def test(self):
"""
Test models
"""
start_time = time.time()
tf.reset_default_graph()
loaded_graph = tf.Graph()
with tf.Session(graph=loaded_graph) as sess:
# Load saved models
loader = tf.train.import_meta_graph(self.checkpoint_path + '.meta')
loader.restore(sess, self.checkpoint_path)
# Get Tensors from loaded models
if self.cfg.TEST_WITH_RECONSTRUCTION:
inputs, labels, loss, accuracy, \
clf_loss, rec_loss, rec_images = \
self._get_tensors(loaded_graph)
else:
inputs, labels, loss, accuracy = self._get_tensors(
loaded_graph)
clf_loss, rec_loss, rec_images = None, None, None
utils.thick_line()
print('Testing on test set...')
utils.thin_line()
print('Calculating loss and accuracy of test set...')
loss_test, clf_loss_test, rec_loss_test, acc_test = \
self._eval_on_batches(
sess, inputs, labels, loss, accuracy,
clf_loss, rec_loss, rec_images,
self.x_test, self.y_test, self.n_batch_test)
# Print losses and accuracy
utils.thin_line()
print('Test_Loss: {:.4f}'.format(loss_test))
if self.cfg.TEST_WITH_RECONSTRUCTION:
print('Test_Classifier_Loss: {:.4f}\n'.format(clf_loss_test),
'Test_Reconstruction_Loss: {:.4f}'.format(rec_loss_test))
print('Test_Accuracy: {:.2f}%'.format(acc_test * 100))
# Save test log
utils.save_test_log(self.test_log_path, loss_test, acc_test,
clf_loss_test, rec_loss_test,
self.cfg.TEST_WITH_RECONSTRUCTION)
utils.thin_line()
print('Testing finished! Using time: {:.2f}'.format(time.time() -
start_time))
utils.thick_line()
def _test_after_training(self, sess):
"""
Evaluate on the test set after training.
"""
test_start_time = time.time()
utils.thick_line()
print('Testing...')
# Check directory of paths
utils.check_dir([self.test_log_path])
if self.cfg.WITH_RECONSTRUCTION:
if self.cfg.TEST_SAVE_IMAGE_STEP is not None:
utils.check_dir([self.test_image_path])
# Load data
utils.thin_line()
print('Loading test set...')
utils.thin_line()
x_test = utils.load_data_from_pkl(
join(self.preprocessed_path, 'x_test.p'))
y_test = utils.load_data_from_pkl(
join(self.preprocessed_path, 'y_test.p'))
n_batch_test = len(y_test) // self.cfg.BATCH_SIZE
utils.thin_line()
print('Calculating loss and accuracy on test set...')
loss_test_all = []
acc_test_all = []
clf_loss_test_all = []
rec_loss_test_all = []
step = 0
_test_batch_generator = utils.get_batches(x_test, y_test,
self.cfg.BATCH_SIZE)
if self.cfg.WITH_RECONSTRUCTION:
for _ in tqdm(range(n_batch_test),
total=n_batch_test,
ncols=100,
unit=' batches'):
step += 1
test_batch_x, test_batch_y = next(_test_batch_generator)
loss_test_i, clf_loss_i, rec_loss_i, acc_test_i = sess.run(
[self.loss, self.clf_loss, self.rec_loss, self.accuracy],
feed_dict={
self.inputs: test_batch_x,
self.labels: test_batch_y,
self.is_training: False
})
loss_test_all.append(loss_test_i)
acc_test_all.append(acc_test_i)
clf_loss_test_all.append(clf_loss_i)
rec_loss_test_all.append(rec_loss_i)
# Save reconstruct images
if self.cfg.TEST_SAVE_IMAGE_STEP is not None:
if step % self.cfg.TEST_SAVE_IMAGE_STEP == 0:
self._save_images(sess,
self.test_image_path,
test_batch_x,
test_batch_y,
step,
silent=False)
clf_loss_test = sum(clf_loss_test_all) / len(clf_loss_test_all)
rec_loss_test = sum(rec_loss_test_all) / len(rec_loss_test_all)
else:
for _ in tqdm(range(n_batch_test),
total=n_batch_test,
ncols=100,
unit=' batches'):
test_batch_x, test_batch_y = next(_test_batch_generator)
loss_test_i, acc_test_i = sess.run(
[self.loss, self.accuracy],
feed_dict={
self.inputs: test_batch_x,
self.labels: test_batch_y,
self.is_training: False
})
loss_test_all.append(loss_test_i)
acc_test_all.append(acc_test_i)
clf_loss_test, rec_loss_test = None, None
loss_test = sum(loss_test_all) / len(loss_test_all)
acc_test = sum(acc_test_all) / len(acc_test_all)
# Print losses and accuracy
utils.thin_line()
print('Test_Loss: {:.4f}\n'.format(loss_test),
'Test_Accuracy: {:.2f}%'.format(acc_test * 100))
if self.cfg.WITH_RECONSTRUCTION:
utils.thin_line()
print('Test_Train_Loss: {:.4f}\n'.format(clf_loss_test),
'Test_Reconstruction_Loss: {:.4f}'.format(rec_loss_test))
# Save test log
utils.save_test_log(self.test_log_path, loss_test, acc_test,
clf_loss_test, rec_loss_test,
self.cfg.WITH_RECONSTRUCTION)
utils.thin_line()
print('Testing finished! Using time: {:.2f}'.format(time.time() -
test_start_time))
def tester(self,
sess,
inputs,
labels,
input_imgs,
is_training,
clf_preds,
rec_imgs,
start_time,
loss=None,
acc=None,
clf_loss=None,
rec_loss=None):
utils.thin_line()
print('Calculating loss and accuracy of test set...')
# Get losses and accuracies
clf_preds_vec_test, loss_test, clf_loss_test, rec_loss_test, acc_test = \
self._eval_on_batches(
sess, inputs, labels, input_imgs, is_training, clf_preds,
loss, acc, clf_loss, rec_loss, rec_imgs)
# Get integer predictions
_ = self._get_preds_int(preds_vec=clf_preds_vec_test)
# Get top N accuracy
if self.cfg.TOP_N_LIST is not None:
acc_top_n_list = self._get_top_n_accuracy(clf_preds_vec_test)
else:
acc_top_n_list = None
# Print losses and accuracy
utils.thin_line()
print('Test Loss: {:.4f}'.format(loss_test))
if self.cfg.TEST_WITH_REC:
print('Test Classifier Loss: {:.4f}\n'.format(clf_loss_test),
'Test Reconstruction Loss: {:.4f}'.format(rec_loss_test))
print('Test Accuracy: {:.4f}%'.format(acc_test * 100))
if self.cfg.TOP_N_LIST is not None:
utils.thin_line()
for i, top_n in enumerate(self.cfg.TOP_N_LIST):
print('Top_{} Test Accuracy: {:.4f}% \n'.format(
top_n, acc_top_n_list[i] * 100))
# Save test log
if self.during_training and (self.epoch_train != 'end'):
utils.save_test_log_is_training(
self.test_log_path, self.epoch_train, self.step_train,
loss_test, acc_test, clf_loss_test, rec_loss_test,
self.cfg.TEST_WITH_REC, self.cfg.TOP_N_LIST, acc_top_n_list)
else:
utils.save_test_log(self.test_log_path, loss_test, acc_test,
clf_loss_test, rec_loss_test,
self.cfg.TEST_WITH_REC, self.cfg.TOP_N_LIST,
acc_top_n_list)
utils.thin_line()
print('Testing finished! Using time: {:.2f}'.format(time.time() -
start_time))
utils.thick_line()