import ffutils
import mygpt
-import sky, wireworld, quizz_machine
+import sky, reasoning, quizz_machine
# world quizzes vs. culture quizzes
parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
-parser.add_argument("--reverse_cleanup", action="store_true", default=False)
-
-parser.add_argument("--validation_forward_only", 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("--min_to_validate", type=int, default=4)
+parser.add_argument("--min_to_validate", type=int, default=None)
-parser.add_argument("--max_to_validate", type=int, default=4)
+parser.add_argument("--max_to_validate", type=int, default=None)
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("--deterministic_validation", action="store_true", default=False)
+
+parser.add_argument("--bidirectional_validation", action="store_true", default=False)
+
parser.add_argument("--dirty_debug", action="store_true", default=False)
+######################################################################
+
parser.add_argument("--sky_height", type=int, default=6)
parser.add_argument("--sky_width", type=int, default=8)
args = parser.parse_args()
+if args.min_to_validate is None:
+ args.min_to_validate = args.nb_gpts - 1
+
+if args.max_to_validate is None:
+ args.max_to_validate = args.nb_gpts - 1
+
if args.result_dir is None:
args.result_dir = f"results_culture"
nb_iterations=args.sky_nb_iterations,
speed=args.sky_speed,
)
-elif args.problem == "wireworld":
- problem = wireworld.Wireworld(height=8, width=10, nb_iterations=2, speed=5)
+ back_accuracy = False
+elif args.problem == "reasoning":
+ problem = reasoning.Reasoning(device=device)
+ back_accuracy = True
else:
raise ValueError
problem=problem,
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
+ back_accuracy=back_accuracy,
batch_size=args.physical_batch_size,
result_dir=args.result_dir,
logger=log_string,
nb_test_samples += input.size(0)
- main_test_accuracy = quizz_machine.produce_results(
+ 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 = quizz_machine.produce_results(
n_epoch=n_epoch,
model=model,
result_dir=args.result_dir,
deterministic_synthesis=deterministic_synthesis,
)
- 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 = main_test_accuracy
-
######################################################################
):
recorded = []
- sum_logits, sum_nb_c_quizzes = 0, 0
-
nb_to_create = nb_for_train + nb_for_test
# ------------------------------------------------------------
nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
)
- while valid_c_quizzes(recorded, standard_validity).size(0) < nb_to_create:
- model_for_generation = models[torch.randint(len(models), (1,))]
-
- c_quizzes, ave_seq_logproba = quizz_machine.generate_quizzes(
- nb_to_create,
- model_for_generation=model_for_generation,
- reverse_cleanup=args.reverse_cleanup,
- )
+ 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(recorded, standard_validity).size(0) < nb_to_create:
+ # Select a model at random to generate the new quizzes
- sum_logits += c_quizzes.size(0) * ave_seq_logproba
- sum_nb_c_quizzes += c_quizzes.size(0)
+ model_for_generation = models[torch.randint(len(models), (1,))]
- nb_correct = quizz_machine.compute_correctness(
- c_quizzes, models, both_directions=not args.validation_forward_only
- )
+ c_quizzes = quizz_machine.generate_quizzes(
+ nb_to_create,
+ model_for_generation=model_for_generation,
+ temperature=args.generation_temperature,
+ )
- if args.dirty_debug:
- nb_correct = torch.randint(
- len(models) + 1, nb_correct.size(), device=c_quizzes.device
+ nb_correct, seq_logproba = quizz_machine.compute_correctness(
+ c_quizzes,
+ models,
+ bidirectional_validation=args.bidirectional_validation,
+ deterministic_validation=args.deterministic_validation,
)
- recorded.append((c_quizzes, nb_correct))
+ 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")
- nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
- nv = " ".join([str(x.item()) for x in nv])
+ if args.dirty_debug:
+ nb_correct = torch.randint(
+ len(models) + 1, nb_correct.size(), device=c_quizzes.device
+ )
- nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
+ recorded.append((c_quizzes, nb_correct))
- log_string(
- f"keep c_quizzes kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
- )
+ 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(recorded, 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
else ""
)
- quizz_machine.problem.save_quizzes(
- valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72],
- args.result_dir,
- f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
- )
+ q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72]
- return sum_logits / sum_nb_c_quizzes
+ if q.size(0) > 0:
+ quizz_machine.save_quizzes(
+ args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
+ )
######################################################################
for n_epoch in range(args.nb_epochs):
log_string(f"--- epoch {n_epoch} ----------------------------------------")
+ 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))
log_string(
f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
)
- # improve it
one_epoch(weakest_model, quizz_machine)
log_string(
f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
)
- # test it
run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
log_string(
f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
)
- cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
- log_string(f"current_test_accuracies {cta}")
+ # Replace a fraction of the w_quizzes with fresh ones
- # replace a fraction of the w_quizzes with a fresh ones
quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+ # If all the models are good enough, generate new quizzes and
+ # re-compute the test errors
+
if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
create_c_quizzes(
models,
nb_for_test=nb_new_c_quizzes_for_test,
)
- # We update everyone
for model in models:
run_tests(model, quizz_machine, deterministic_synthesis=False)