from torch.nn import functional as F
import ffutils
-import mygpt, tasks
+import mygpt, quizz_machine
# world quizzes vs. culture quizzes
assert args.nb_train_samples % args.batch_size == 0
assert args.nb_test_samples % args.batch_size == 0
-task = tasks.World(
+quizz_machine = quizz_machine.QuizzMachine(
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
batch_size=args.physical_batch_size,
log_string(f"device {device}")
-vocabulary_size = task.vocabulary_size()
+vocabulary_size = quizz_machine.vocabulary_size()
log_string(f"vocabulary_size {vocabulary_size}")
# Compute the entropy of the training tokens
token_count = 0
-for input in task.batches(split="train", desc="train-entropy"):
- token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
+for input in quizz_machine.batches(split="train", desc="train-entropy"):
+ token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
+ (0, 1)
+ )
token_probas = token_count / token_count.sum()
entropy = -torch.xlogy(token_probas, token_probas).sum()
train_set_perplexity = math.exp(entropy)
nb_test, nb_in_train = 0, 0
for test_subset in subsets_as_tuples(
- task.batches(split="test", desc="test-check"), 25000
+ quizz_machine.batches(split="test", desc="test-check"), 25000
):
in_train = set()
for train_subset in subsets_as_tuples(
- task.batches(split="train", desc="train-check"), 25000
+ quizz_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, task):
+def one_epoch(model, quizz_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 task.batches(split="train"):
+ for input in quizz_machine.batches(split="train"):
input = input.to(device)
if nb_train_samples % args.batch_size == 0:
######################################################################
-def run_tests(model, task, deterministic_synthesis):
+def run_tests(model, quizz_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 task.batches(split="test"):
+ for input in quizz_machine.batches(split="test"):
input = input.to(device)
bs = model(mygpt.BracketedSequence(input))
nb_test_samples += input.size(0)
- main_test_accuracy = task.produce_results(
+ main_test_accuracy = quizz_machine.produce_results(
n_epoch=n_epoch,
model=model,
result_dir=args.result_dir,
def create_c_quizzes(
model,
other_models,
- task,
+ quizz_machine,
nb_for_train=1000,
nb_for_test=100,
min_ave_seq_logproba=None,
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)
- new_c_quizzes, nb_correct, ave_seq_logproba = task.create_c_quizzes(
+ new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
n_epoch=n_epoch,
result_dir=args.result_dir,
logger=log_string,
new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
- task.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
- task.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
+ 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)
- task.save_quizzes(
+ quizz_machine.save_quizzes(
new_c_quizzes[:72],
args.result_dir,
f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
)
# improve it
- one_epoch(model, task)
+ one_epoch(model, quizz_machine)
- task.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+ quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
log_string(
- f"train_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
+ f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
)
# test it
- run_tests(model, task, deterministic_synthesis=False)
+ run_tests(model, quizz_machine, deterministic_synthesis=False)
log_string(
- f"test_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
+ 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:
ave_seq_logproba = create_c_quizzes(
model,
other_models,
- task,
+ quizz_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 update everyone
for model in models:
- run_tests(model, task, deterministic_synthesis=False)
+ run_tests(model, quizz_machine, deterministic_synthesis=False)
######################################################################
):
assert input.size() == ar_mask.size()
- batches = zip(input.split(batch_size), ar_mask.split(batch_size))
+ batches = zip(
+ input.split(batch_size),
+ ar_mask.split(batch_size),
+ seq_logproba.split(batch_size),
+ )
if progress_bar_desc is not None:
batches = tqdm.tqdm(
t = model.training
model.eval()
- for input, ar_mask in batches:
+ for input, ar_mask, seq_logproba in batches:
model.masked_inplace_autoregression(
input=input,
ar_mask=ar_mask,
import world
-class World(Task):
+class QuizzMachine(Task):
def save_image(self, input, result_dir, filename, logger):
img = world.seq2img(input.to("cpu"), self.height, self.width)
image_name = os.path.join(result_dir, filename)
input = input[:nmax]
ar_mask = self.make_ar_mask(input)
result = input.clone() * (1 - ar_mask)
+ seq_logproba = torch.empty(input.size(0), device=self.device)
masked_inplace_autoregression(
model=model,
batch_size=self.batch_size,
input=result,
ar_mask=ar_mask,
- seq_logproba=None,
+ seq_logproba=seq_logproba,
temperature=1.0,
deterministic_synthesis=deterministic_synthesis,
progress_bar_desc=None,
input = self.test_w_quizzes[:96]
ar_mask = self.make_ar_mask(input)
result = input.clone() * (1 - ar_mask)
+ seq_logproba = torch.empty(input.size(0), device=self.device)
masked_inplace_autoregression(
model=model,
batch_size=self.batch_size,
input=result,
ar_mask=ar_mask,
- seq_logproba=None,
+ seq_logproba=seq_logproba,
temperature=1.0,
deterministic_synthesis=deterministic_synthesis,
progress_bar_desc=None,
nb,
model,
other_models,
- min_ave_seq_logproba=None,
+ min_ave_seq_logproba,
):
###############################################################
# Generate quizzes with model
)
ar_mask = torch.full(c_quizzes.size(), 1, device=self.device)
- seq_logproba = torch.empty(nb, device=self.device)
+ seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
temperature = 1
d_temperature = 1
# Oh man that's ugly
if ave_seq_logproba < min_ave_seq_logproba * 1.1:
if d_temperature > 0:
- d_temperature *= -0.5
+ d_temperature *= -1 / 3
temperature += d_temperature
elif ave_seq_logproba > min_ave_seq_logproba:
if d_temperature < 0:
- d_temperature *= -0.5
+ d_temperature *= -1 / 3
temperature += d_temperature
else:
break
)
ar_mask = self.make_ar_mask(c_quizzes)
+ seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
###############################################################
# Check how many of the other models can solve them in both
batch_size=self.batch_size,
input=result,
ar_mask=ar_mask,
- seq_logproba=None,
+ seq_logproba=seq_logproba,
temperature=1.0,
deterministic_synthesis=True,
progress_bar_desc="solving c_quizzes",
batch_size=self.batch_size,
input=reverse_result,
ar_mask=ar_mask,
- seq_logproba=None,
+ seq_logproba=seq_logproba,
temperature=1.0,
deterministic_synthesis=True,
progress_bar_desc="solving reversed c_quizzes",
projects / culture.git / commitdiff