"learning_rate {learning_rate}")
958 if args.optim == "sgd":
959 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
960 elif args.optim == "adam":
961 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
962 elif args.optim == "adamw":
963 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
965 raise ValueError(f"Unknown optimizer {args.optim}.")
969 nb_train_samples, acc_train_loss = 0, 0.0
971 for input in task.batches(split="train"):
972 input = input.to(device)
973 output = model(mygpt.BracketedSequence(input)).x
974 loss = F.cross_entropy(output.transpose(1, 2), input)
975 acc_train_loss += loss.item() * input.size(0)
976 nb_train_samples += input.size(0)
977 nb_samples_seen += input.size(0)
979 optimizer.zero_grad()
983 with torch.autograd.no_grad():
986 nb_test_samples, acc_test_loss = 0, 0.0
988 for input in task.batches(split="test"):
989 input = input.to(device)
991 output = model(mygpt.BracketedSequence(input)).x
992 loss = F.cross_entropy(output.transpose(1, 2), input)
993 acc_test_loss += loss.item() * input.size(0)
994 nb_test_samples += input.size(0)
996 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
997 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
1000 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1003 task.produce_results(n_epoch, model)
1006 "nb_epochs_finished": n_epoch + 1,
1007 "model_state": model.state_dict(),
1008 "rng_state": torch.get_rng_state(),
1011 if torch.cuda.is_available():
1012 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1014 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1015 torch.save(checkpoint, checkpoint_name)
1016 log_string(f"saved checkpoint {checkpoint_name}")
1018 ######################################################################