import threading
-# world quizzes vs. culture quizzes
-
-######################################################################
-
-if torch.cuda.is_available():
- device = torch.device("cuda")
- torch.backends.cuda.matmul.allow_tf32 = True
-else:
- device = torch.device("cpu")
+import torch.multiprocessing as mp
######################################################################
parser.add_argument("--nb_threads", type=int, default=1)
-parser.add_argument("--nb_gpus", 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.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=args.nb_gpus * args.nb_train_samples // 100,
+ 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 == "grids":
problem = grids.Grids(
- max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
+ 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()
######################################################################
-######################################################################
-
-
-def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
- if local_device is None:
- local_device = device
-
+def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
with torch.autograd.no_grad():
model.eval().to(local_device)
test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
- log_string(f"test_perplexity {n_epoch} {test_perplexity}")
+ log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
model.main_test_accuracy = quiz_machine.produce_results(
n_epoch=n_epoch,
)
-def one_epoch(model, quiz_machine, local_device=None):
- if local_device is None:
- local_device = device
-
+def one_epoch(model, quiz_machine, local_device=main_device):
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
model.to(local_device).train()
train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
- log_string(f"train_perplexity {n_epoch} {train_perplexity}")
+ log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
run_tests(model, quiz_machine, deterministic_synthesis=False)
- model.TRAINING_LOCK.release()
+ model.to(main_device)
######################################################################
def standard_validity(logproba):
l = logproba.sort(dim=-1).values
- return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
- # warnings.warn("TEST!!!", RuntimeWarning)
- # print(l.exp())
- # return (l[:, 0] < math.log(0.99))
+ return (l[:, 0] < math.log(args.proba_not_understands)) & (
+ l[:, 1] > math.log(args.proba_understands)
+ )
def valid_c_quizzes(recorded, criteria):
quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
-######################################################################
-
-
-def create_c_quizzes_(
- models,
- quiz_machine,
- nb_for_train=1000,
- nb_for_test=100,
-):
- quizzes_and_nb_correct_records = []
-
- nb_to_create = nb_for_train + nb_for_test
-
- # ------------------------------------------------------------
-
- standard_validity = lambda nb_correct: (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)
- < nb_to_create
- ):
- # Select a model at random to generate the new quizzes
-
- model_for_generation = models[torch.randint(len(models), (1,))]
-
- c_quizzes = quiz_machine.generate_quizzes(
- nb_to_create,
- model_for_generation=model_for_generation,
- temperature=args.generation_temperature,
- )
-
- # if args.prediction_correctness:
-
- # else:
- # logproba = quiz_machine.new(quiz_machine.size(0), len(models))
- # for q,l in zip(quizzes.split(args.batch_size), logits.split(args.batch_size)):
- # for model in models:
- # l[...] = F.cross_entropy(model(q))
-
- c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
-
- if c_quizzes.size(0) > 0:
- 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")
-
- 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])
-
- nb_validated = valid_c_quizzes(
- quizzes_and_nb_correct_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}"
- )
-
- # store the new c_quizzes which have been validated
-
- new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
-
- quiz_machine.reverse_random_half_in_place(new_c_quizzes)
-
- quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
- quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
-
- # 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]
-
- quiz_machine.reverse_random_half_in_place(q)
-
- if q.size(0) > 0:
- quiz_machine.save_quizzes(
- args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
- )
-
-
######################################################################
models = []
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.TRAINING_LOCK = threading.Lock()
- model.train_w_quizzes = quiz_machine.generate_token_sequences(
- args.nb_train_samples
- ).to(device)
+ 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
- ).to(device)
+ 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_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):
ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
- weakest_models = ranked_models[: args.nb_gpus]
+ weakest_models = ranked_models[: len(gpus)]
- for gpu_id, model in enumerate(weakest_models):
- model.TRAINING_LOCK.acquire()
+ threads = []
- log_string(
- f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
+ 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)
)
- threading.Thread(
- target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
- ).start()
+ threads.append(t)
- for model in weakest_models:
- model.TRAINING_LOCK.acquire()
- model.TRAINING_LOCK.release()
+ t.start()
+
+ for t in threads:
+ t.join()
##################################################
# Replace a fraction of the w_quizzes with fresh ones