import ffutils
import mygpt
-import sky, quizz_machine
+import sky, reasoning, quiz_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("--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_correct_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=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)
+
+parser.add_argument("--sky_nb_birds", type=int, default=3)
+
+parser.add_argument("--sky_nb_iterations", type=int, default=2)
+
+parser.add_argument("--sky_speed", type=int, default=3)
+
+######################################################################
+
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"
######################################################################
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
assert args.nb_train_samples % args.batch_size == 0
assert args.nb_test_samples % args.batch_size == 0
-quizz_machine = quizz_machine.QuizzMachine(
- problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
+if args.problem == "sky":
+ problem = sky.Sky(
+ height=args.sky_height,
+ width=args.sky_width,
+ nb_birds=args.sky_nb_birds,
+ nb_iterations=args.sky_nb_iterations,
+ speed=args.sky_speed,
+ )
+ back_accuracy = False
+elif args.problem == "reasoning":
+ problem = reasoning.Reasoning(device=device)
+ back_accuracy = True
+else:
+ raise ValueError
+
+quiz_machine = quiz_machine.QuizMachine(
+ 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,
log_string(f"device {device}")
-vocabulary_size = quizz_machine.vocabulary_size()
+vocabulary_size = quiz_machine.vocabulary_size()
log_string(f"vocabulary_size {vocabulary_size}")
# Compute the entropy of the training tokens
token_count = 0
-for input in quizz_machine.batches(split="train", desc="train-entropy"):
- token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
+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()
nb_test, nb_in_train = 0, 0
for test_subset in subsets_as_tuples(
- quizz_machine.batches(split="test", desc="test-check"), 25000
+ quiz_machine.batches(split="test", desc="test-check"), 25000
):
in_train = set()
for train_subset in subsets_as_tuples(
- quizz_machine.batches(split="train", desc="train-check"), 25000
+ quiz_machine.batches(split="train", desc="train-check"), 25000
):
in_train.update(test_subset.intersection(train_subset))
nb_in_train += len(in_train)
##############################
-def one_epoch(model, quizz_machine):
+def one_epoch(model, quiz_machine):
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
model.train()
nb_train_samples, acc_train_loss = 0, 0.0
- for input in quizz_machine.batches(split="train"):
+ for input in quiz_machine.batches(split="train"):
input = input.to(device)
if nb_train_samples % args.batch_size == 0:
######################################################################
-def run_tests(model, quizz_machine, deterministic_synthesis):
+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 quizz_machine.batches(split="test"):
+ for input in quiz_machine.batches(split="test"):
input = input.to(device)
bs = model(mygpt.BracketedSequence(input))
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 = quiz_machine.produce_results(
n_epoch=n_epoch,
model=model,
result_dir=args.result_dir,
- logger=log_string,
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
+def valid_c_quizzes(recorded, criteria):
+ result = [q[criteria(c)] for q, c in recorded]
+ return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
######################################################################
def create_c_quizzes(
models,
- quizz_machine,
+ quiz_machine,
nb_for_train=1000,
nb_for_test=100,
- min_ave_seq_logproba=None,
):
- # We will store the generated quizzes for each number of
- # correct prediction
- recorded = dict([(n, []) for n in range(len(models) + 1)])
+ quizzes_and_nb_correct_records = []
- model_indexes = []
- sum_logits, sum_nb_c_quizzes = 0, 0
+ nb_to_create = nb_for_train + nb_for_test
- while (
- sum([x.size(0) for x in recorded[args.nb_correct_to_validate]])
- < nb_for_train + nb_for_test
- ):
- nb_to_validate = nb_for_train + nb_for_test
+ # ------------------------------------------------------------
- if len(model_indexes) == 0:
- model_indexes = [i.item() for i in torch.randperm(len(models))]
+ standard_validity = lambda nb_correct: torch.logical_and(
+ nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
+ )
- model = models[model_indexes.pop()]
+ 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
- new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
- nb=nb_to_validate,
- model_for_generation=model,
- models_for_validation=models,
- min_ave_seq_logproba=min_ave_seq_logproba,
- n_epoch=n_epoch,
- result_dir=args.result_dir,
- logger=log_string,
- )
+ model_for_generation = models[torch.randint(len(models), (1,))]
- sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
- sum_nb_c_quizzes += new_c_quizzes.size(0)
+ c_quizzes = quiz_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=new_c_quizzes.device
+ nb_correct, seq_logproba = quiz_machine.compute_correctness(
+ c_quizzes,
+ models,
+ bidirectional_validation=args.bidirectional_validation,
+ deterministic_validation=args.deterministic_validation,
)
- for n in range(nb_correct.max() + 1):
- recorded[n].append(new_c_quizzes[nb_correct == n].clone())
+ 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")
- nb_validated = sum([x.size(0) for x in recorded[args.nb_correct_to_validate]])
- nb_generated = sum(
- [sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()]
- )
+ if args.dirty_debug:
+ nb_correct = torch.randint(
+ len(models) + 1, nb_correct.size(), device=c_quizzes.device
+ )
- log_string(
- f"keep c_quizzes {nb_validated*100/nb_generated:.02f}% kept total {nb_validated}/{nb_to_validate}"
- )
+ 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
- # 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 = recorded[args.nb_correct_to_validate][: 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)
-
- for n in recorded.keys():
- s = "_validated" if n == args.nb_correct_to_validate else ""
- quizz_machine.problem.save_quizzes(
- recorded[n][:72],
- args.result_dir,
- f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
+ 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 ""
)
- return sum_logits / sum_nb_c_quizzes
+ q = valid_c_quizzes(
+ quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
+ )[:72]
+
+ if q.size(0) > 0:
+ quiz_machine.save_quizzes(
+ args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
+ )
######################################################################
######################################################################
-min_ave_seq_logproba = None
-
for n_epoch in range(args.nb_epochs):
log_string(f"--- epoch {n_epoch} ----------------------------------------")
- 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}")
+ cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
+ log_string(f"current_test_accuracies {cta}")
- # select the model with lowest accuracy
- models.sort(key=lambda model: model.main_test_accuracy)
- model = models[0]
+ # Select, improve, and eval the worst model
+
+ weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
log_string(
- f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
+ f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
)
- # improve it
- one_epoch(model, quizz_machine)
-
- quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+ one_epoch(weakest_model, quiz_machine)
log_string(
- f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
+ f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
)
- # test it
- run_tests(model, quizz_machine, deterministic_synthesis=False)
+ run_tests(weakest_model, quiz_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}"
+ f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
)
- if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
- ave_seq_logproba = create_c_quizzes(
+ # Replace a fraction of the w_quizzes with fresh ones
+
+ quiz_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,
- quizz_machine,
+ quiz_machine,
nb_for_train=nb_new_c_quizzes_for_train,
nb_for_test=nb_new_c_quizzes_for_test,
- min_ave_seq_logproba=min_ave_seq_logproba,
)
- # 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}"
- # )
-
- # We update everyone
for model in models:
- run_tests(model, quizz_machine, deterministic_synthesis=False)
+ run_tests(model, quiz_machine, deterministic_synthesis=False)
######################################################################