import threading
+import torch.multiprocessing as mp
+
# world quizzes vs. culture quizzes
######################################################################
else:
raise ValueError
+problem.save_some_examples(args.result_dir)
+
quiz_machine = quiz_machine.QuizMachine(
problem=problem,
nb_train_samples=args.nb_train_samples,
train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
- log_string(f"train_perplexity {n_epoch} {train_perplexity}")
+ log_string(f"train_perplexity {n_epoch} model.id {model.id} {train_perplexity}")
run_tests(model, quiz_machine, deterministic_synthesis=False)
- model.TRAINING_LOCK.release()
-
######################################################################
def standard_validity(logproba):
l = logproba.sort(dim=-1).values
return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
- # warnings.warn("TEST!!!", RuntimeWarning)
- # print(l.exp())
- # return (l[:, 0] < math.log(0.99))
def valid_c_quizzes(recorded, criteria):
model.main_test_accuracy = 0.0
model.id = k
- model.TRAINING_LOCK = threading.Lock()
- model.train_w_quizzes = quiz_machine.generate_token_sequences(
- args.nb_train_samples
- ).to(device)
+ model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
- model.test_w_quizzes = quiz_machine.generate_token_sequences(
- args.nb_test_samples
- ).to(device)
+ model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
models.append(model)
nb_new_c_quizzes_for_train = 100
nb_new_c_quizzes_for_test = 10
+ def standard_validity(logproba):
+ l = logproba.sort(dim=-1).values
+ return l[:, 0] < math.log(0.5)
+
+
######################################################################
for n_epoch in range(args.nb_epochs):
weakest_models = ranked_models[: args.nb_gpus]
+ threads = []
+
for gpu_id, model in enumerate(weakest_models):
- model.TRAINING_LOCK.acquire()
+ log_string(f"training model {model.id}")
- log_string(
- f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
+ t = threading.Thread(
+ target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
)
- threading.Thread(
- target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
- ).start()
+ threads.append(t)
- for model in weakest_models:
- model.TRAINING_LOCK.acquire()
- model.TRAINING_LOCK.release()
+ t.start()
+
+ for t in threads:
+ t.join()
##################################################
# Replace a fraction of the w_quizzes with fresh ones