from torch.nn import functional as F
import ffutils
-import mygpt, tasks
+import mygpt
+import sky, quizz_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
######################################################################
parser.add_argument("--nb_test_samples", type=int, default=None)
-parser.add_argument("--learning_rate", type=float, default=1e-4)
+parser.add_argument("--learning_rate", type=float, default=1e-3)
########################################
parser.add_argument("--nb_gpts", type=int, default=5)
+parser.add_argument("--nb_correct_to_validate", type=int, default=4)
+
parser.add_argument("--dirty_debug", action="store_true", default=False)
######################################################################
######################################################################
+if args.dirty_debug:
+ accuracy_to_make_c_quizzes = 0.0
+ nb_new_c_quizzes_for_train = 100
+ nb_new_c_quizzes_for_test = 10
+
+######################################################################
+
default_args = {
"model": "37M",
"batch_size": 100,
- "nb_train_samples": 250000,
+ "nb_train_samples": 100000,
"nb_test_samples": 10000,
}
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(
+ problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
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_quizzes(
- model,
- other_models,
- task,
+def create_c_quizzes(
+ models,
+ quizz_machine,
nb_for_train=1000,
nb_for_test=100,
- desired_average_logits=None,
+ min_ave_seq_logproba=None,
):
- kept = []
+ # We will store the generated quizzes for each number of
+ # correct prediction
+ recorded = dict([(n, []) for n in range(len(models) + 1)])
- sum_logits, sum_nb_quizzes = 0, 0
+ model_indexes = []
+ sum_logits, sum_nb_c_quizzes = 0, 0
- 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)
+ 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
- new_quizzes, nb_correct, average_logits = task.create_new_quizzes(
+ if len(model_indexes) == 0:
+ model_indexes = [i.item() for i in torch.randperm(len(models))]
+
+ model = models[model_indexes.pop()]
+
+ 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,
- nb=nb_to_generate,
- model=model,
- other_models=other_models,
- desired_average_logits=desired_average_logits,
)
- sum_logits += new_quizzes.size(0) * average_logits
- sum_nb_quizzes += new_quizzes.size(0)
-
- to_keep = new_quizzes[nb_correct == len(other_models) - 1]
+ sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
+ sum_nb_c_quizzes += new_c_quizzes.size(0)
if args.dirty_debug:
- to_keep = new_quizzes
+ nb_correct = torch.randint(
+ len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
+ )
+
+ for n in range(nb_correct.max() + 1):
+ recorded[n].append(new_c_quizzes[nb_correct == n].clone())
+
+ 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()]
+ )
log_string(
- f"keep {to_keep.size(0)}/{new_quizzes.size(0)} quizzes ({to_keep.size(0)*100/new_quizzes.size(0):.02f}%)"
+ f"keep c_quizzes {nb_validated*100/nb_generated:.02f}% kept total {nb_validated}/{nb_to_validate}"
)
- kept.append(to_keep)
+ # 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_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
+ new_c_quizzes = recorded[args.nb_correct_to_validate][: nb_for_train + nb_for_test]
- task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
- task.store_new_quizzes(new_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_image(
- new_quizzes[:72],
- args.result_dir,
- f"world_quiz_{n_epoch:04d}_{model.id:02d}.png",
- log_string,
- )
+ 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}",
+ )
- return sum_logits / sum_nb_quizzes
+ return sum_logits / sum_nb_c_quizzes
######################################################################
######################################################################
-accuracy_to_make_quizzes = 0.975
-nb_new_quizzes_for_train = 1000
-nb_new_quizzes_for_test = 100
-
-if args.dirty_debug:
- accuracy_to_make_quizzes = 0.0
- nb_new_quizzes_for_train = 100
- nb_new_quizzes_for_test = 10
-
-desired_average_logits = None
+min_ave_seq_logproba = None
for n_epoch in range(args.nb_epochs):
log_string(f"--- epoch {n_epoch} ----------------------------------------")
)
# improve it
- one_epoch(model, task)
+ one_epoch(model, quizz_machine)
- task.renew_samples(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 world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_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 world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_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_quizzes:
- other_models = models.copy()
- other_models.remove(model)
-
- average_logits = create_quizzes(
- model,
- other_models,
- task,
- nb_for_train=nb_new_quizzes_for_train,
- nb_for_test=nb_new_quizzes_for_test,
- desired_average_logits=desired_average_logits,
+ if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
+ ave_seq_logproba = create_c_quizzes(
+ models,
+ 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 keep the first average logits as a reference
- if desired_average_logits is None:
- desired_average_logits = average_logits
- else:
- log_string(
- f"desired_average_logits {desired_average_logits} average_logits {average_logits}"
- )
+ # 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, task, deterministic_synthesis=False)
+ run_tests(model, quizz_machine, deterministic_synthesis=False)
######################################################################