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, itertools, 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 to solve a toy geometric reasoning task."
33 parser.add_argument("--log_filename", type=str, default="train.log")
35 parser.add_argument("--result_dir", type=str, default="results_default")
37 parser.add_argument("--seed", type=int, default=0)
39 parser.add_argument("--nb_epochs", type=int, default=25)
41 parser.add_argument("--batch_size", type=int, default=100)
43 parser.add_argument("--data_size", type=int, default=-1)
45 parser.add_argument("--optim", type=str, default="adam")
47 parser.add_argument("--learning_rate", type=float, default=1e-3)
50 "--learning_rate_schedule", type=str, default="10: 2e-4,20: 4e-5,30: 8e-6"
53 parser.add_argument("--dim_model", type=int, default=512)
55 parser.add_argument("--dim_keys", type=int, default=64)
57 parser.add_argument("--dim_hidden", type=int, default=2048)
59 parser.add_argument("--nb_heads", type=int, default=8)
61 parser.add_argument("--nb_blocks", type=int, default=12)
63 parser.add_argument("--dropout", type=float, default=0.1)
65 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
67 parser.add_argument("--no_checkpoint", action="store_true", default=False)
69 parser.add_argument("--overwrite_results", action="store_true", default=False)
71 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
73 ##############################
76 parser.add_argument("--world_height", type=int, default=23)
78 parser.add_argument("--world_width", type=int, default=31)
80 parser.add_argument("--world_nb_walls", type=int, default=15)
82 ######################################################################
84 args = parser.parse_args()
86 assert args.prune_properties in {"none", "train+eval", "eval"}
89 os.mkdir(args.result_dir)
90 except FileExistsError:
91 if not args.overwrite_results:
92 print(f"result directory {args.result_dir} already exists")
95 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
98 # torch.backends.cudnn.deterministic = True
99 # torch.backends.cudnn.benchmark = False
100 # torch.use_deterministic_algorithms(True)
101 torch.manual_seed(args.seed)
102 if torch.cuda.is_available():
103 torch.cuda.manual_seed_all(args.seed)
105 ######################################################################
109 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
111 if log_file is not None:
112 log_file.write(t + s + "\n")
120 log_string(f"args.{n} {getattr(args, n)}")
122 ######################################################################
125 def masked_inplace_autoregression(
126 model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu")
129 for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
130 i = (ar_mask.sum(0) > 0).nonzero()
133 mygpt.BracketedSequence(input, 0, i.min())
134 ) # Needed to initialize the model's cache
135 for s in range(i.min(), i.max() + 1):
136 output = model(mygpt.BracketedSequence(input, s, 1)).x
137 logits = output[:, s]
138 if forbidden_tokens is not None:
139 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
140 if args.deterministic_synthesis:
141 t_next = logits.argmax(1)
143 dist = torch.distributions.categorical.Categorical(logits=logits)
144 t_next = dist.sample()
145 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
148 ######################################################################
152 def batches(self, split="train"):
155 def vocabulary_size(self):
158 def produce_results(self, n_epoch, model):
162 ######################################################################
167 class TaskPicoCLVR(Task):
169 # Make a tensor from a list of strings
170 def tensorize(self, descr):
171 token_descr = [s.strip().split(" ") for s in descr]
172 l = max([len(s) for s in token_descr])
173 token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
174 id_descr = [[self.token2id[u] for u in s] for s in token_descr]
175 return torch.tensor(id_descr, device=self.device)
177 # Make a list of strings from a tensor
178 def detensorize(self, x):
179 return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
181 # trim all the tensors in the tuple z to remove as much token from
182 # left and right in the first tensor. If z is a tuple, all its
183 # elements are trimed according to the triming for the first
184 def trim(self, z, token="<nul>"):
185 n = self.token2id[token]
188 i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
189 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
190 return tuple([t[:, a:b] for t in z])
192 i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
193 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
196 ######################
197 # Not the cleanest part of the code
199 # Extract the last image of each sequence, from the last <img>
200 # included, and set to <nul> all the tokens from the beginning of
201 # that image to the end
202 def excise_last_image(self, input):
203 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
204 nb_img_tokens = self.height * self.width + 1
206 input = input.clone()
207 t = (input == t_img).long()
208 tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
209 i = (t * tail_masks).nonzero(as_tuple=True)
212 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
214 images = self.trim(input[j])
216 loss_masks = 1 - tail_masks
217 input, loss_masks = self.trim((input, loss_masks))
218 return input, loss_masks, images
220 def add_true_image(self, input, images, loss_masks):
221 t_nul = self.token2id["<nul>"]
222 nb_img_tokens = self.height * self.width + 1
223 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
224 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
225 t = (input == t_nul).long()
226 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
229 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
233 input, loss_masks = self.trim((input, loss_masks))
234 return input, loss_masks
236 def add_generated_image(self, input, loss_masks, model):
237 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
238 nb_img_tokens = self.height * self.width + 1
240 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
241 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
242 t = (input == t_nul).long()
243 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
250 + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
252 ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
255 torch.arange(self.vocabulary_size(), device=input.device) == t_nul
257 with torch.autograd.no_grad():
260 masked_inplace_autoregression(
270 input, loss_masks = self.trim((input, loss_masks))
272 return input, loss_masks
274 ######################
282 device=torch.device("cpu"),
286 def generate_descr(nb, cache_suffix, pruner):
287 return picoclvr.generate(
297 self.batch_size = batch_size
299 nb = args.data_size if args.data_size > 0 else 250000
300 self.pruner_train = pruner_train
301 self.pruner_eval = pruner_eval
307 "nb_colors": nb_colors,
308 "batch_size": batch_size,
309 "rng_state": list(torch.get_rng_state()),
312 log_string(f"generating {nb} samples (can take some time)")
313 self.train_descr = generate_descr(
314 (nb * 4) // 5, "train", pruner=self.pruner_train
316 self.test_descr = generate_descr((nb * 1) // 5, "test", pruner=None)
318 # Build the tokenizer
319 tokens = {"<nul>", "<img>"}
320 for d in [self.train_descr, self.test_descr]:
322 for t in s.strip().split(" "):
324 # make this set a sorted list to get the same tensors given
326 tokens = list(tokens)
328 self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
329 self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
331 # Tokenize the train and test sets
332 self.train_input = self.tensorize(self.train_descr)
333 self.test_input = self.tensorize(self.test_descr)
335 def batches(self, split="train"):
336 assert split in {"train", "test"}
337 input = self.train_input if split == "train" else self.test_input
338 for batch in tqdm.tqdm(
339 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
341 yield self.trim(batch)
343 def vocabulary_size(self):
344 return len(self.token2id)
346 def compute_missing_properties(self, n_epoch, model, pruner=None):
348 acc_nb_requested_properties = []
349 acc_nb_missing_properties = []
352 for input in tqdm.tqdm(
353 self.test_input.split(self.batch_size),
355 desc=f"test-properties",
357 tape, loss_masks, _ = self.excise_last_image(input)
358 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
359 result_descr = self.detensorize(tape)
360 np = picoclvr.nb_properties(
366 nb_requested_properties, _, nb_missing_properties = zip(*np)
367 acc_nb_requested_properties += nb_requested_properties
368 acc_nb_missing_properties += nb_missing_properties
369 acc_nb_results += len(result_descr)
371 nb_requested_properties = sum(acc_nb_requested_properties)
372 nb_missing_properties = sum(acc_nb_missing_properties)
374 prefix = "" if pruner is None else "pruned_"
375 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
377 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
380 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
383 ######################################################################
385 def produce_results(self, n_epoch, model):
387 self.compute_missing_properties(n_epoch, model)
389 if self.pruner_eval is not None:
390 self.compute_missing_properties(n_epoch, model, self.pruner_eval)
392 nb_tokens_to_generate = self.height * self.width + 3
397 for primer_descr in [
398 "red above green <sep> green top <sep> blue right of red",
399 "there is red <sep> there is yellow <sep> there is blue",
400 "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
401 "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
403 primer += [primer_descr] * nb_per_primer
405 tape = self.tensorize(primer)
406 loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
407 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
408 result_descr = self.detensorize(tape)
410 np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
412 acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
413 acc_nb_results = len(result_descr)
415 nb_requested_properties = sum(acc_nb_requested_properties)
416 nb_missing_properties = sum(acc_nb_missing_properties)
419 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
421 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
424 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
427 img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
431 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
435 torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
441 image_name = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png")
442 torchvision.utils.save_image(
443 img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
445 log_string(f"wrote {image_name}")
448 ######################################################################
450 log_string(f"device {device}")
453 def pruner_horizontal_green(p):
454 return not ("green" in p and ("left" in p or "right" in p))
458 batch_size=args.batch_size,
461 nb_colors=args.nb_colors,
463 pruner_train=pruner_horizontal_green
464 if args.prune_properties in {"train+eval"}
466 pruner_eval=(lambda p: not pruner_horizontal_green(p))
467 if args.prune_properties in {"train+eval", "eval"}
471 vocabulary_size = task.vocabulary_size()
473 log_string(f"vocabulary_size {vocabulary_size}")
475 ##############################
478 vocabulary_size=vocabulary_size,
479 dim_model=args.dim_model,
480 dim_keys=args.dim_keys,
481 dim_hidden=args.dim_hidden,
482 nb_heads=args.nb_heads,
483 nb_blocks=args.nb_blocks,
485 dropout=args.dropout,
490 nb_parameters = sum(p.numel() for p in model.parameters())
491 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
493 ######################################################################
495 nb_epochs_finished = 0
497 if args.no_checkpoint:
498 log_string(f"not trying to load checkpoint.")
502 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
503 checkpoint = torch.load(checkpoint_name)
504 nb_epochs_finished = checkpoint["nb_epochs_finished"]
505 model.load_state_dict(checkpoint["model_state"])
506 torch.set_rng_state(checkpoint["rng_state"])
507 if torch.cuda.is_available():
508 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
510 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
512 except FileNotFoundError:
513 log_string("starting from scratch.")
516 log_string("error when loading the checkpoint.")
519 ######################################################################
521 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
524 for input in task.batches(split="train"):
525 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
526 token_probas = token_count / token_count.sum()
527 entropy = -torch.xlogy(token_probas, token_probas).sum()
528 train_set_perplexity = math.exp(entropy)
530 ##############################
532 if args.learning_rate_schedule == "cos":
533 learning_rate_schedule = {}
534 for n_epoch in range(args.nb_epochs):
535 u = n_epoch / args.nb_epochs * math.pi
536 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
541 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
545 learning_rate_schedule = {}
546 learning_rate = args.learning_rate
547 for n_epoch in range(args.nb_epochs):
549 learning_rate = u[n_epoch]
550 learning_rate_schedule[n_epoch] = learning_rate
552 log_string(f"learning_rate_schedule {learning_rate_schedule}")
554 ##############################
558 if nb_epochs_finished >= nb_epochs:
559 task.produce_results(nb_epochs_finished, model)
561 for n_epoch in range(nb_epochs_finished, nb_epochs):
563 learning_rate = learning_rate_schedule[n_epoch]
565 log_string(f"learning_rate {learning_rate}")
567 if args.optim == "sgd":
568 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
569 elif args.optim == "adam":
570 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
571 elif args.optim == "adamw":
572 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
574 raise ValueError(f"Unknown optimizer {args.optim}.")
578 nb_train_samples, acc_train_loss = 0, 0.0
580 for input in task.batches(split="train"):
581 input = input.to(device)
582 output = model(mygpt.BracketedSequence(input)).x
583 loss = F.cross_entropy(output.transpose(1, 2), input)
584 acc_train_loss += loss.item() * input.size(0)
585 nb_train_samples += input.size(0)
586 nb_samples_seen += input.size(0)
588 optimizer.zero_grad()
592 with torch.autograd.no_grad():
596 nb_test_samples, acc_test_loss = 0, 0.0
598 for input in task.batches(split="test"):
599 input = input.to(device)
601 # input, loss_masks, true_images = task.excise_last_image(input)
602 # input, loss_masks = task.add_true_image(input, true_images, loss_masks)
604 output = model(mygpt.BracketedSequence(input)).x
605 loss = F.cross_entropy(output.transpose(1, 2), input)
606 acc_test_loss += loss.item() * input.size(0)
607 nb_test_samples += input.size(0)
609 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
610 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
613 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
616 task.produce_results(n_epoch, model)
619 "nb_epochs_finished": n_epoch + 1,
620 "model_state": model.state_dict(),
621 "rng_state": torch.get_rng_state(),
624 if torch.cuda.is_available():
625 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
627 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
628 torch.save(checkpoint, checkpoint_name)
629 log_string(f"saved checkpoint {checkpoint_name}")
631 ######################################################################