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
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,
38 help="picoclvr, mnist, maze, snake, stack, expr, world",
41 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
43 parser.add_argument("--result_dir", type=str, default=None)
45 parser.add_argument("--seed", type=int, default=0)
47 parser.add_argument("--nb_epochs", type=int, default=None)
49 parser.add_argument("--batch_size", type=int, default=None)
51 parser.add_argument("--nb_train_samples", type=int, default=None)
53 parser.add_argument("--nb_test_samples", type=int, default=None)
55 parser.add_argument("--optim", type=str, default="adam")
57 parser.add_argument("--learning_rate", type=float, default=1e-4)
59 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
61 parser.add_argument("--dim_model", type=int, default=512)
63 parser.add_argument("--dim_keys", type=int, default=64)
65 parser.add_argument("--dim_hidden", type=int, default=2048)
67 parser.add_argument("--nb_heads", type=int, default=8)
69 parser.add_argument("--nb_blocks", type=int, default=12)
71 parser.add_argument("--dropout", type=float, default=0.1)
73 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
75 parser.add_argument("--no_checkpoint", action="store_true", default=False)
77 parser.add_argument("--overwrite_results", action="store_true", default=False)
79 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
81 ##############################
84 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
86 parser.add_argument("--picoclvr_height", type=int, default=12)
88 parser.add_argument("--picoclvr_width", type=int, default=16)
90 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
92 ##############################
95 parser.add_argument("--maze_height", type=int, default=23)
97 parser.add_argument("--maze_width", type=int, default=39)
99 parser.add_argument("--maze_nb_walls", type=int, default=45)
101 ##############################
104 parser.add_argument("--snake_height", type=int, default=6)
106 parser.add_argument("--snake_width", type=int, default=8)
108 parser.add_argument("--snake_nb_colors", type=int, default=5)
110 parser.add_argument("--snake_length", type=int, default=200)
112 ##############################
115 parser.add_argument("--stack_nb_steps", type=int, default=100)
117 parser.add_argument("--stack_nb_stacks", type=int, default=3)
119 parser.add_argument("--stack_nb_digits", type=int, default=3)
121 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
123 ##############################
126 parser.add_argument("--expr_nb_variables", type=int, default=5)
128 parser.add_argument("--expr_sequence_length", type=int, default=40)
130 parser.add_argument("--expr_operand_max", type=int, default=9)
132 parser.add_argument("--expr_result_max", type=int, default=99)
134 parser.add_argument("--expr_input_file", type=str, default=None)
136 ##############################
139 parser.add_argument("--world_vqae_nb_epochs", type=int, default=10)
141 ######################################################################
143 args = parser.parse_args()
145 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
147 if args.result_dir is None:
148 args.result_dir = f"results_{args.task}"
150 ######################################################################
156 "nb_train_samples": 250000,
157 "nb_test_samples": 10000,
162 "nb_train_samples": 250000,
163 "nb_test_samples": 10000,
168 "nb_train_samples": 250000,
169 "nb_test_samples": 10000,
174 "nb_train_samples": 250000,
175 "nb_test_samples": 10000,
180 "nb_train_samples": 100000,
181 "nb_test_samples": 1000,
186 "nb_train_samples": 1000000,
187 "nb_test_samples": 10000,
192 "nb_train_samples": 10000,
193 "nb_test_samples": 1000,
197 if args.task in default_args:
198 for k, v in default_args[args.task].items():
199 if getattr(args, k) is None:
202 ######################################################################
205 os.mkdir(args.result_dir)
206 except FileExistsError:
207 if not args.overwrite_results:
208 print(f"result directory {args.result_dir} already exists")
211 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
214 # torch.backends.cudnn.deterministic = True
215 # torch.backends.cudnn.benchmark = False
216 # torch.use_deterministic_algorithms(True)
217 torch.manual_seed(args.seed)
218 if torch.cuda.is_available():
219 torch.cuda.manual_seed_all(args.seed)
221 ######################################################################
225 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
227 if log_file is not None:
228 log_file.write(t + s + "\n")
236 log_string(f"args.{n} {getattr(args, n)}")
239 ######################################################################
242 def picoclvr_pruner_horizontal_green(p):
243 return not ("green" in p and ("left" in p or "right" in p))
246 picoclvr_pruner_train = (
247 picoclvr_pruner_horizontal_green
248 if args.picocvlr_prune_properties in {"train+eval"}
252 picoclvr_pruner_eval = (
253 (lambda p: not picoclvr_pruner_horizontal_green(p))
254 if args.picocvlr_prune_properties in {"train+eval", "eval"}
258 ######################################################################
260 if args.task == "picoclvr":
261 task = tasks.PicoCLVR(
262 nb_train_samples=args.nb_train_samples,
263 nb_test_samples=args.nb_test_samples,
264 batch_size=args.batch_size,
265 height=args.picoclvr_height,
266 width=args.picoclvr_width,
267 nb_colors=args.picoclvr_nb_colors,
270 pruner_train=picoclvr_pruner_train,
271 pruner_eval=picoclvr_pruner_eval,
274 elif args.task == "mnist":
276 nb_train_samples=args.nb_train_samples,
277 nb_test_samples=args.nb_test_samples,
278 batch_size=args.batch_size,
282 elif args.task == "maze":
284 nb_train_samples=args.nb_train_samples,
285 nb_test_samples=args.nb_test_samples,
286 batch_size=args.batch_size,
287 height=args.maze_height,
288 width=args.maze_width,
289 nb_walls=args.maze_nb_walls,
293 elif args.task == "snake":
295 nb_train_samples=args.nb_train_samples,
296 nb_test_samples=args.nb_test_samples,
297 batch_size=args.batch_size,
298 height=args.snake_height,
299 width=args.snake_width,
300 nb_colors=args.snake_nb_colors,
301 length=args.snake_length,
302 prompt_length=args.snake_length // 2,
306 elif args.task == "stack":
308 nb_train_samples=args.nb_train_samples,
309 nb_test_samples=args.nb_test_samples,
310 batch_size=args.batch_size,
312 nb_steps=args.stack_nb_steps,
313 nb_stacks=args.stack_nb_stacks,
314 nb_digits=args.stack_nb_digits,
315 fraction_values_for_train=args.stack_fraction_values_for_train,
319 elif args.task == "expr":
321 nb_train_samples=args.nb_train_samples,
322 nb_test_samples=args.nb_test_samples,
323 nb_variables=args.expr_nb_variables,
324 sequence_length=args.expr_sequence_length,
325 operand_max=args.expr_operand_max,
326 result_max=args.expr_result_max,
327 batch_size=args.batch_size,
331 elif args.task == "world":
333 nb_train_samples=args.nb_train_samples,
334 nb_test_samples=args.nb_test_samples,
335 batch_size=args.batch_size,
336 vqae_nb_epochs=args.world_vqae_nb_epochs,
341 raise ValueError(f"Unknown task {args.task}")
343 ######################################################################
345 log_string(f"device {device}")
347 vocabulary_size = task.vocabulary_size()
349 log_string(f"vocabulary_size {vocabulary_size}")
351 ##############################
354 vocabulary_size=vocabulary_size,
355 dim_model=args.dim_model,
356 dim_keys=args.dim_keys,
357 dim_hidden=args.dim_hidden,
358 nb_heads=args.nb_heads,
359 nb_blocks=args.nb_blocks,
361 dropout=args.dropout,
366 nb_parameters = sum(p.numel() for p in model.parameters())
367 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
369 ######################################################################
371 nb_epochs_finished = 0
373 if args.no_checkpoint:
374 log_string(f"not trying to load checkpoint.")
378 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
379 checkpoint = torch.load(checkpoint_name)
380 nb_epochs_finished = checkpoint["nb_epochs_finished"]
381 model.load_state_dict(checkpoint["model_state"])
382 torch.set_rng_state(checkpoint["rng_state"])
383 if torch.cuda.is_available():
384 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
386 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
388 except FileNotFoundError:
389 log_string("starting from scratch.")
392 log_string("error when loading the checkpoint.")
395 ######################################################################
397 if args.task == "expr" and args.expr_input_file is not None:
398 task.produce_results(
403 args.deterministic_synthesis,
404 args.expr_input_file,
409 ######################################################################
411 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
413 # Compute the entropy of the training tokens
416 for input in task.batches(split="train"):
417 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
418 token_probas = token_count / token_count.sum()
419 entropy = -torch.xlogy(token_probas, token_probas).sum()
420 train_set_perplexity = math.exp(entropy)
422 ##############################
424 # A bit of paranoia never hurts
429 for input in task.batches(split="train"):
430 assert input.dim() == 2 and input.dtype == torch.int64
432 train_examples[x.sum().item()] = x
434 nb_total, nb_collisions = 0, 0
435 for input in task.batches(split="test"):
436 assert input.dim() == 2 and input.dtype == torch.int64
439 y = train_examples.get(x.sum().item())
441 if x.size() == y.size() and (x - y).abs().sum() == 0:
447 f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
450 ##############################
452 if args.learning_rate_schedule == "cos":
453 learning_rate_schedule = {}
454 for n_epoch in range(args.nb_epochs):
455 u = n_epoch / args.nb_epochs * math.pi
456 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
461 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
465 learning_rate_schedule = {}
466 learning_rate = args.learning_rate
467 for n_epoch in range(args.nb_epochs):
469 learning_rate = u[n_epoch]
470 learning_rate_schedule[n_epoch] = learning_rate
472 log_string(f"learning_rate_schedule {learning_rate_schedule}")
474 ##############################
478 if nb_epochs_finished >= nb_epochs:
479 task.produce_results(
484 args.deterministic_synthesis,
487 for n_epoch in range(nb_epochs_finished, nb_epochs):
488 learning_rate = learning_rate_schedule[n_epoch]
490 log_string(f"learning_rate {learning_rate}")
492 if args.optim == "sgd":
493 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
494 elif args.optim == "adam":
495 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
496 elif args.optim == "adamw":
497 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
499 raise ValueError(f"Unknown optimizer {args.optim}.")
503 nb_train_samples, acc_train_loss = 0, 0.0
505 for input in task.batches(split="train"):
506 input = input.to(device)
507 output = model(mygpt.BracketedSequence(input)).x
508 loss = F.cross_entropy(output.transpose(1, 2), input)
509 acc_train_loss += loss.item() * input.size(0)
510 nb_train_samples += input.size(0)
511 nb_samples_seen += input.size(0)
513 optimizer.zero_grad()
517 with torch.autograd.no_grad():
520 nb_test_samples, acc_test_loss = 0, 0.0
522 for input in task.batches(split="test"):
523 input = input.to(device)
525 output = model(mygpt.BracketedSequence(input)).x
526 loss = F.cross_entropy(output.transpose(1, 2), input)
527 acc_test_loss += loss.item() * input.size(0)
528 nb_test_samples += input.size(0)
530 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
531 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
534 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
537 task.produce_results(
538 n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
542 "nb_epochs_finished": n_epoch + 1,
543 "model_state": model.state_dict(),
544 "rng_state": torch.get_rng_state(),
547 if torch.cuda.is_available():
548 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
550 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
551 torch.save(checkpoint, checkpoint_name)
552 log_string(f"saved checkpoint {checkpoint_name}")
554 ######################################################################