def train(self, epoch):
os.makedirs('tb/%s' % self.model_name)
os.makedirs('trained_model/%s' % self.model_name)
tf.summary.image('input x_image', self.x_image, 4)
tf.summary.image('y_prediction', self.y_conv, 4)
tf.summary.image('y_GT', self.y_, 4)
tf.summary.image('y_pred_softmax', self.y_soft, 4)
tf.summary.scalar('cross_entropy', self.cross_entropy)
tf.summary.scalar('learning rate', self.lr)
sess = tf.Session()
saver = tf.train.Saver()
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('/home/lsmjn/Drone-CNN/tb/%s' % self.model_name, sess.graph)
conn, cur = data.get_db_connection()
steps = data.get_steps(self.batch_size)
filename_queue = tf.train.string_input_producer(['Drone-CNN.tfrecords'])
image, annotation = data.read_and_decode(filename_queue, self.batch_size)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
print('Training...')
for k in tqdm(range(0, steps*epoch)):
x_batch, y_batch = sess.run([image, annotation])
summary, _ = sess.run([merged, self.train_step], feed_dict={self.x_image: x_batch, self.y_: y_batch, self.lr:self.lr_value, self.m:self.m_value, self.keep_prob: 0.5})
k = k + 1
train_writer.add_summary(summary, k)
cur.close()
conn.close()
save_path = saver.save(sess, "trained_model/%s/Drone_CNN.ckpt" % self.model_name)
print('Model saved in file: %s' % save_path)
f_log.close()
train_writer.close()
def drone_prediction(self, mode, window_sliding_stride=8, interest_label='Building'):
conn, cur = data.get_db_connection()
drone_dir = data.get_drone_dir_all()
y_conv_argmax = tf.argmax(self.y_conv, 1)
y_conv_softmax= tf.nn.softmax(self.y_conv)
with self.restore() as sess:
#for each drone ortho-images
for row in drone_dir:
print('Current Dataset: %s (%s)' % (row[0], mode))
curr_image = cv2.imread(row[1])
x_batch_drone = []
k = 0
prob_list = []
curr_image_h = len(curr_image)
curr_image_w = len(curr_image[0])
result_image_h = curr_image_h - curr_image_h%window_sliding_stride - self.input_patch_size
result_image_w = curr_image_w - curr_image_w%window_sliding_stride - self.input_patch_size
#For each patch...
for i in range(0, result_image_h, window_sliding_stride):
for j in range(0, result_image_w, window_sliding_stride):
k = k + 1
patch = np.array(curr_image[i:i+self.input_patch_size, j:j+self.input_patch_size])
x_batch_drone.append(patch)
#Run tensorflow with 128 patches
if k%self.batch_size==0:
self.run_prediction(sess, x_batch_drone, y_conv_argmax, y_conv_softmax, mode, interest_label, prob_list)
x_batch_drone = []
#Run prediction for rest of data
self.run_prediction(sess, x_batch_drone, y_conv_argmax, y_conv_softmax, mode, interest_label, prob_list)
result = np.reshape(prob_list, (int(result_image_h/window_sliding_stride), int(result_image_w/window_sliding_stride)))
if mode == 'PROB_OF_INTEREST':
result = cv2.resize(result, (len(curr_image[0]), len(curr_image)), interpolation=cv2.INTER_LINEAR) * 255
cv2.imwrite('result_%s_%s.png' % (row[0], mode), result)
print('Prediction Complete.')
def drone_prediction(self):
y_conv_argmax = tf.argmax(self.y_conv, 1)
conn, cur = data.get_db_connection()
drone_dir = data.get_drone_dir_all()
with self.restore() as sess:
for i in range(0, len(drone_dir)):
curr_dataset_name = drone_dir[i][0]
print('Current Dataset: %s' % curr_dataset_name)
end_idx = data.get_patch_num(curr_dataset_name)
for start_idx in range(0, end_idx, 20):
x_batch_drone = data.make_batch_drone(conn, cur, start_idx, 20)
y_prediction = sess.run(y_conv_argmax, feed_dict={self.x_image:x_batch_drone, self.keep_prob: 1.0})
'''
for pred_result in y_prediction:
if pred_result == 0:
y_label = 'Building'
elif pred_result== 1:
y_label = 'Road'
elif pred_result== 2:
y_label = 'Otherwise'
else:
y_label = "???"
print(y_label)
'''
f, axarr = plt.subplots(2, 10)
f.suptitle('Prediction Result of %s Dataset' % curr_dataset_name)
for i in range(0, 20):
if y_prediction[i] == 0:
y_label = 'Building'
elif y_prediction[i] == 1:
y_label = 'Road'
elif y_prediction[i] == 2:
y_label = 'Otherwise'
else:
y_label = "???"
axarr[0, i].imshow(x_batch_drone[i]) if i < 10 else axarr[1, i-10].imshow(x_batch_drone[i])
axarr[0, i].set_title(y_label) if i < 10 else axarr[1, i-10].set_title(y_label)
plt.show()
def build_tfrecords(target_dataset, augment_op=False):
conn, cur = data.get_db_connection()
filename_pairs, _, _ = data.get_patch_all(conn, cur, target_dataset)
tfrecords_filename = '/media/lsmjn/56fcc20e-a0ee-45e0-8df1-bf8b2e9a43b2/tfrecords/NGII_%s.tfrecords' % target_dataset
writer = tf.python_io.TFRecordWriter(tfrecords_filename)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for img_path, annotation_path in tqdm(filename_pairs):
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
annotation = cv2.imread(annotation_path)
annotation = cv2.cvtColor(annotation, cv2.COLOR_BGR2RGB)
if augment_op:
augmented_img_list = augment_image(img, sess)
augmented_annotation_list = augment_image(annotation, sess)
for i in range(0, 7):
img_raw = augmented_img_list[i].tostring()
annotation_raw = augmented_annotation_list[i][:, :,
0].tostring()
example = tf.train.Example(features=tf.train.Features(
feature={
'image_raw': _bytes_feature(img_raw),
'mask_raw': _bytes_feature(annotation_raw)
}))
writer.write(example.SerializeToString())
else:
img_raw = img.tostring()
annotation_raw = annotation[:, :, 0].tostring()
example = tf.train.Example(features=tf.train.Features(
feature={
'image_raw': _bytes_feature(img_raw),
'mask_raw': _bytes_feature(annotation_raw)
}))
writer.write(example.SerializeToString())
writer.close()
def create_test_patches(self):
y_conv_argmax = tf.argmax(self.y_conv, 1)
with self.restore() as sess:
conn, cur = data.get_db_connection()
x_batch_test, _, y_batch_test = data.make_batch(conn, cur, 'test', 20)
y_prediction, test_accuracy = sess.run([y_conv_argmax, self.accuracy], feed_dict={self.x_image:x_batch_test, self.y_:y_batch_test, self.keep_prob: 1.0})
f, axarr = plt.subplots(2, 10)
f.suptitle('Test Result of Test Dataset (Accuracy: %f)' % test_accuracy)
for i in range(0, 20):
if y_prediction[i] == 0:
y_label = 'Building'
elif y_prediction[i] == 1:
y_label = 'Road'
elif y_prediction[i] == 2:
y_label = 'Otherwise'
else:
y_label = "???"
axarr[0, i].imshow(x_batch_test[i]) if i < 10 else axarr[1, i-10].imshow(x_batch_test[i])
axarr[0, i].set_title(y_label) if i < 10 else axarr[1, i-10].set_title(y_label)
plt.show()
ground_truth = cv2.imread(y_index_filename_list[i])
ground_truth = y_index_one_hot(ground_truth)
prediction = cv2.imread(
p_filename_list[i][interest_category])[:, :,
interest_category]
prediction = prediction > t
n_TP_sum = n_TP_sum + np.sum((ground_truth == True)
& (prediction == True))
n_FP_sum = n_FP_sum + np.sum((ground_truth == False)
& (prediction == True))
n_FN_sum = n_FN_sum + np.sum((ground_truth == True)
& (prediction == False))
recall = n_TP_sum / (n_TP_sum + n_FN_sum)
precision = n_TP_sum / (n_TP_sum + n_FP_sum)
line = '%f,%f,%f\n' % (t, recall, precision)
print(line)
f.write(line)
if __name__ == '__main__':
conn, cur = data.get_db_connection()
patch_filename_list, _, _ = data.get_patch_all(conn, cur, 'test')
x_filename_list, y_index_filename_list, p_filename_list = create_prediction_result_patches(
patch_filename_list)
for i in range(0, 3):
recall_precision(y_index_filename_list, p_filename_list, i)
def train(self, epoch):
os.makedirs('tb/%s' % self.model_name)
os.makedirs('trained_model/%s' % self.model_name)
tf.summary.image('input x_image', self.x_image)
tf.summary.scalar('train_accuracy', self.accuracy)
tf.summary.scalar('cross_entropy', self.cross_entropy)
tf.summary.scalar('learning rate', self.lr)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
saver = tf.train.Saver()
f_log = open('trained_model/%s/log.csv' % self.model_name, 'w')
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('/home/lsmjn/Drone-CNN/tb/%s' % self.model_name, sess.graph)
ngii_dir_training = data.get_ngii_dir('training')
ngii_dir_test = data.get_ngii_dir('test')
conn, cur = data.get_db_connection()
steps = data.get_steps(self.batch_size)
k = 0
print('\nCurrent Model: %s' % self.model_name)
for i in range(0, epoch):
for j in range(0, steps):
x_batch, y_batch, _ = data.make_batch(conn, cur, 'training', self.batch_size, 'Building')
if k%10 == 0:
print('\nstep %d, epoch %d' % (k, i))
train_xe, train_accuracy = sess.run([self.cross_entropy, self.accuracy], feed_dict={self.x_image:x_batch, self.y_:y_batch, self.keep_prob: 1.0})
print('Train XE:')
print(train_xe)
print('Train Accuracy:')
print(train_accuracy)
for l in range(0, len(ngii_dir_test)):
dataset_test_name = ngii_dir_test[l][0]
x_batch_test, y_batch_test, _ = data.make_batch(conn, cur, 'test', self.batch_size, 'Building')
test_accuracy = sess.run(self.accuracy, feed_dict={self.x_image:x_batch_test, self.y_:y_batch_test, self.keep_prob: 1.0})
print('Test Accuracy:')
print(test_accuracy)
f_log.write('%d,%d,%f,%f,%f\n' % (i, k, train_xe, train_accuracy, test_accuracy))
if k%self.lr_decay_freq == 0:
self.lr_value = self.lr_value * 0.1
print('Learning rate:')
print(self.lr_value)
summary, _ = sess.run([merged, self.train_step], feed_dict={self.x_image: x_batch, self.y_: y_batch, self.lr:self.lr_value, self.m:self.m_value, self.keep_prob: 0.5})
#sess.run(self.train_step, feed_dict={self.x_image: x_batch, self.y_: y_batch, self.lr:self.lr_value, self.m:self.m_value, self.keep_prob: 0.5})
train_writer.add_summary(summary, k)
k = k + 1
cur.close()
conn.close()
save_path = saver.save(sess, "trained_model/%s/Drone_CNN.ckpt" % self.model_name)
print('Model saved in file: %s' % save_path)
#train_writer.close()
f_log.close()
