import ffutils
import mygpt
-import sky, reasoning, quizz_machine
+import sky, reasoning, quiz_machine
# world quizzes vs. culture quizzes
else:
raise ValueError
-quizz_machine = quizz_machine.QuizzMachine(
+quiz_machine = quiz_machine.QuizMachine(
problem=problem,
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
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))
log_string(f"test_perplexity {n_epoch} {test_perplexity}")
- model.main_test_accuracy = quizz_machine.produce_results(
+ model.main_test_accuracy = quiz_machine.produce_results(
n_epoch=n_epoch,
model=model,
result_dir=args.result_dir,
def create_c_quizzes(
models,
- quizz_machine,
+ quiz_machine,
nb_for_train=1000,
nb_for_test=100,
):
- recorded = []
+ quizzes_and_nb_correct_records = []
nb_to_create = nb_for_train + nb_for_test
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:
+ 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 = quizz_machine.generate_quizzes(
+ c_quizzes = quiz_machine.generate_quizzes(
nb_to_create,
model_for_generation=model_for_generation,
temperature=args.generation_temperature,
)
- nb_correct, seq_logproba = quizz_machine.compute_correctness(
+ nb_correct, seq_logproba = quiz_machine.compute_correctness(
c_quizzes,
models,
bidirectional_validation=args.bidirectional_validation,
len(models) + 1, nb_correct.size(), device=c_quizzes.device
)
- recorded.append((c_quizzes, nb_correct))
+ 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(recorded, standard_validity).size(0)
+ 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(recorded, standard_validity)
+ new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
+
+ quiz_machine.reverse_random_half_in_place(new_c_quizzes)
- 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)
+ 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
else ""
)
- q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72]
+ q = valid_c_quizzes(
+ quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
+ )[:72]
if q.size(0) > 0:
- quizz_machine.save_quizzes(
+ quiz_machine.save_quizzes(
args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
)
f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
)
- one_epoch(weakest_model, quizz_machine)
+ 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}"
)
- run_tests(weakest_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}"
)
# Replace a fraction of the w_quizzes with fresh ones
- quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+ 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,
)
for model in models:
- run_tests(model, quizz_machine, deterministic_synthesis=False)
+ run_tests(model, quiz_machine, deterministic_synthesis=False)
######################################################################
######################################################################
-class QuizzMachine:
+class QuizMachine:
def indices_forward_and_backward(self, quizzes):
i_forward = quizzes[:, 0] == self.token_forward
j_forward = quizzes[:, 1 + self.prompt_len] == self.token_forward
return m * forward_to_backward + (1 - m) * backward_to_forward
+ def reverse_random_half_in_place(self, quizzes):
+ i = torch.rand(quizzes.size(0)) < 0.5
+ if i.any():
+ quizzes[i] = self.reverse_time(quizzes[i])
+
def make_ar_mask(self, quizzes, first=False):
i_forward, i_backward = self.indices_forward_and_backward(quizzes)
result = []
for prompt, answer in zip(prompts, answers):
- if torch.rand(1) < 0.5:
- a = [
- torch.tensor([self.token_forward]),
- prompt,
- torch.tensor([self.token_forward]),
- answer,
- ]
- else:
- a = [
- torch.tensor([self.token_backward]),
- answer,
- torch.tensor([self.token_backward]),
- prompt,
- ]
+ a = [
+ torch.tensor([self.token_forward]),
+ prompt,
+ torch.tensor([self.token_forward]),
+ answer,
+ ]
result.append(torch.cat(a, dim=0)[None, :])
self.prompt_len = None
self.answer_len = None
- self.train_w_quizzes = self.generate_token_sequences(nb_train_samples).to(
- device
- )
+ self.train_w_quizzes = self.generate_token_sequences(nb_train_samples)
+ self.reverse_random_half_in_place(self.train_w_quizzes)
+ self.train_w_quizzes = self.train_w_quizzes.to(device)
self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
+ self.reverse_random_half_in_place(self.test_w_quizzes)
+ self.test_w_quizzes = self.test_w_quizzes.to(device)
self.train_c_quizzes = []
self.test_c_quizzes = []
result_dir,
"culture_w_quizzes",
self.train_w_quizzes[:72],
- n_backward=self.train_w_quizzes[:72, 0] == self.token_backward,
)
def save_quizzes(
result_dir,
filename_prefix,
quizzes,
- n_backward=None,
mistakes=None,
):
quizzes = quizzes.clone()
- forward = quizzes[quizzes[:, 0] == self.token_forward]
- ib = quizzes[:, 0] == self.token_backward
- backward = quizzes[ib]
- assert forward.size(0) + backward.size(0) == quizzes.size(0)
- quizzes[ib] = self.reverse_time(quizzes[ib])
-
- if n_backward is None:
- predicted_prompts = None
- predicted_answers = None
+ n_forward = quizzes[quizzes[:, 0] == self.token_forward]
+ n_backward = quizzes[:, 0] == self.token_backward
+ backward = quizzes[n_backward]
+ assert n_forward.size(0) + backward.size(0) == quizzes.size(0)
+ quizzes[n_backward] = self.reverse_time(quizzes[n_backward])
+
+ predicted_prompts = n_backward.long()
+ predicted_answers = 1 - predicted_prompts
+ if mistakes is not None:
+ # 0/-1/+1 ~ not-to-predict / predicted wrong / predicted correct
+ predicted_prompts *= mistakes
+ predicted_answers *= mistakes
else:
- predicted_prompts = n_backward.long()
- predicted_answers = 1 - predicted_prompts
- if mistakes is not None:
- # 0/-1/+1 ~ not-to-predict / predicted wrong / predicted correct
- predicted_prompts *= mistakes
- predicted_answers *= mistakes
- else:
- # 0/2 ~ not-to-predict / to predict
- predicted_prompts *= 2
- predicted_answers *= 2
+ # 0/2 ~ not-to-predict / to predict
+ predicted_prompts *= 2
+ predicted_answers *= 2
self.problem.save_quizzes(
result_dir,
back_input[:, 2 + self.prompt_len :] = input[
n_backward, 1 : 1 + self.answer_len
]
- result[n_backward], correct[n_backward] = compute_accuracy(back_input)
+ _, correct[n_backward] = compute_accuracy(back_input)
if log_prefix is not None:
forward_nb_correct = correct[n_forward].sum()
result_dir,
f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
quizzes=test_result[:72],
- n_backward=self.test_w_quizzes[:72, 0] == self.token_backward,
mistakes=test_correct[:72] * 2 - 1,
)
input = self.train_w_quizzes if for_train else self.test_w_quizzes
nb = min(nb, input.size(0))
input[:-nb] = input[nb:].clone()
- input[-nb:] = self.generate_token_sequences(nb).to(self.device)
+ fresh_w_quizzes = self.generate_token_sequences(nb)
+ self.reverse_random_half_in_place(fresh_w_quizzes)
+ input[-nb:] = fresh_w_quizzes.to(self.device)
def store_c_quizzes(self, new_c_quizzes, for_train=True):
if for_train:
projects / culture.git / commitdiff