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("--both_directions", 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)
+ 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,
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}")
-
######################################################################
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,))]
+ 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
- c_quizzes, ave_seq_logproba = quizz_machine.generate_quizzes(
- nb_to_create,
- model_for_generation=model_for_generation,
- )
+ model_for_generation = models[torch.randint(len(models), (1,))]
- nb_correct = quizz_machine.compute_correctness(
- c_quizzes, models, both_directions=args.both_directions
- )
+ 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 model {model_for_generation.id} 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