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",
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_input_file", type=str, default=None)
132 ######################################################################
134 args = parser.parse_args()
136 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
138 if args.result_dir is None:
139 args.result_dir = f"results_{args.task}"
141 ######################################################################
147 "nb_train_samples": 250000,
148 "nb_test_samples": 10000,
153 "nb_train_samples": 250000,
154 "nb_test_samples": 10000,
159 "nb_train_samples": 250000,
160 "nb_test_samples": 10000,
165 "nb_train_samples": 250000,
166 "nb_test_samples": 10000,
171 "nb_train_samples": 100000,
172 "nb_test_samples": 1000,
177 "nb_train_samples": 1000000,
178 "nb_test_samples": 10000,
182 if args.task in default_args:
183 for k, v in default_args[args.task].items():
184 if getattr(args, k) is None:
187 ######################################################################
190 os.mkdir(args.result_dir)
191 except FileExistsError:
192 if not args.overwrite_results:
193 print(f"result directory {args.result_dir} already exists")
196 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
199 # torch.backends.cudnn.deterministic = True
200 # torch.backends.cudnn.benchmark = False
201 # torch.use_deterministic_algorithms(True)
202 torch.manual_seed(args.seed)
203 if torch.cuda.is_available():
204 torch.cuda.manual_seed_all(args.seed)
206 ######################################################################
210 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
212 if log_file is not None:
213 log_file.write(t + s + "\n")
221 log_string(f"args.{n} {getattr(args, n)}")
224 ######################################################################
227 def picoclvr_pruner_horizontal_green(p):
228 return not ("green" in p and ("left" in p or "right" in p))
231 picoclvr_pruner_train = (
232 picoclvr_pruner_horizontal_green
233 if args.picocvlr_prune_properties in {"train+eval"}
237 picoclvr_pruner_eval = (
238 (lambda p: not picoclvr_pruner_horizontal_green(p))
239 if args.picocvlr_prune_properties in {"train+eval", "eval"}
243 ######################################################################
245 if args.task == "picoclvr":
246 task = tasks.PicoCLVR(
247 nb_train_samples=args.nb_train_samples,
248 nb_test_samples=args.nb_test_samples,
249 batch_size=args.batch_size,
250 height=args.picoclvr_height,
251 width=args.picoclvr_width,
252 nb_colors=args.picoclvr_nb_colors,
255 pruner_train=picoclvr_pruner_train,
256 pruner_eval=picoclvr_pruner_eval,
259 elif args.task == "mnist":
261 nb_train_samples=args.nb_train_samples,
262 nb_test_samples=args.nb_test_samples,
263 batch_size=args.batch_size,
267 elif args.task == "maze":
269 nb_train_samples=args.nb_train_samples,
270 nb_test_samples=args.nb_test_samples,
271 batch_size=args.batch_size,
272 height=args.maze_height,
273 width=args.maze_width,
274 nb_walls=args.maze_nb_walls,
278 elif args.task == "snake":
280 nb_train_samples=args.nb_train_samples,
281 nb_test_samples=args.nb_test_samples,
282 batch_size=args.batch_size,
283 height=args.snake_height,
284 width=args.snake_width,
285 nb_colors=args.snake_nb_colors,
286 length=args.snake_length,
287 prompt_length=args.snake_length // 2,
291 elif args.task == "stack":
293 nb_train_samples=args.nb_train_samples,
294 nb_test_samples=args.nb_test_samples,
295 batch_size=args.batch_size,
297 nb_steps=args.stack_nb_steps,
298 nb_stacks=args.stack_nb_stacks,
299 nb_digits=args.stack_nb_digits,
300 fraction_values_for_train=args.stack_fraction_values_for_train,
304 elif args.task == "expr":
306 nb_train_samples=args.nb_train_samples,
307 nb_test_samples=args.nb_test_samples,
308 nb_variables=args.expr_nb_variables,
309 sequence_length=args.expr_sequence_length,
310 batch_size=args.batch_size,
315 raise ValueError(f"Unknown task {args.task}")
317 ######################################################################
319 log_string(f"device {device}")
321 vocabulary_size = task.vocabulary_size()
323 log_string(f"vocabulary_size {vocabulary_size}")
325 ##############################
328 vocabulary_size=vocabulary_size,
329 dim_model=args.dim_model,
330 dim_keys=args.dim_keys,
331 dim_hidden=args.dim_hidden,
332 nb_heads=args.nb_heads,
333 nb_blocks=args.nb_blocks,
335 dropout=args.dropout,
340 nb_parameters = sum(p.numel() for p in model.parameters())
341 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
343 ######################################################################
345 nb_epochs_finished = 0
347 if args.no_checkpoint:
348 log_string(f"not trying to load checkpoint.")
352 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
353 checkpoint = torch.load(checkpoint_name)
354 nb_epochs_finished = checkpoint["nb_epochs_finished"]
355 model.load_state_dict(checkpoint["model_state"])
356 torch.set_rng_state(checkpoint["rng_state"])
357 if torch.cuda.is_available():
358 torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
360 log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
362 except FileNotFoundError:
363 log_string("starting from scratch.")
366 log_string("error when loading the checkpoint.")
369 ######################################################################
371 if args.task == "expr" and args.expr_input_file is not None:
372 task.produce_results(
377 args.deterministic_synthesis,
378 args.expr_input_file,
383 ######################################################################
385 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
387 # Compute the entropy of the training tokens
390 for input in task.batches(split="train"):
391 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
392 token_probas = token_count / token_count.sum()
393 entropy = -torch.xlogy(token_probas, token_probas).sum()
394 train_set_perplexity = math.exp(entropy)
396 ##############################
398 # A bit of paranoia never hurts
403 for input in task.batches(split="train"):
404 assert input.dim() == 2 and input.dtype == torch.int64
406 train_examples[x.sum().item()] = x
408 nb_total, nb_collisions = 0, 0
409 for input in task.batches(split="test"):
410 assert input.dim() == 2 and input.dtype == torch.int64
413 y = train_examples.get(x.sum().item())
415 if x.size() == y.size() and (x - y).abs().sum() == 0:
421 f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
424 ##############################
426 if args.learning_rate_schedule == "cos":
427 learning_rate_schedule = {}
428 for n_epoch in range(args.nb_epochs):
429 u = n_epoch / args.nb_epochs * math.pi
430 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
435 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
439 learning_rate_schedule = {}
440 learning_rate = args.learning_rate
441 for n_epoch in range(args.nb_epochs):
443 learning_rate = u[n_epoch]
444 learning_rate_schedule[n_epoch] = learning_rate
446 log_string(f"learning_rate_schedule {learning_rate_schedule}")
448 ##############################
452 if nb_epochs_finished >= nb_epochs:
453 task.produce_results(
458 args.deterministic_synthesis,
461 for n_epoch in range(nb_epochs_finished, nb_epochs):
462 learning_rate = learning_rate_schedule[n_epoch]
464 log_string(f"learning_rate {learning_rate}")
466 if args.optim == "sgd":
467 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
468 elif args.optim == "adam":
469 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
470 elif args.optim == "adamw":
471 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
473 raise ValueError(f"Unknown optimizer {args.optim}.")
477 nb_train_samples, acc_train_loss = 0, 0.0
479 for input in task.batches(split="train"):
480 input = input.to(device)
481 output = model(mygpt.BracketedSequence(input)).x
482 loss = F.cross_entropy(output.transpose(1, 2), input)
483 acc_train_loss += loss.item() * input.size(0)
484 nb_train_samples += input.size(0)
485 nb_samples_seen += input.size(0)
487 optimizer.zero_grad()
491 with torch.autograd.no_grad():
494 nb_test_samples, acc_test_loss = 0, 0.0
496 for input in task.batches(split="test"):
497 input = input.to(device)
499 output = model(mygpt.BracketedSequence(input)).x
500 loss = F.cross_entropy(output.transpose(1, 2), input)
501 acc_test_loss += loss.item() * input.size(0)
502 nb_test_samples += input.size(0)
504 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
505 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
508 f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
511 task.produce_results(
512 n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
516 "nb_epochs_finished": n_epoch + 1,
517 "model_state": model.state_dict(),
518 "rng_state": torch.get_rng_state(),
521 if torch.cuda.is_available():
522 checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
524 checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
525 torch.save(checkpoint, checkpoint_name)
526 log_string(f"saved checkpoint {checkpoint_name}")
528 ######################################################################