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("--nb_colors", type=int, default=5)
78 parser.add_argument("--height", type=int, default=12)
80 parser.add_argument("--width", type=int, default=16)
82 parser.add_argument("--prune_properties", type=str, default="none")
84 ######################################################################
86 args = parser.parse_args()
88 assert args.prune_properties in {"none", "train+eval", "eval"}
91 os.mkdir(args.result_dir)
92 except FileExistsError:
93 if not args.overwrite_results:
94 print(f"result directory {args.result_dir} already exists")
97 log_file = open(os.path.join(args.result_dir, args.log_filename), "w")
100 # torch.backends.cudnn.deterministic = True
101 # torch.backends.cudnn.benchmark = False
102 # torch.use_deterministic_algorithms(True)
103 torch.manual_seed(args.seed)
104 if torch.cuda.is_available():
105 torch.cuda.manual_seed_all(args.seed)
107 ######################################################################
111 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
113 if log_file is not None:
114 log_file.write(t + s + "\n")
122 log_string(f"args.{n} {getattr(args, n)}")
124 ######################################################################
127 def masked_inplace_autoregression(
128 model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu")
131 for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
132 i = (ar_mask.sum(0) > 0).nonzero()
135 mygpt.BracketedSequence(input, 0, i.min())
136 ) # Needed to initialize the model's cache
137 for s in range(i.min(), i.max() + 1):
138 output = model(mygpt.BracketedSequence(input, s, 1)).x
139 logits = output[:, s]
140 if forbidden_tokens is not None:
141 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
142 if args.deterministic_synthesis:
143 t_next = logits.argmax(1)
145 dist = torch.distributions.categorical.Categorical(logits=logits)
146 t_next = dist.sample()
147 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
150 ######################################################################
154 def batches(self, split="train"):
157 def vocabulary_size(self):
160 def produce_results(self, n_epoch, model):
164 ######################################################################
169 class TaskPicoCLVR(Task):
171 # Make a tensor from a list of strings
172 def tensorize(self, descr):
173 token_descr = [s.strip().split(" ") for s in descr]
174 l = max([len(s) for s in token_descr])
175 token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
176 id_descr = [[self.token2id[u] for u in s] for s in token_descr]
177 return torch.tensor(id_descr, device=self.device)
179 # Make a list of strings from a tensor
180 def detensorize(self, x):
181 return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
183 # trim all the tensors in the tuple z to remove as much token from
184 # left and right in the first tensor. If z is a tuple, all its
185 # elements are trimed according to the triming for the first
186 def trim(self, z, token="<nul>"):
187 n = self.token2id[token]
190 i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
191 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
192 return tuple([t[:, a:b] for t in z])
194 i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
195 a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
198 ######################
199 # Not the cleanest part of the code
201 # Extract the last image of each sequence, from the last <img>
202 # included, and set to <nul> all the tokens from the beginning of
203 # that image to the end
204 def excise_last_image(self, input):
205 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
206 nb_img_tokens = self.height * self.width + 1
208 input = input.clone()
209 t = (input == t_img).long()
210 tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
211 i = (t * tail_masks).nonzero(as_tuple=True)
214 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
216 images = self.trim(input[j])
218 loss_masks = 1 - tail_masks
219 input, loss_masks = self.trim((input, loss_masks))
220 return input, loss_masks, images
222 def add_true_image(self, input, images, loss_masks):
223 t_nul = self.token2id["<nul>"]
224 nb_img_tokens = self.height * self.width + 1
225 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
226 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
227 t = (input == t_nul).long()
228 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
231 i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
235 input, loss_masks = self.trim((input, loss_masks))
236 return input, loss_masks
238 def add_generated_image(self, input, loss_masks, model):
239 t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
240 nb_img_tokens = self.height * self.width + 1
242 input = F.pad(input, (0, nb_img_tokens), value=t_nul)
243 loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
244 t = (input == t_nul).long()
245 i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
252 + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
254 ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
257 torch.arange(self.vocabulary_size(), device=input.device) == t_nul
259 with torch.autograd.no_grad():
262 masked_inplace_autoregression(
272 input, loss_masks = self.trim((input, loss_masks))
274 return input, loss_masks
276 ######################
284 device=torch.device("cpu"),
288 def generate_descr(nb, cache_suffix, pruner):
289 return picoclvr.generate(
299 self.batch_size = batch_size
301 nb = args.data_size if args.data_size > 0 else 250000
302 self.pruner_train = pruner_train
303 self.pruner_eval = pruner_eval
309 "nb_colors": nb_colors,
310 "batch_size": batch_size,
311 "rng_state": list(torch.get_rng_state()),
314 log_string(f"generating {nb} samples (can take some time)")
315 self.train_descr = generate_descr(
316 (nb * 4) // 5, "train", pruner=self.pruner_train
318 self.test_descr = generate_descr((nb * 1) // 5, "test", pruner=None)
320 # Build the tokenizer
321 tokens = {"<nul>", "<img>"}
322 for d in [self.train_descr, self.test_descr]:
324 for t in s.strip().split(" "):
326 # make this set a sorted list to get the same tensors given
328 tokens = list(tokens)
330 self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
331 self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
333 # Tokenize the train and test sets
334 self.train_input = self.tensorize(self.train_descr)
335 self.test_input = self.tensorize(self.test_descr)
337 def batches(self, split="train"):
338 assert split in {"train", "test"}
339 input = self.train_input if split == "train" else self.test_input
340 for batch in tqdm.tqdm(
341 input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
343 yield self.trim(batch)
345 def vocabulary_size(self):
346 return len(self.token2id)
348 def compute_missing_properties(self, n_epoch, model, pruner=None):
350 acc_nb_requested_properties = []
351 acc_nb_missing_properties = []
354 for input in tqdm.tqdm(
355 self.test_input.split(self.batch_size),
357 desc=f"test-properties",
359 tape, loss_masks, _ = self.excise_last_image(input)
360 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
361 result_descr = self.detensorize(tape)
362 np = picoclvr.nb_properties(
368 nb_requested_properties, _, nb_missing_properties = zip(*np)
369 acc_nb_requested_properties += nb_requested_properties
370 acc_nb_missing_properties += nb_missing_properties
371 acc_nb_results += len(result_descr)
373 nb_requested_properties = sum(acc_nb_requested_properties)
374 nb_missing_properties = sum(acc_nb_missing_properties)
376 prefix = "" if pruner is None else "pruned_"
377 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
379 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
382 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
385 ######################################################################
387 def produce_results(self, n_epoch, model):
389 self.compute_missing_properties(n_epoch, model)
391 if self.pruner_eval is not None:
392 self.compute_missing_properties(n_epoch, model, self.pruner_eval)
394 nb_tokens_to_generate = self.height * self.width + 3
399 for primer_descr in [
400 "red above green <sep> green top <sep> blue right of red",
401 "there is red <sep> there is yellow <sep> there is blue",
402 "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
403 "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
405 primer += [primer_descr] * nb_per_primer
407 tape = self.tensorize(primer)
408 loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
409 tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
410 result_descr = self.detensorize(tape)
412 np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
414 acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
415 acc_nb_results = len(result_descr)
417 nb_requested_properties = sum(acc_nb_requested_properties)
418 nb_missing_properties = sum(acc_nb_missing_properties)
421 log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
423 f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
426 f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
429 img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
433 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
437 torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
443 image_name = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png")
444 torchvision.utils.save_image(
445 img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
447 log_string(f"wrote {image_name}")
450 ######################################################################
452 log_string(f"device {device}")
455 def pruner_horizontal_green(p):
456 return not ("green" in p and ("left" in p or "right" in p))
460 batch_size=args.batch_size,
463 nb_colors=args.nb_colors,
465 pruner_train=pruner_horizontal_green
466 if args.prune_properties in {"train+eval"}
468 pruner_eval=(lambda p: not pruner_horizontal_green(p))
469 if args.prune_properties in {"train+eval", "eval"}
473 vocabulary_size = task.vocabulary_size()
475 log_string(f"vocabulary_size {vocabulary_size}")
477 ##############################
480 vocabulary_size=vocabulary_size,
481 dim_model=args.dim_model,
482 dim_keys=args.dim_keys,
483 dim_hidden=args.dim_hidden,
484 nb_heads=args.nb_heads,
485 nb_blocks=args.nb_blocks,
487 dropout=args.dropout,
492 nb_parameters = sum(p.numel() for p in model.parameters())
493 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
495 ######################################################################
497 nb_epochs_finished = 0
499 if args.no_checkpoint:
500 log_string(f"not trying to load checkpoint.")
504 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
505 checkpoint = torch.load(checkpoint_name)
506 nb_epochs_finished = checkpoint["nb_epochs_finished"]
507 model.load_state_dict(checkpoint["model_state"])
508 torch.set_rng_state(checkpoint["rng_state"])
509 if torch.cuda.is_available():
510 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
512 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
514 except FileNotFoundError:
515 log_string("starting from scratch.")
518 log_string("error when loading the checkpoint.")
521 ######################################################################
523 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
526 for input in task.batches(split="train"):
527 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
528 token_probas = token_count / token_count.sum()
529 entropy = -torch.xlogy(token_probas, token_probas).sum()
530 train_set_perplexity = math.exp(entropy)
532 ##############################
534 if args.learning_rate_schedule == "cos":
535 learning_rate_schedule = {}
536 for n_epoch in range(args.nb_epochs):
537 u = n_epoch / args.nb_epochs * math.pi
538 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
543 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
547 learning_rate_schedule = {}
548 learning_rate = args.learning_rate
549 for n_epoch in range(args.nb_epochs):
551 learning_rate = u[n_epoch]
552 learning_rate_schedule[n_epoch] = learning_rate
554 log_string(f"learning_rate_schedule {learning_rate_schedule}")
556 ##############################
560 if nb_epochs_finished >= nb_epochs:
561 task.produce_results(nb_epochs_finished, model)
563 for n_epoch in range(nb_epochs_finished, nb_epochs):
565 learning_rate = learning_rate_schedule[n_epoch]
567 log_string(f"learning_rate {learning_rate}")
569 if args.optim == "sgd":
570 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
571 elif args.optim == "adam":
572 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
573 elif args.optim == "adamw":
574 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
576 raise ValueError(f"Unknown optimizer {args.optim}.")
580 nb_train_samples, acc_train_loss = 0, 0.0
582 for input in task.batches(split="train"):
583 input = input.to(device)
584 output = model(mygpt.BracketedSequence(input)).x
585 loss = F.cross_entropy(output.transpose(1, 2), input)
586 acc_train_loss += loss.item() * input.size(0)
587 nb_train_samples += input.size(0)
588 nb_samples_seen += input.size(0)
590 optimizer.zero_grad()
594 with torch.autograd.no_grad():
598 nb_test_samples, acc_test_loss = 0, 0.0
600 for input in task.batches(split="test"):
601 input = input.to(device)
603 # input, loss_masks, true_images = task.excise_last_image(input)
604 # input, loss_masks = task.add_true_image(input, true_images, loss_masks)
606 output = model(mygpt.BracketedSequence(input)).x
607 loss = F.cross_entropy(output.transpose(1, 2), input)
608 acc_test_loss += loss.item() * input.size(0)
609 nb_test_samples += input.size(0)
611 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
612 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
615 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
618 task.produce_results(n_epoch, model)
621 "nb_epochs_finished": n_epoch + 1,
622 "model_state": model.state_dict(),
623 "rng_state": torch.get_rng_state(),
626 if torch.cuda.is_available():
627 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
629 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
630 torch.save(checkpoint, checkpoint_name)
631 log_string(f"saved checkpoint {checkpoint_name}")
633 ######################################################################