n_epochs = 10
n_al_epochs = 10
only_real = True
y_columns = list(range(data_set.y.shape[1]))
y_columns = [2, 11, 23]
net.build_multi_reg_k(len(y_columns),
activation=tf.keras.layers.LeakyReLU())
net1.build_multi_reg_k(len(y_columns),
activation=tf.keras.layers.LeakyReLU())
data_set.split_dataset_random(train_perc=0.0,
val_perc=0.1,
notused_perc=0.9)
data_set.make_dataset_tensors(y_columns=y_columns, only_real=only_real)
val_sample_size = data_set.val_labels.shape[0]
validation_steps = val_sample_size // n_batch
val_gen = data_set.dataset_generator_multi(data_set.val_datas,
data_set.val_labels, n_batch)
for al in range(n_al_epochs):
al_res = dataset_predict(net.pre_model, data_set, role=0)[:, 1]
al_res = al_res.reshape((-1, len(y_columns)))
al_res1 = dataset_predict(net1.pre_model, data_set, role=0)[:, 1]
al_res1 = al_res1.reshape((-1, len(y_columns)))
delta = np.abs(al_res - al_res1).mean(axis=1)
al_as = delta.argsort()[::-1]
new_train_num = min(np.sum(delta > 0.02), 200)
idx = data_set.notused_indices[al_as[:new_train_num]]
data_set.sample_prop.role[idx] = 1
print("add %d new train samples." % new_train_num)
data_set.make_dataset_tensors(y_columns=y_columns, only_real=only_real)
train_sample_size = data_set.train_labels.shape[0]
dt_test_begin = datetime.datetime(2018, 8, 26)
dt_test_end = datetime.datetime(2018, 9, 1) - datetime.timedelta(seconds=1)
ids = data_set.split_dataset_dt(dt_train_begin=dt_train_begin,
dt_train_end=dt_train_end,
dt_test_begin=dt_test_begin,
dt_test_end=dt_test_end,
val_perc=0.1)
assert (0 in ids) and (1 in ids)
data_set.make_dataset_tensors(y_columns=y_columns, only_real=only_real)
train_data, train_labels, _ = data_set.get_dataset(0)
train_sample_size = train_labels.shape[0]
steps_per_epoch = train_sample_size // n_batch
val_data, val_labels, _ = data_set.get_dataset(1)
val_sample_size = val_labels.shape[0]
validation_steps = val_sample_size // n_batch
train_gen = data_set.dataset_generator_multi(train_data, train_labels, n_batch)
val_gen = data_set.dataset_generator_multi(val_data, val_labels, n_batch)
dt_run_start = datetime.datetime.now()
with timer("Timer training"):
history = net.train_model.fit_generator(train_gen,
epochs=n_epochs,
steps_per_epoch=steps_per_epoch,
validation_data=val_gen,
validation_steps=validation_steps)
test_labels, pre = dataset_predict(net.pre_model, data_set, role=2)
save_model(res_path, args.res_type, net.pre_model, suffix='pb')
# save_model(res_path, args.res_type, net.pre_model, suffix='json')
write_input(data_set, res_path + "/input.txt")
write_output(data_set.y.columns[y_columns], res_path + "/output.txt")
sys.exit(0)