from torch.nn import functional as F
import ffutils
-import mygpt
-import sky, reasoning, quiz_machine
-
-# world quizzes vs. culture quizzes
-######################################################################
+import mygpt
+import sky, grids, quiz_machine
-nb_new_c_quizzes_for_train = 1000
-nb_new_c_quizzes_for_test = 100
+import threading
-######################################################################
+import torch.multiprocessing as mp
-if torch.cuda.is_available():
- device = torch.device("cuda")
- torch.backends.cuda.matmul.allow_tf32 = True
-else:
- device = torch.device("cpu")
+# world quizzes vs. culture quizzes
######################################################################
parser = argparse.ArgumentParser(
- description="An implementation of GPT with cache.",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument("--seed", type=int, default=0)
-parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
+parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
########################################
parser.add_argument("--nb_test_samples", type=int, default=None)
-parser.add_argument("--learning_rate", type=float, default=1e-3)
+parser.add_argument("--learning_rate", type=float, default=5e-4)
########################################
parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
-parser.add_argument("--problem", type=str, default="sky")
+parser.add_argument("--problem", type=str, default="grids")
+
+parser.add_argument("--nb_threads", type=int, default=1)
+
+parser.add_argument("--gpus", type=str, default="all")
parser.add_argument("--nb_gpts", type=int, default=5)
parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
-parser.add_argument("--generation_temperature", type=float, default=2.0)
+parser.add_argument("--proba_understands", type=float, default=0.99)
-parser.add_argument("--deterministic_validation", action="store_true", default=False)
+parser.add_argument("--proba_not_understands", type=float, default=0.5)
-parser.add_argument("--bidirectional_validation", action="store_true", default=False)
+parser.add_argument("--generation_temperature", type=float, default=2.0)
parser.add_argument("--dirty_debug", action="store_true", default=False)
######################################################################
+grids_tasks = ", ".join(
+ [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
+)
+
+parser.add_argument(
+ "--grids_tasks",
+ type=str,
+ default=None,
+ help="A comma-separated subset of: " + grids_tasks + ", or None for all.",
+)
+
+######################################################################
+
parser.add_argument("--sky_height", type=int, default=6)
parser.add_argument("--sky_width", type=int, default=8)
######################################################################
-if args.dirty_debug:
- args.accuracy_to_make_c_quizzes = 0.0
- nb_new_c_quizzes_for_train = 100
- nb_new_c_quizzes_for_test = 10
-
-######################################################################
-
default_args = {
"model": "37M",
- "batch_size": 100,
+ "batch_size": 25,
"nb_train_samples": 100000,
"nb_test_samples": 10000,
}
######################################################################
+if args.gpus == "all":
+ gpus_idx = range(torch.cuda.device_count())
+else:
+ gpus_idx = [int(k) for k in args.gpus.split(",")]
+
+gpus = [torch.device(f"cuda:{n}") for n in gpus_idx]
+
+if torch.cuda.is_available():
+ main_device = gpus[0]
+else:
+ assert len(gpus) == 0
+ main_device = torch.device("cpu")
+
if args.dirty_debug:
args.nb_train_samples = 2500
args.nb_test_samples = 100
nb_birds=args.sky_nb_birds,
nb_iterations=args.sky_nb_iterations,
speed=args.sky_speed,
+ max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
+ chunk_size=100,
+ nb_threads=args.nb_threads,
)
back_accuracy = False
-elif args.problem == "reasoning":
- problem = reasoning.Reasoning(device=device)
+elif args.problem == "grids":
+ problem = grids.Grids(
+ max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
+ chunk_size=100,
+ nb_threads=args.nb_threads,
+ tasks=args.grids_tasks,
+ )
back_accuracy = True
else:
raise ValueError
+problem.save_some_examples(args.result_dir)
+
quiz_machine = quiz_machine.QuizMachine(
problem=problem,
nb_train_samples=args.nb_train_samples,
batch_size=args.physical_batch_size,
result_dir=args.result_dir,
logger=log_string,
- device=device,
+ device=main_device,
)
######################################################################
-log_string(f"device {device}")
+log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
vocabulary_size = quiz_machine.vocabulary_size()
######################################################################
-# Compute the entropy of the training tokens
-token_count = 0
-for input in quiz_machine.batches(split="train", desc="train-entropy"):
- token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
- (0, 1)
- )
-token_probas = token_count / token_count.sum()
-entropy = -torch.xlogy(token_probas, token_probas).sum()
-train_set_perplexity = math.exp(entropy)
+def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
+ with torch.autograd.no_grad():
+ model.eval().to(local_device)
-######################################################################
-# A bit of paranoia never hurts
+ nb_test_samples, acc_test_loss = 0, 0.0
+ nb_samples_accumulated = 0
-if args.max_percents_of_test_in_train >= 0:
+ for input in quiz_machine.batches(model, split="test"):
+ input = input.to(local_device)
- def subsets_as_tuples(batches, cs):
- s = set()
- for batch in batches:
- for x in batch:
- s.add(tuple([v.item() for v in x]))
- if len(s) == cs:
- yield s
- s = set()
- yield s
+ bs = model(mygpt.BracketedSequence(input))
+ output = bs.x
- nb_test, nb_in_train = 0, 0
- for test_subset in subsets_as_tuples(
- quiz_machine.batches(split="test", desc="test-check"), 25000
- ):
- in_train = set()
- for train_subset in subsets_as_tuples(
- quiz_machine.batches(split="train", desc="train-check"), 25000
- ):
- in_train.update(test_subset.intersection(train_subset))
- nb_in_train += len(in_train)
- nb_test += len(test_subset)
+ loss = F.cross_entropy(output.transpose(1, 2), input)
- log_string(
- f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
- )
+ acc_test_loss += loss.item() * input.size(0)
- assert (
- nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
- ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
+ nb_test_samples += input.size(0)
-##############################
+ test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+ log_string(f"test_perplexity {n_epoch} {test_perplexity}")
-def one_epoch(model, quiz_machine):
+ model.main_test_accuracy = quiz_machine.produce_results(
+ n_epoch=n_epoch,
+ model=model,
+ result_dir=args.result_dir,
+ deterministic_synthesis=deterministic_synthesis,
+ )
+
+
+def one_epoch(model, quiz_machine, local_device=main_device):
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
- model.train()
+ model.to(local_device).train()
nb_train_samples, acc_train_loss = 0, 0.0
- for input in quiz_machine.batches(split="train"):
- input = input.to(device)
+ for input in quiz_machine.batches(model, split="train"):
+ input = input.to(local_device)
if nb_train_samples % args.batch_size == 0:
optimizer.zero_grad()
train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
- log_string(f"train_perplexity {n_epoch} {train_perplexity}")
-
-
-######################################################################
-
-
-def run_tests(model, quiz_machine, deterministic_synthesis):
- with torch.autograd.no_grad():
- model.eval()
-
- nb_test_samples, acc_test_loss = 0, 0.0
- nb_samples_accumulated = 0
-
- for input in quiz_machine.batches(split="test"):
- input = input.to(device)
-
- bs = model(mygpt.BracketedSequence(input))
- output = bs.x
-
- loss = F.cross_entropy(output.transpose(1, 2), input)
+ log_string(f"train_perplexity {n_epoch} model.id {model.id} {train_perplexity}")
- acc_test_loss += loss.item() * input.size(0)
+ run_tests(model, quiz_machine, deterministic_synthesis=False)
- nb_test_samples += input.size(0)
- test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
-
- log_string(f"test_perplexity {n_epoch} {test_perplexity}")
-
- model.main_test_accuracy = quiz_machine.produce_results(
- n_epoch=n_epoch,
- model=model,
- result_dir=args.result_dir,
- deterministic_synthesis=deterministic_synthesis,
- )
+######################################################################
-######################################################################
+def standard_validity(logproba):
+ l = logproba.sort(dim=-1).values
+ return (l[:, 0] < math.log(args.proba_not_understands)) & (
+ l[:, 1] > math.log(args.proba_understands)
+ )
def valid_c_quizzes(recorded, criteria):
- result = [q[criteria(c)] for q, c in recorded]
+ result = [q[criteria(lp)] for q, lp in recorded]
return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
nb_for_train=1000,
nb_for_test=100,
):
- quizzes_and_nb_correct_records = []
+ quizzes_and_logproba_records = []
nb_to_create = nb_for_train + nb_for_test
# ------------------------------------------------------------
- standard_validity = lambda nb_correct: torch.logical_and(
- nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
- )
-
file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
+
with open(file_name, "w") as logp_file:
while (
- valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
+ valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
< nb_to_create
):
# Select a model at random to generate the new quizzes
temperature=args.generation_temperature,
)
- nb_correct, seq_logproba = quiz_machine.compute_correctness(
- c_quizzes,
- models,
- bidirectional_validation=args.bidirectional_validation,
- deterministic_validation=args.deterministic_validation,
- )
-
- for n, l in zip(nb_correct, seq_logproba):
- s = " ".join([str(x.item()) for x in l])
- logp_file.write(f"{n} {s}\n")
+ c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
- if args.dirty_debug:
- nb_correct = torch.randint(
- len(models) + 1, nb_correct.size(), device=c_quizzes.device
- )
-
- quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
-
- nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
- nv = " ".join([str(x.item()) for x in nv])
+ if c_quizzes.size(0) > 0:
+ logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
+ for l in logproba:
+ s = " ".join([str(x.item()) for x in l])
+ logp_file.write(s + "\n")
+ quizzes_and_logproba_records.append((c_quizzes, logproba))
nb_validated = valid_c_quizzes(
- quizzes_and_nb_correct_records, standard_validity
+ quizzes_and_logproba_records, standard_validity
).size(0)
log_string(
- f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
+ f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
)
# store the new c_quizzes which have been validated
- new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
+ new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
quiz_machine.reverse_random_half_in_place(new_c_quizzes)
# save a bunch of images to investigate what quizzes with a
# certain nb of correct predictions look like
- for n in range(len(models) + 1):
- s = (
- "_validated"
- if n >= args.min_to_validate and n <= args.max_to_validate
- else ""
- )
-
- q = valid_c_quizzes(
- quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
- )[:72]
+ q = new_c_quizzes[:72]
- if q.size(0) > 0:
- quiz_machine.save_quizzes(
- args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
- )
+ if q.size(0) > 0:
+ quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
######################################################################
models = []
for k in range(args.nb_gpts):
+ log_string(f"creating model {k} and its w_quizzes")
model = mygpt.MyGPT(
vocabulary_size=vocabulary_size,
dim_model=args.dim_model,
nb_blocks=args.nb_blocks,
causal=True,
dropout=args.dropout,
- ).to(device)
+ ).to(main_device)
model.main_test_accuracy = 0.0
model.id = k
+ model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
+ quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
+ model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
+ quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
+
models.append(model)
nb_parameters = sum(p.numel() for p in models[0].parameters())
log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
+######################################################################
+
+# Compute the entropy of the training tokens
+
+token_count = 0
+for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
+ token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
+ (0, 1)
+ )
+token_probas = token_count / token_count.sum()
+entropy = -torch.xlogy(token_probas, token_probas).sum()
+train_set_perplexity = math.exp(entropy)
+
+######################################################################
+# A bit of paranoia never hurts
+
+if args.max_percents_of_test_in_train >= 0:
+
+ def subsets_as_tuples(batches, cs):
+ s = set()
+ for batch in batches:
+ for x in batch:
+ s.add(tuple([v.item() for v in x]))
+ if len(s) == cs:
+ yield s
+ s = set()
+ yield s
+
+ nb_test, nb_in_train = 0, 0
+ for test_subset in subsets_as_tuples(
+ quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
+ ):
+ in_train = set()
+ for train_subset in subsets_as_tuples(
+ quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
+ ):
+ in_train.update(test_subset.intersection(train_subset))
+ nb_in_train += len(in_train)
+ nb_test += len(test_subset)
+
+ log_string(
+ f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
+ )
+
+ assert (
+ nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
+ ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
+
+######################################################################
+
+nb_new_c_quizzes_for_train = args.nb_train_samples // 50
+nb_new_c_quizzes_for_test = args.nb_test_samples // 50
+
+log_string(
+ f"nb_new_c_quizzes_for_train {nb_new_c_quizzes_for_train} nb_new_c_quizzes_for_test {nb_new_c_quizzes_for_test}"
+)
+
+######################################################################
+
+if args.dirty_debug:
+ args.accuracy_to_make_c_quizzes = 0.0
+ args.nb_gpts = 2
+ nb_new_c_quizzes_for_train = 100
+ nb_new_c_quizzes_for_test = 10
+
+ def standard_validity(logproba):
+ l = logproba.sort(dim=-1).values
+ return l[:, 0] < math.log(0.5)
+
+
######################################################################
for n_epoch in range(args.nb_epochs):
cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
log_string(f"current_test_accuracies {cta}")
+ ##################################################
# Select, improve, and eval the worst model
- weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
+ ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
- log_string(
- f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
- )
+ weakest_models = ranked_models[: len(gpus)]
- one_epoch(weakest_model, quiz_machine)
+ threads = []
- log_string(
- f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
- )
+ for gpu, model in zip(gpus, weakest_models):
+ log_string(f"training model {model.id}")
+
+ t = threading.Thread(
+ target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
+ )
+
+ threads.append(t)
- run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)
+ t.start()
+
+ for t in threads:
+ t.join()
+
+ ##################################################
+ # Replace a fraction of the w_quizzes with fresh ones
log_string(
- f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
+ f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
)
- # Replace a fraction of the w_quizzes with fresh ones
+ # Renew entirely the train set
- quiz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+ for model in weakest_models:
+ quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
+ ##################################################
# If all the models are good enough, generate new quizzes and
# re-compute the test errors
nb_for_test=nb_new_c_quizzes_for_test,
)
- for model in models:
- run_tests(model, quiz_machine, deterministic_synthesis=False)
-
-
######################################################################