def testing():
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.6)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
ssd_model = ssd300.SSD300(sess,
False,
class_size=class_size,
graph_config=graph_config)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(var_list=tf.trainable_variables())
if os.path.exists('./session_params/session.ckpt.index'):
saver.restore(sess, './session_params/session.ckpt')
image, actual, file_list = next(dataset.next_test())
pred_class, pred_class_val, pred_location = ssd_model.run(
image, None, file_list)
print('file_list:' + str(file_list))
#
for index, act in zip(range(len(image)), actual):
for a in act:
print('【img-' + str(index) + ' actual】:' + str(a))
print('pred_class:' + str(pred_class[index]))
print('pred_class_val:' + str(pred_class_val[index]))
print('pred_location:' + str(pred_location[index]))
else:
print('No Data Exists!')
sess.close()
def training():
batch_size = FLAGS.batch_size
running_count = 0
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
ssd_model = ssd300.SSD300(sess, True)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(var_list=tf.trainable_variables())
if FLAGS.finetune:
print('Finetune from %s' % FLAGS.finetune)
ckpt = tf.train.get_checkpoint_state(FLAGS.finetune)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
raise ValueError('Failed to load model.')
print('-------------------------')
print('\nStart Training')
min_loss_location = 100000.
min_loss_class = 100000.
while ((min_loss_location + min_loss_class) > 0.001
and running_count < FLAGS.max_steps):
running_count += 1
train_data, actual_data, _ = get_traindata(batch_size)
if len(train_data) > 0:
loss_all, loss_class, loss_location, pred_class, pred_location = ssd_model.run(
train_data, actual_data)
l = np.sum(loss_location)
c = np.sum(loss_class)
if min_loss_location > l:
min_loss_location = l
if min_loss_class > c:
min_loss_class = c
if running_count % FLAGS.log_frequency == 0:
# print('Step: %d | Loss All: %.4f(min) %.4f | Location: %.4f | Class: %.4f | pred_class: %.4f(all) %.4f(max) %.4f(min) | pred_location: %.4f(all) %.4f(max) %.4f(min)'
# %(running_count, min_loss_location + min_loss_class, loss_all, np.sum(loss_location), np.sum(loss_class), np.sum(pred_class), np.amax(pred_class), np.min(pred_class), np.sum(pred_location), np.amax(pred_location), np.min(pred_location)))
print(
'Step: %d | Loss All: %.4f(min) %.4f | Location: %.4f | Class: %.4f'
%
(running_count, min_loss_location + min_loss_class,
loss_all, np.sum(loss_location), np.sum(loss_class)))
# 定期保存ckpt
if running_count % 1000 == 0:
saver.save(sess, FLAGS.model_dir + 'session.ckpt')
print('session.ckpt has been saved.')
gc.collect()
else:
print('No Data Exists!')
break
saver.save(sess, FLAGS.model_dir + 'session.ckpt')
sess.close()
gc.collect()
print('End Training')
def training():
batch_size = 15
running_count = 0
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.9)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
ssd_model = ssd300.SSD300(sess, True)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(var_list=tf.trainable_variables())
if os.path.exists('./session_params/session.ckpt.index'):
print('\nStart Restore')
saver.restore(sess, './session_params/session.ckpt')
print('\nEnd Restore')
print('\nStart Training')
min_loss_location = 100000.
min_loss_class = 100000.
while ((min_loss_location + min_loss_class) > 0.001
and running_count < 100000):
running_count += 1
train_data, actual_data, _ = get_traindata_voc2007(batch_size)
if len(train_data) > 0:
loss_all, loss_class, loss_location, pred_class, pred_location = ssd_model.run(
train_data, actual_data)
l = np.sum(loss_location)
c = np.sum(loss_class)
if min_loss_location > l:
min_loss_location = l
if min_loss_class > c:
min_loss_class = c
print('Running:【' + str(running_count) + '】|Loss All:【' +
str(min_loss_location + min_loss_class) + '|' +
str(loss_all) + '】|Location:【' +
str(np.sum(loss_location)) + '】|Class:【' +
str(np.sum(loss_class)) + '】|pred_class:【' +
str(np.sum(pred_class)) + '|' +
str(np.amax(pred_class)) + '|' +
str(np.min(pred_class)) + '】|pred_location:【' +
str(np.sum(pred_location)) + '|' +
str(np.amax(pred_location)) + '|' +
str(np.min(pred_location)) + '】')
# 定期保存ckpt
if running_count % 100 == 0:
saver.save(sess, './session_params/session.ckpt')
print('session.ckpt has been saved.')
gc.collect()
else:
print('No Data Exists!')
break
saver.save(sess, './session_params/session.ckpt')
sess.close()
gc.collect()
print('End Training')
def testing():
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
ssd_model = ssd300.SSD300(sess, False)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(var_list=tf.trainable_variables())
if os.path.exists(FLAGS.model_dir + 'session.ckpt.index'):
saver.restore(sess, FLAGS.model_dir + 'session.ckpt')
img_path, image_data, actual, file_list = get_traindata(1, False)
pred_class, pred_class_val, pred_location = ssd_model.run(
image_data, None)
print('file_list:' + str(file_list))
image = cv2.imread(img_path)
height = image.shape[0]
width = image.shape[1]
for index, act in zip(range(len(image_data)), actual):
for a in act:
print('【img-' + str(index) + ' actual】:' + str(a))
print('pred_class:' + str(pred_class[index]))
print('pred_class_val:' + str(pred_class_val[index]))
print('pred_location:' + str(pred_location[index]))
for i in range(len(pred_location[index])):
w = int(pred_location[index][i].tolist()[2] * width)
h = int(pred_location[index][i].tolist()[3] * height)
x = max(
0,
int(pred_location[index][i].tolist()[0] * width -
w / 2))
y = max(
0,
int(pred_location[index][i].tolist()[1] * height -
h / 2))
cv2.rectangle(image, (x, y), (x + w, y + h), (255, 0, 0),
2)
cv2.putText(image, lable_arr[pred_class[index][i]],
(x + 3, y + 10), cv2.FONT_HERSHEY_SIMPLEX, .5,
(0, 0, 255), 2)
else:
print('No Data Exists!')
sess.close()
cv2.imshow("image", image)
cv2.waitKey()
cv2.destroyAllWindows()
