from torch.nn import functional as F
import ffutils
-import mygpt, tasks
+import mygpt
+import sky, wireworld, quizz_machine
+
+# world quizzes vs. culture quizzes
+
+######################################################################
+
+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("--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("--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("--check", action="store_true", default=False)
+parser.add_argument("--nb_models_for_generation", type=int, default=1)
+
+parser.add_argument("--generation_mode", type=str, default="groupthink")
+
+parser.add_argument("--min_to_validate", type=int, default=4)
+
+parser.add_argument("--max_to_validate", type=int, default=4)
+
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
+
+parser.add_argument("--dirty_debug", action="store_true", default=False)
######################################################################
######################################################################
+if args.dirty_debug:
+ args.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,
}
######################################################################
-if args.check:
- args.nb_train_samples = 500
+if args.dirty_debug:
+ args.nb_train_samples = 2500
args.nb_test_samples = 100
if args.physical_batch_size is None:
assert args.nb_train_samples % args.batch_size == 0
assert args.nb_test_samples % args.batch_size == 0
-task = tasks.World(
+if args.problem == "sky":
+ problem = sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2, speed=3)
+elif args.problem == "wireworld":
+ problem = wireworld.Wireworld(height=8, width=10, nb_iterations=2, speed=5)
+else:
+ raise ValueError
+
+quizz_machine = quizz_machine.QuizzMachine(
+ problem=problem,
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,
- logger=log_string,
deterministic_synthesis=deterministic_synthesis,
)
######################################################################
-def create_quizzes(
- model,
- other_models,
- task,
+def create_c_quizzes(
+ models,
+ quizz_machine,
nb_for_train=1000,
nb_for_test=100,
+ 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)])
+
+ model_indexes = []
+ sum_logits, sum_nb_c_quizzes = 0, 0
+
+ def nb_generated():
+ return sum([sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()])
+
+ def nb_validated():
+ return sum(
+ [
+ sum([x.size(0) for x in recorded[n]])
+ for n in range(args.min_to_validate, args.max_to_validate + 1)
+ ]
+ )
- while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
- new_quizzes, nb_correct = task.create_new_quizzes(
+ nb_to_create = nb_for_train + nb_for_test
+
+ while nb_validated() < nb_to_create:
+ (
+ new_c_quizzes,
+ nb_correct,
+ ave_seq_logproba,
+ ) = quizz_machine.gang_create_c_quizzes(
+ nb=nb_to_create,
+ nb_models_for_generation=args.nb_models_for_generation,
+ models=models,
+ mode=args.generation_mode,
+ min_ave_seq_logproba=min_ave_seq_logproba,
n_epoch=n_epoch,
result_dir=args.result_dir,
- logger=log_string,
- nb=4 * (nb_for_train + nb_for_test),
- model=model,
- other_models=other_models,
)
- print(nb_correct)
+ sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
+ sum_nb_c_quizzes += new_c_quizzes.size(0)
- to_keep = new_quizzes[nb_correct == len(other_models) - 1]
- log_string(f"keep {to_keep.size(0)} quizzes")
- kept.append(to_keep)
+ if args.dirty_debug:
+ nb_correct = torch.randint(
+ len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
+ )
- new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
+ for n in range(nb_correct.max() + 1):
+ recorded[n].append(new_c_quizzes[nb_correct == n].clone())
- task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
- task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
+ log_string(
+ f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_create}"
+ )
- task.save_image(
- new_quizzes[:72],
- args.result_dir,
- f"world_quiz_{n_epoch:04d}_{model.id:02d}.png",
- log_string,
+ # 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 = torch.cat(
+ [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)],
+ dim=0,
)
+ new_c_quizzes = new_c_quizzes[
+ torch.randperm(new_c_quizzes.size(0), device=new_c_quizzes.device)[
+ : 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.min_to_validate and n <= args.max_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_c_quizzes
+
######################################################################
######################################################################
-accuracy_to_make_quizzes = 0.975
-nb_new_quizzes_for_train = 1000
-nb_new_quizzes_for_test = 100
-
-if args.check:
- accuracy_to_make_quizzes = 0.0
- nb_new_quizzes_for_train = 10
- nb_new_quizzes_for_test = 10
+min_ave_seq_logproba = None
for n_epoch in range(args.nb_epochs):
- a = [(model.id, model.main_test_accuracy.item()) for model in models]
+ 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}")
+ s = " ".join([f"{p[1]*100:.02f}%" for p in a])
+ log_string(f"current accuracies {s}")
# select the model with lowest accuracy
models.sort(key=lambda model: model.main_test_accuracy)
)
# improve it
- one_epoch(model, task)
+ one_epoch(model, quizz_machine)
+
+ 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)
-
- if model.main_test_accuracy >= accuracy_to_make_quizzes:
- other_models = models.copy()
- other_models.remove(model)
-
- create_quizzes(
- model,
- other_models,
- task,
- nb_for_train=nb_new_quizzes_for_train,
- nb_for_test=nb_new_quizzes_for_test,
+ run_tests(model, quizz_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}"
+ )
+
+ if min([m.main_test_accuracy for m in models]) >= args.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 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)
######################################################################