exp_name = 'exp' + str(exp_idx) + '_' + exp_name + '_LeNet5'
model = LeNet5()
optimizer = SGD(learning_rate = learning_rate)
''' =============== Learning Setting =============== '''
loss_object = SparseCategoricalCrossentropy()
path_dict = dir_setting(exp_name, CONTINUE_LEARNING)
model, losses_accs, start_epoch = continue_setting(CONTINUE_LEARNING, path_dict, model=model)
train_ds, validation_ds, test_ds = load_processing_mnist(train_ratio, train_batch_size, test_batch_size)
metric_objects = get_classification_metrics()
import time
start_time = time.time()
for epoch in range(start_epoch, epochs):
train(train_ds, model, loss_object, optimizer, metric_objects)
validation(validation_ds, model, loss_object, metric_objects)
training_reporter(epoch, losses_accs, metric_objects, exp_name=exp_name)
save_metrics_model(epoch, model, losses_accs, path_dict, save_period)
metric_visualizer(losses_accs, path_dict['cp_path'])
resetter(metric_objects)
end_time = time.time()
elapsed_time = end_time - start_time
with open(path_dict['cp_path'] + './elapsed_time.txt', 'w') as f:
f.write(str(elapsed_time))
test(test_ds, model, loss_object, metric_objects, path_dict)
CONTINUE_LEARNING = False
train_ratio = 0.8
train_batch_size, test_batch_size = 128, 128
epochs = 10
save_interval = 2
learning_rate = 0.01
model = LeNet5()
optimizer = SGD(learning_rate=learning_rate)
loss_object = SparseCategoricalCrossentropy()
path_dict = dir_setting(exp_name, CONTINUE_LEARNING)
model, losses_accs, start_epoch = continue_setting(CONTINUE_LEARNING,
path_dict, model)
train_ds, validation_ds, test_ds = get_ds(train_ratio, train_batch_size,
test_batch_size, "mnist")
metric_objects = get_classification_metrics()
for epoch in range(start_epoch, epochs):
train(train_ds, model, loss_object, optimizer, metric_objects)
validation(validation_ds, model, loss_object, metric_objects)
training_reporter(epoch, losses_accs, metric_objects)
save_metrics_model(epoch, model, losses_accs, path_dict, save_interval)
metric_visualizer(losses_accs, path_dict['cp_path'])
resetter(metric_objects)
test(test_ds, model, loss_object, metric_objects, path_dict)
if DONE:
# 각 에피소드 당 학습 정보를 기록
max_q_value = np.max(
(agent.model1(state).numpy() + agent.model1(state).numpy()) /
2)
scores.append(score)
max_q_values.append(max_q_value)
score_avg(score)
max_q_avg(max_q_value)
loss_avg(loss.numpy())
if episode % report_period == 0:
if score_avg.result() >= 399:
print('Saturated!')
save_metrics_model(episode, model, max_q_values_scores, path_dict)
sys.exit()
save_learning_curve(max_q_values_scores, report_period,
path_dict['cp_path'])
training_reporter(episode,
max_q_values_scores,
metric_objects,
exp_name=exp_name)
agent_reporter(agent)
resetter(metric_objects)
# save_period 에피소드마다 모델 저장
if episode % save_period == 0:
save_metrics_model(episode, model, max_q_values_scores, path_dict)