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
20 ######################################################################
22 if torch.cuda.is_available():
23 device = torch.device("cuda")
24 torch.backends.cuda.matmul.allow_tf32 = True
26 device = torch.device("cpu")
28 ######################################################################
30 parser = argparse.ArgumentParser(
31 description="An implementation of GPT with cache.",
32 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
39 help="sandbox, picoclvr, mnist, maze, snake, stack, expr, world",
42 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
44 parser.add_argument("--result_dir", type=str, default=None)
46 parser.add_argument("--seed", type=int, default=0)
48 parser.add_argument("--nb_epochs", type=int, default=None)
50 parser.add_argument("--batch_size", type=int, default=None)
52 parser.add_argument("--nb_train_samples", type=int, default=None)
54 parser.add_argument("--nb_test_samples", type=int, default=None)
56 parser.add_argument("--optim", type=str, default="adam")
58 parser.add_argument("--learning_rate", type=float, default=1e-4)
60 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
62 parser.add_argument("--dim_model", type=int, default=512)
64 parser.add_argument("--dim_keys", type=int, default=64)
66 parser.add_argument("--dim_hidden", type=int, default=2048)
68 parser.add_argument("--nb_heads", type=int, default=8)
70 parser.add_argument("--nb_blocks", type=int, default=12)
72 parser.add_argument("--dropout", type=float, default=0.1)
74 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
76 parser.add_argument("--no_checkpoint", action="store_true", default=False)
78 parser.add_argument("--overwrite_results", action="store_true", default=False)
80 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
82 ##############################
85 parser.add_argument("--sandbox_level", type=int, default=0)
87 parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
89 parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
91 parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
93 ##############################
96 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
98 parser.add_argument("--picoclvr_height", type=int, default=12)
100 parser.add_argument("--picoclvr_width", type=int, default=16)
102 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
104 ##############################
107 parser.add_argument("--maze_height", type=int, default=23)
109 parser.add_argument("--maze_width", type=int, default=39)
111 parser.add_argument("--maze_nb_walls", type=int, default=45)
113 ##############################
116 parser.add_argument("--snake_height", type=int, default=6)
118 parser.add_argument("--snake_width", type=int, default=8)
120 parser.add_argument("--snake_nb_colors", type=int, default=5)
122 parser.add_argument("--snake_length", type=int, default=200)
124 ##############################
127 parser.add_argument("--stack_nb_steps", type=int, default=100)
129 parser.add_argument("--stack_nb_stacks", type=int, default=3)
131 parser.add_argument("--stack_nb_digits", type=int, default=3)
133 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
135 ##############################
138 parser.add_argument("--expr_nb_variables", type=int, default=5)
140 parser.add_argument("--expr_sequence_length", type=int, default=40)
142 parser.add_argument("--expr_operand_max", type=int, default=9)
144 parser.add_argument("--expr_result_max", type=int, default=99)
146 parser.add_argument("--expr_input_file", type=str, default=None)
148 ##############################
151 parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
153 ######################################################################
155 args = parser.parse_args()
157 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
159 if args.result_dir is None:
160 args.result_dir = f"results_{args.task}"
162 ######################################################################
168 "nb_train_samples": 100000,
169 "nb_test_samples": 10000,
174 "nb_train_samples": 250000,
175 "nb_test_samples": 10000,
180 "nb_train_samples": 250000,
181 "nb_test_samples": 10000,
186 "nb_train_samples": 250000,
187 "nb_test_samples": 10000,
192 "nb_train_samples": 250000,
193 "nb_test_samples": 10000,
198 "nb_train_samples": 100000,
199 "nb_test_samples": 1000,
204 "nb_train_samples": 1000000,
205 "nb_test_samples": 10000,
210 "nb_train_samples": 25000,
211 "nb_test_samples": 1000,
215 if args.task in default_args:
216 for k, v in default_args[args.task].items():
217 if getattr(args, k) is None:
220 ######################################################################
223 os.mkdir(args.result_dir)
224 except FileExistsError:
225 if not args.overwrite_results:
226 print(f"result directory {args.result_dir} already exists")
229 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
232 # torch.backends.cudnn.deterministic = True
233 # torch.backends.cudnn.benchmark = False
234 # torch.use_deterministic_algorithms(True)
235 torch.manual_seed(args.seed)
236 if torch.cuda.is_available():
237 torch.cuda.manual_seed_all(args.seed)
239 ######################################################################
243 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
245 if log_file is not None:
246 log_file.write(t + s + "\n")
254 log_string(f"args.{n} {getattr(args, n)}")
257 ######################################################################
260 def picoclvr_pruner_horizontal_green(p):
261 return not ("green" in p and ("left" in p or "right" in p))
264 picoclvr_pruner_train = (
265 picoclvr_pruner_horizontal_green
266 if args.picocvlr_prune_properties in {"train+eval"}
270 picoclvr_pruner_eval = (
271 (lambda p: not picoclvr_pruner_horizontal_green(p))
272 if args.picocvlr_prune_properties in {"train+eval", "eval"}
276 ######################################################################
278 if args.task == "sandbox":
279 if args.sandbox_level == 0:
280 problem = tasks.ProblemLevel0(
281 nb_sentences=args.sandbox_levels_nb_items,
282 len_prompt=args.sandbox_levels_len_source,
283 len_result=args.sandbox_levels_len_result,
285 elif args.sandbox_level == 1:
286 problem = tasks.ProblemLevel1(
287 nb_operators=args.sandbox_levels_nb_items,
288 len_source=args.sandbox_levels_len_source,
289 len_result=args.sandbox_levels_len_result,
291 elif args.sandbox_level == 2:
292 problem = tasks.ProblemLevel2(
293 len_source=args.sandbox_levels_len_source,
294 len_result=args.sandbox_levels_len_result,
297 raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
299 task = tasks.SandBox(
301 # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
302 nb_train_samples=args.nb_train_samples,
303 nb_test_samples=args.nb_test_samples,
304 batch_size=args.batch_size,
309 elif args.task == "picoclvr":
310 task = tasks.PicoCLVR(
311 nb_train_samples=args.nb_train_samples,
312 nb_test_samples=args.nb_test_samples,
313 batch_size=args.batch_size,
314 height=args.picoclvr_height,
315 width=args.picoclvr_width,
316 nb_colors=args.picoclvr_nb_colors,
319 pruner_train=picoclvr_pruner_train,
320 pruner_eval=picoclvr_pruner_eval,
323 elif args.task == "mnist":
325 nb_train_samples=args.nb_train_samples,
326 nb_test_samples=args.nb_test_samples,
327 batch_size=args.batch_size,
331 elif args.task == "maze":
333 nb_train_samples=args.nb_train_samples,
334 nb_test_samples=args.nb_test_samples,
335 batch_size=args.batch_size,
336 height=args.maze_height,
337 width=args.maze_width,
338 nb_walls=args.maze_nb_walls,
342 elif args.task == "snake":
344 nb_train_samples=args.nb_train_samples,
345 nb_test_samples=args.nb_test_samples,
346 batch_size=args.batch_size,
347 height=args.snake_height,
348 width=args.snake_width,
349 nb_colors=args.snake_nb_colors,
350 length=args.snake_length,
351 prompt_length=args.snake_length // 2,
355 elif args.task == "stack":
357 nb_train_samples=args.nb_train_samples,
358 nb_test_samples=args.nb_test_samples,
359 batch_size=args.batch_size,
361 nb_steps=args.stack_nb_steps,
362 nb_stacks=args.stack_nb_stacks,
363 nb_digits=args.stack_nb_digits,
364 fraction_values_for_train=args.stack_fraction_values_for_train,
368 elif args.task == "expr":
370 nb_train_samples=args.nb_train_samples,
371 nb_test_samples=args.nb_test_samples,
372 nb_variables=args.expr_nb_variables,
373 sequence_length=args.expr_sequence_length,
374 operand_max=args.expr_operand_max,
375 result_max=args.expr_result_max,
376 batch_size=args.batch_size,
380 elif args.task == "world":
382 nb_train_samples=args.nb_train_samples,
383 nb_test_samples=args.nb_test_samples,
384 batch_size=args.batch_size,
385 vqae_nb_epochs=args.world_vqae_nb_epochs,
391 raise ValueError(f"Unknown task {args.task}")
393 ######################################################################
395 log_string(f"device {device}")
397 vocabulary_size = task.vocabulary_size()
399 log_string(f"vocabulary_size {vocabulary_size}")
401 ##############################
404 vocabulary_size=vocabulary_size,
405 dim_model=args.dim_model,
406 dim_keys=args.dim_keys,
407 dim_hidden=args.dim_hidden,
408 nb_heads=args.nb_heads,
409 nb_blocks=args.nb_blocks,
411 dropout=args.dropout,
416 nb_parameters = sum(p.numel() for p in model.parameters())
417 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
419 ######################################################################
421 nb_epochs_finished = 0
423 if args.no_checkpoint:
424 log_string(f"not trying to load checkpoint.")
428 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
429 checkpoint = torch.load(checkpoint_name)
430 nb_epochs_finished = checkpoint["nb_epochs_finished"]
431 model.load_state_dict(checkpoint["model_state"])
432 torch.set_rng_state(checkpoint["rng_state"])
433 if torch.cuda.is_available():
434 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
436 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
438 except FileNotFoundError:
439 log_string("starting from scratch.")
442 log_string("error when loading the checkpoint.")
445 ######################################################################
447 if args.task == "expr" and args.expr_input_file is not None:
448 task.produce_results(
453 args.deterministic_synthesis,
454 args.expr_input_file,
459 ######################################################################
461 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
463 # Compute the entropy of the training tokens
466 for input in task.batches(split="train"):
467 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
468 token_probas = token_count / token_count.sum()
469 entropy = -torch.xlogy(token_probas, token_probas).sum()
470 train_set_perplexity = math.exp(entropy)
472 ##############################
474 # A bit of paranoia never hurts
479 for input in task.batches(split="train"):
480 assert input.dim() == 2 and input.dtype == torch.int64
482 train_examples[x.sum().item()] = x
484 nb_total, nb_collisions = 0, 0
485 for input in task.batches(split="test"):
486 assert input.dim() == 2 and input.dtype == torch.int64
489 y = train_examples.get(x.sum().item())
491 if x.size() == y.size() and (x - y).abs().sum() == 0:
497 f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
500 ##############################
502 if args.learning_rate_schedule == "cos":
503 learning_rate_schedule = {}
504 for n_epoch in range(args.nb_epochs):
505 u = n_epoch / args.nb_epochs * math.pi
506 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
511 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
515 learning_rate_schedule = {}
516 learning_rate = args.learning_rate
517 for n_epoch in range(args.nb_epochs):
519 learning_rate = u[n_epoch]
520 learning_rate_schedule[n_epoch] = learning_rate
522 log_string(f"learning_rate_schedule {learning_rate_schedule}")
524 ##############################
528 if nb_epochs_finished >= nb_epochs:
529 task.produce_results(
534 args.deterministic_synthesis,
537 for n_epoch in range(nb_epochs_finished, nb_epochs):
538 learning_rate = learning_rate_schedule[n_epoch]
540 log_string(f"learning_rate {learning_rate}")
542 if args.optim == "sgd":
543 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
544 elif args.optim == "adam":
545 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
546 elif args.optim == "adamw":
547 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
549 raise ValueError(f"Unknown optimizer {args.optim}.")
553 nb_train_samples, acc_train_loss = 0, 0.0
555 for input in task.batches(split="train"):
556 input = input.to(device)
557 output = model(mygpt.BracketedSequence(input)).x
558 loss = F.cross_entropy(output.transpose(1, 2), input)
559 acc_train_loss += loss.item() * input.size(0)
560 nb_train_samples += input.size(0)
561 nb_samples_seen += input.size(0)
563 optimizer.zero_grad()
567 with torch.autograd.no_grad():
570 nb_test_samples, acc_test_loss = 0, 0.0
572 for input in task.batches(split="test"):
573 input = input.to(device)
575 output = model(mygpt.BracketedSequence(input)).x
576 loss = F.cross_entropy(output.transpose(1, 2), input)
577 acc_test_loss += loss.item() * input.size(0)
578 nb_test_samples += input.size(0)
580 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
581 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
584 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
587 task.produce_results(
588 n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
592 "nb_epochs_finished": n_epoch + 1,
593 "model_state": model.state_dict(),
594 "rng_state": torch.get_rng_state(),
597 if torch.cuda.is_available():
598 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
600 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
601 torch.save(checkpoint, checkpoint_name)
602 log_string(f"saved checkpoint {checkpoint_name}")
604 ######################################################################