def __init__(self,model_path,sizes = [12,24,48]):
self.sizes = sizes
# load network
self.net_12 = model.detect_12Net(is_train = False,size = (sizes[0],sizes[0],3))
self.net_24 = model.detect_24Net(is_train = False,size = (sizes[1],sizes[1],3))
self.net_48 = model.detect_48Net(is_train = False)
# create session
self.sess = tf.Session()
self.restore(model_path)
restorer_12_calib = tf.train.Saver(
[v for v in tf.global_variables() if "12calib_" in v.name])
restorer_12_calib.restore(sess, param.model_dir + "12-calib-net.ckpt")
if sys.argv[1] == str(param.img_size_48):
#24net
input_24_node = tf.placeholder(
"float",
[None, param.img_size_24, param.img_size_24, param.input_channel])
from_12_node = tf.placeholder("float", [None, 16])
target_24_node = tf.placeholder("float", [None, 1])
inputs_24 = np.zeros((param.mini_batch, param.img_size_24,
param.img_size_24, param.input_channel), np.float32)
net_24 = model.detect_24Net(input_24_node, target_24_node, from_12_node)
net_24_calib = model.calib_24Net(input_24_node, target_24_node)
restorer_24 = tf.train.Saver(
[v for v in tf.global_variables() if "24det_" in v.name])
restorer_24.restore(sess, param.model_dir + "24-net.ckpt")
restorer_24_calib = tf.train.Saver(
[v for v in tf.global_variables() if "24calib_" in v.name])
restorer_24_calib.restore(sess, param.model_dir + "24-calib-net.ckpt")
neg_file_list = [f for f in os.listdir(param.neg_dir) if f.endswith(".jpg")]
#hard neg mining
neg_db_sz = 0
neg_db = [0 for _ in range(1000)]
for nid, img_name in enumerate(neg_file_list):
def train_det_net():
# get all training sample
data_info = parse_data_info(only_positive = False)
# data_info = [<image-path str>,[<nonface/face int>,<pattern-id int>]]
# training configuration
batch = 500
size = (48,48,3)
start_epoch = 0
end_epoch = 1000
train_validation_rate = 0.9 # training set / all sample
# load the pretrained model , set None if you don't have
pretrained = 'models/48_net_6.ckpt'
# load data iterater
dataset = DataSet(data_info,train_rate = train_validation_rate)
_ , train_op , val_op , next_ele = dataset.get_iterator(batch,size)
# load network
# learning rate is great impact in training models
net_12 = model.detect_12Net(lr = 0.001,size = (12,12,3))
net_24 = model.detect_24Net(lr = 0.001,size = (24,24,3))
net_48 = model.detect_48Net(lr = 0.001,size = (48,48,3))
sess = tf.InteractiveSession()
saver = tf.train.Saver()
if pretrained:
saver.restore(sess , pretrained)
else:
sess.run(tf.global_variables_initializer())
for epoch in xrange(start_epoch,end_epoch):
loss = 0
iteration = 0
sess.run(train_op)
# get each element of the training dataset until the end is reached
while True:
try:
# default of the size returned from data iterator is 48
inputs,clss ,pattern = sess.run(next_ele)
# <ndarray> , <0/1> , <one-hot of 45-class>
clss = clss.reshape(batch,2)
pattern = pattern.reshape(batch,45)
# resize image to fit each net
inputs_12 = np.array([cv2.resize(img,(net_12.size[0],net_12.size[1])) for img in inputs])
inputs_24 = np.array([cv2.resize(img,(net_24.size[0],net_24.size[1])) for img in inputs])
inputs_48 = np.array([cv2.resize(img,(net_48.size[0],net_48.size[1])) for img in inputs])
# forward 12net
net_12_fc = net_12.get_fc(inputs_12)
# forward 24net
net_24_fc = net_24.get_fc(inputs_24,net_12_fc)
train_nets = [net_12,net_24,net_48]
net_feed_dict = {net_12.inputs:inputs_12 , net_12.targets:clss,\
net_24.inputs:inputs_24 , net_24.targets:clss,net_24.from_12:net_12_fc,\
net_48.inputs:inputs_48 , net_48.targets:clss,net_48.from_24:net_24_fc}
# training net
sess.run([net.train_step for net in train_nets],\
feed_dict = net_feed_dict)
# loss computation
losses = sess.run([net.loss for net in train_nets],\
feed_dict = net_feed_dict)
if iteration % 100 == 0:
net_12_eva = net_12.evaluate(inputs_12,clss)
net_12_acc = sum(net_12_eva)/len(net_12_eva)
net_24_eva = net_24.evaluate(inputs_24,clss,net_12_fc)
net_24_acc = sum(net_24_eva)/len(net_24_eva)
net_48_eva = net_48.evaluate(inputs_48,clss,net_24_fc)
net_48_acc = sum(net_48_eva)/len(net_48_eva)
print ('Training Epoch {} --- Iter {} --- Training Accuracy: {}%,{}%,{}% --- Training Loss: {}'\
.format(epoch , iteration , net_12_acc , net_24_acc , net_48_acc , losses))
iteration += 1
except tf.errors.OutOfRangeError:
# print("End of training dataset.")
break
# get each element of the validation dataset until the end is reached
sess.run(val_op)
net_12_acc = []
net_24_acc = []
net_48_acc = []
while True:
try:
# the size returned from data iterator is 48
inputs,clss ,pattern = sess.run(next_ele)
clss = clss.reshape(batch,2)
# resize image to fit each net
inputs_12 = np.array([cv2.resize(img,(net_12.size[0],net_12.size[1])) for img in inputs])
inputs_24 = np.array([cv2.resize(img,(net_24.size[0],net_24.size[1])) for img in inputs])
inputs_48 = np.array([cv2.resize(img,(net_48.size[0],net_48.size[1])) for img in inputs])
# forward 12net
net_12_fc = net_12.get_fc(inputs_12)
# forward 24net
net_24_fc = net_24.get_fc(inputs_24,net_12_fc)
net_12_eva = net_12.evaluate(inputs_12,clss)
net_24_eva = net_24.evaluate(inputs_24,clss,net_12_fc)
net_48_eva = net_48.evaluate(inputs_48,clss,net_24_fc)
for i in range(len(net_12_eva)):
net_12_acc.append(net_12_eva[i])
net_24_acc.append(net_24_eva[i])
net_48_acc.append(net_48_eva[i])
except tf.errors.OutOfRangeError:
# print("End of validation dataset.")
break
print ('Validation Epoch {} Validation Accuracy: {}%,{}%,{}%'\
.format(epoch , sum(net_12_acc)/len(net_12_acc),\
sum(net_24_acc)/len(net_24_acc),\
sum(net_48_acc)/len(net_48_acc)))
saver = tf.train.Saver()
save_path = saver.save(sess, "models/48_net_{}.ckpt".format(epoch))
print ("Model saved in file: ", save_path)