def eger(cfg):
gen = data_gen.get_batch(batch_size=cfg.batch_size)
images, true_box, true_label = next(gen)
print(true_label)
loct, conft = np_utils.get_loc_conf(true_box, true_label, batch_size=cfg.batch_size)
get_loss(images, conft, loct)
def train():
img = tf.placeholder(shape=[config.batch_size, config.Config['min_dim'], config.Config['min_dim'], 3], dtype=tf.float32)
#ig = AddCoords(x_dim=512,y_dim=512)(img)
anchors_num = sum(
[config.Config['feature_maps'][s] ** 2 * config.Config['aspect_num'][s] for s in range(5)])
loc = tf.placeholder(shape=[config.batch_size, anchors_num, 4], dtype=tf.float32)
conf = tf.placeholder(shape=[config.batch_size, anchors_num], dtype=tf.float32)
pred_loc, pred_confs, vbs = retinanet.model(img,config)
train_tensors = get_loss(conf, loc, pred_loc, pred_confs,config)
gen = data_gen.get_batch_inception(batch_size=config.batch_size,image_size=config.Config['min_dim'],max_detect=50)
global_step = slim.get_or_create_global_step()
lr = tf.train.exponential_decay(
learning_rate=0.001,
global_step=global_step,
decay_steps=40000,
decay_rate=0.7,
staircase=True)
tf.summary.scalar('lr', lr)
sum_op = tf.summary.merge_all()
optimizer = tf.train.MomentumOptimizer(learning_rate=lr,momentum=0.9)
train_op = slim.learning.create_train_op(train_tensors, optimizer)
vbs = []
for s in slim.get_variables():
print(s.name)
if 'resnet_v2_50' in s.name and 'Momentum' not in s.name:
print(s.name)
vbs.append(s)
saver = tf.train.Saver(vbs)
def restore(sess):
saver.restore(sess, config.check_dir)
sv = tf.train.Supervisor(logdir=config.save_dir, summary_op=None, init_fn=restore)
with sv.managed_session() as sess:
for step in range(200000):
print(' '+' '.join(['*']*(step%10)))
images, true_box, true_label = q.get()
loct, conft = np_utils.get_loc_conf(true_box, true_label, batch_size=config.batch_size,cfg=config.Config)
feed_dict = {img: images, loc: loct,
conf: conft}
ls, step = sess.run([train_op, global_step], feed_dict=feed_dict)
if step % 10 == 0:
print('step:' + str(step) +
' ' + 'class_loss:' + str(ls[0]) +
' ' + 'loc_loss:' + str(ls[1])
)
summaries = sess.run(sum_op, feed_dict=feed_dict)
sv.summary_computed(sess, summaries)
def train():
img = tf.placeholder(shape=[
config.batch_size, config.Config['min_dim'], config.Config['min_dim'],
3
],
dtype=tf.float32)
anchors_num = sum([
config.Config['feature_maps'][s]**2 * config.Config['aspect_num'][s]
for s in range(6)
])
loc = tf.placeholder(shape=[config.batch_size, anchors_num, 4],
dtype=tf.float32)
conf = tf.placeholder(shape=[config.batch_size, anchors_num],
dtype=tf.float32)
pred_loc, pred_confs, vbs = inception_500_ince.inception_v2_ssd(
img, config)
train_tensors, sum_op = get_loss(conf, loc, pred_loc, pred_confs, config)
gen = data_gen.get_batch_inception(batch_size=config.batch_size,
image_size=config.Config['min_dim'],
max_detect=50)
optimizer = tf.train.MomentumOptimizer(learning_rate=0.001, momentum=0.9)
train_op = slim.learning.create_train_op(train_tensors, optimizer)
saver = tf.train.Saver(vbs)
def restore(sess):
saver.restore(sess, '/home/dsl/all_check/inception_v2.ckpt')
sv = tf.train.Supervisor(logdir='/home/dsl/all_check/face_detect/voc-1',
summary_op=None,
init_fn=restore)
with sv.managed_session() as sess:
for step in range(1000000000):
images, true_box, true_label = q.get()
loct, conft = np_utils.get_loc_conf(true_box,
true_label,
batch_size=config.batch_size,
cfg=config.Config)
feed_dict = {img: images, loc: loct, conf: conft}
ls = sess.run(train_op, feed_dict=feed_dict)
if step % 10 == 0:
summaries = sess.run(sum_op, feed_dict=feed_dict)
sv.summary_computed(sess, summaries)
print(ls)
def train():
img = tf.placeholder(shape=[config.batch_size, config.Config['min_dim'], config.Config['min_dim'], 3], dtype=tf.float32)
anchors_num = sum(
[config.Config['feature_maps'][s] ** 2 * config.Config['aspect_num'][s] for s in range(6)])
loc = tf.placeholder(shape=[config.batch_size, anchors_num, 4], dtype=tf.float32)
conf = tf.placeholder(shape=[config.batch_size, anchors_num], dtype=tf.float32)
pred_loc, pred_confs, vbs = inceptionv3_500_ince.inception_v2_ssd(img,config)
train_tensors = get_loss(conf, loc, pred_loc, pred_confs,config)
global_step = get_or_create_global_step()
# Define your exponentially decaying learning rate
lr = tf.train.exponential_decay(
learning_rate=0.001,
global_step=global_step,
decay_steps=20000,
decay_rate=0.7,
staircase=True)
tf.summary.scalar('lr',lr)
sum_op = tf.summary.merge_all()
gen = data_gen.get_batch_inception(batch_size=config.batch_size,image_size=config.Config['min_dim'],max_detect=50)
optimizer = tf.train.MomentumOptimizer(learning_rate=lr,momentum=0.9)
train_op = slim.learning.create_train_op(train_tensors, optimizer)
saver = tf.train.Saver(vbs)
def restore(sess):
saver.restore(sess, '/home/dsl/all_check/inception_v3.ckpt')
sv = tf.train.Supervisor(logdir='/home/dsl/all_check/face_detect/voc-v32', summary_op=None, init_fn=restore)
with sv.managed_session() as sess:
for step in range(1000000000):
images, true_box, true_label = q.get()
loct, conft = np_utils.get_loc_conf(true_box, true_label, batch_size=config.batch_size,cfg=config.Config)
feed_dict = {img: images, loc: loct,
conf: conft}
t = time.time()
ls,step = sess.run([train_op,global_step], feed_dict=feed_dict)
if step % 10 == 0:
print(time.time()-t)
summaries = sess.run(sum_op, feed_dict=feed_dict)
sv.summary_computed(sess, summaries)
print(ls)
def train():
img = tf.placeholder(
shape=[config.batch_size, cfg["min_dim"], cfg["min_dim"], 3], dtype=tf.float32
)
anchors_num = sum(
[cfg["feature_maps"][s] ** 2 * cfg["aspect_num"][s] for s in range(6)]
)
loc = tf.placeholder(shape=[config.batch_size, anchors_num, 4], dtype=tf.float32)
conf = tf.placeholder(shape=[config.batch_size, anchors_num], dtype=tf.float32)
pred_loc, pred_confs, vbs = mobile.nana_mobile(img, config)
train_tensors, sum_op = get_loss(conf, loc, pred_loc, pred_confs, config)
gen = data_gen.get_batch(batch_size=config.batch_size, image_size=cfg["min_dim"])
optimizer = tf.train.MomentumOptimizer(learning_rate=0.01, momentum=0.9)
train_op = slim.learning.create_train_op(train_tensors, optimizer)
saver = tf.train.Saver(vbs)
def restore(sess):
saver.restore(sess, "/home/dsl/all_check/nasnet-a_mobile_04_10_2017/model.ckpt")
sv = tf.train.Supervisor(
logdir="/home/dsl/all_check/face_detect/nana", summary_op=None, init_fn=restore
)
with sv.managed_session() as sess:
for step in range(1000000000):
images, true_box, true_label = next(gen)
loct, conft = np_utils.get_loc_conf(
true_box, true_label, batch_size=config.batch_size, cfg=cfg
)
feed_dict = {img: images, loc: loct, conf: conft}
ls = sess.run(train_op, feed_dict=feed_dict)
if step % 10 == 0:
summaries = sess.run(sum_op, feed_dict=feed_dict)
sv.summary_computed(sess, summaries)
print(ls)
def train(cfg):
img = tf.placeholder(shape=[cfg.batch_size, 300, 300, 3], dtype=tf.float32)
#boxs = tf.placeholder(shape=[batch_size, 50, 4], dtype=tf.float32)
#label = tf.placeholder(shape=[batch_size, 50], dtype=tf.int32)
loc = tf.placeholder(shape=[cfg.batch_size, 7512,4], dtype=tf.float32)
conf = tf.placeholder(shape=[cfg.batch_size, 7512], dtype=tf.float32)
pred_loc, pred_confs, vbs = model(img)
train_tensors,sum_op = get_loss(conf,loc,pred_loc, pred_confs)
gen = data_gen.get_batch(batch_size=cfg.batch_size)
optimizer = tf.train.MomentumOptimizer(learning_rate=0.001, momentum=0.9)
train_op = slim.learning.create_train_op(train_tensors, optimizer)
saver = tf.train.Saver(vbs)
def restore(sess):
saver.restore(sess,'/home/dsl/all_check/vgg_16.ckpt')
sv = tf.train.Supervisor(logdir='/home/dsl/all_check/face_detect', summary_op=None, init_fn=restore)
with sv.managed_session() as sess:
for step in range(1000000000):
images, true_box, true_label = next(gen)
loct,conft = np_utils.get_loc_conf(true_box,true_label,batch_size=cfg.batch_size)
feed_dict = {img: images, loc: loct,
conf: conft}
ls = sess.run(train_op, feed_dict=feed_dict)
if step%10==0:
summaries = sess.run(sum_op,feed_dict=feed_dict)
sv.summary_computed(sess, summaries)
print(ls)
#train()
#tf.enable_eager_execution()
#eger()
#detect('/media/dsl/20d6b919-92e1-4489-b2be-a092290668e4/VOCdevkit/VOCdevkit/VOC2007/JPEGImages/000133.jpg')
#video()