import ffutils
import mygpt
-import sky, quizz_machine
+import sky, wireworld, quizz_machine
# world quizzes vs. culture quizzes
######################################################################
-accuracy_to_make_c_quizzes = 0.975
nb_new_c_quizzes_for_train = 1000
nb_new_c_quizzes_for_test = 100
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
-parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
+parser.add_argument("--log_filename", type=str, default="train.log")
parser.add_argument("--result_dir", type=str, default=None)
parser.add_argument("--nb_test_samples", type=int, default=None)
-parser.add_argument("--learning_rate", type=float, default=1e-4)
+parser.add_argument("--learning_rate", type=float, default=1e-3)
########################################
parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
+parser.add_argument("--problem", type=str, default="sky")
+
parser.add_argument("--nb_gpts", type=int, default=5)
+parser.add_argument("--nb_models_for_generation", type=int, default=1)
+
+parser.add_argument("--generation_mode", type=str, default="groupthink")
+
+parser.add_argument("--min_to_validate", type=int, default=4)
+
+parser.add_argument("--max_to_validate", type=int, default=4)
+
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
+
parser.add_argument("--dirty_debug", action="store_true", default=False)
######################################################################
######################################################################
if args.dirty_debug:
- accuracy_to_make_c_quizzes = 0.0
+ 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,
- "nb_train_samples": 250000,
+ "nb_train_samples": 100000,
"nb_test_samples": 10000,
}
assert args.nb_train_samples % args.batch_size == 0
assert args.nb_test_samples % args.batch_size == 0
+if args.problem=="sky":
+ problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2, speed=2),
+elif args.problem="wireworld":
+ problem=wireworld.Wireworld(height=10, width=15, nb_iterations=4)
+else:
+ raise ValueError
+
quizz_machine = quizz_machine.QuizzMachine(
- sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
+ problem=problem,
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
batch_size=args.physical_batch_size,
n_epoch=n_epoch,
model=model,
result_dir=args.result_dir,
- logger=log_string,
deterministic_synthesis=deterministic_synthesis,
)
def create_c_quizzes(
- model,
- other_models,
+ models,
quizz_machine,
nb_for_train=1000,
nb_for_test=100,
min_ave_seq_logproba=None,
):
- kept = []
+ # We will store the generated quizzes for each number of
+ # correct prediction
+ recorded = dict([(n, []) for n in range(len(models) + 1)])
+ model_indexes = []
sum_logits, sum_nb_c_quizzes = 0, 0
- while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
- nb_to_generate = 4 * (nb_for_train + nb_for_test)
+ def nb_generated():
+ return sum([sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()])
+
+ def nb_validated():
+ return sum(
+ [
+ sum([x.size(0) for x in recorded[n]])
+ for n in range(args.min_to_validate, args.max_to_validate + 1)
+ ]
+ )
- new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
+ nb_to_create = nb_for_train + nb_for_test
+
+ while nb_validated() < nb_to_create:
+ (
+ new_c_quizzes,
+ nb_correct,
+ ave_seq_logproba,
+ ) = quizz_machine.gang_create_c_quizzes(
+ nb=nb_to_create,
+ nb_models_for_generation=args.nb_models_for_generation,
+ models=models,
+ mode=args.generation_mode,
+ min_ave_seq_logproba=min_ave_seq_logproba,
n_epoch=n_epoch,
result_dir=args.result_dir,
- logger=log_string,
- nb=nb_to_generate,
- model=model,
- other_models=other_models,
- min_ave_seq_logproba=min_ave_seq_logproba,
)
sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
sum_nb_c_quizzes += new_c_quizzes.size(0)
- to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
-
if args.dirty_debug:
- to_keep = new_c_quizzes
+ nb_correct = torch.randint(
+ len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
+ )
+
+ for n in range(nb_correct.max() + 1):
+ recorded[n].append(new_c_quizzes[nb_correct == n].clone())
log_string(
- f"keep {to_keep.size(0)}/{new_c_quizzes.size(0)} c_quizzes ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%)"
+ f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_create}"
)
- kept.append(to_keep)
+ # concatenate and shuffle
+ for n in recorded.keys():
+ if len(recorded[n]) > 0:
+ q = torch.cat(recorded[n], dim=0)
+ q = q[torch.randperm(q.size(0), device=q.device)]
+ recorded[n] = q
+ else:
+ del recorded[n]
+
+ new_c_quizzes = torch.cat(
+ [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)],
+ dim=0,
+ )
- new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
+ new_c_quizzes = new_c_quizzes[
+ torch.randperm(new_c_quizzes.size(0), device=new_c_quizzes.device)[
+ : nb_for_train + nb_for_test
+ ]
+ ]
quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
- quizz_machine.problem.save_quizzes(
- new_c_quizzes[:72],
- args.result_dir,
- f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
- )
+ for n in recorded.keys():
+ s = (
+ "_validated"
+ if n >= args.min_to_validate and n <= args.max_to_validate
+ else ""
+ )
+ quizz_machine.problem.save_quizzes(
+ recorded[n][:72],
+ args.result_dir,
+ f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
+ )
return sum_logits / sum_nb_c_quizzes
a = [(model.id, float(model.main_test_accuracy)) for model in models]
a.sort(key=lambda p: p[0])
- log_string(f"current accuracies {a}")
+ s = " ".join([f"{p[1]*100:.02f}%" for p in a])
+ log_string(f"current accuracies {s}")
# select the model with lowest accuracy
models.sort(key=lambda model: model.main_test_accuracy)
f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
)
- if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
- other_models = models.copy()
- other_models.remove(model)
-
+ if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
ave_seq_logproba = create_c_quizzes(
- model,
- other_models,
+ models,
quizz_machine,
nb_for_train=nb_new_c_quizzes_for_train,
nb_for_test=nb_new_c_quizzes_for_test,
)
# We keep the first average logits as a reference
- if min_ave_seq_logproba is None:
- min_ave_seq_logproba = ave_seq_logproba
- else:
- log_string(
- f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
- )
+ # if min_ave_seq_logproba is None:
+ # min_ave_seq_logproba = ave_seq_logproba
+ # else:
+ # log_string(
+ # f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
+ # )
# We update everyone
for model in models: