Skip to content

Commit c0260bf

Browse files
author
Ubuntu
committed
Latest Code
1 parent c11e760 commit c0260bf

File tree

2 files changed

+49
-48
lines changed

2 files changed

+49
-48
lines changed

‎main.py

+9-11
Original file line numberDiff line numberDiff line change
@@ -3,22 +3,20 @@
33
from model import *
44

55

6-
weights_path = '/Users/shashank/Tensorflow/SPN/weights/'
7-
imgs_path = '/Users/shashank/Tensorflow/CSE252C-Hyperface/git/truth_data.npy'
86

9-
if not os.path.exists('./logs'):
10-
os.makedirs('./logs')
7+
if not os.path.exists('../logs'):
8+
os.makedirs('../logs')
119

12-
if not os.path.exists('../../checkpoint'):
13-
os.makedirs('../../checkpoint')
10+
if not os.path.exists('../checkpoint'):
11+
os.makedirs('../checkpoint')
1412

15-
if not os.path.exists('../../best_checkpoint'):
16-
os.makedirs('../../best_checkpoint')
13+
if not os.path.exists('../best_checkpoint'):
14+
os.makedirs('../best_checkpoint')
1715

18-
map(os.unlink, (os.path.join( './logs',f) for f in os.listdir('./logs')) )
16+
map(os.unlink, (os.path.join( '../logs',f) for f in os.listdir('../logs')) )
1917

20-
net = HyperFace(False, tf_record_file_path='../../aflw_train_new.tfrecords',model_save_path='../../checkpoint/',best_model_save_path='../../best_checkpoint/',
21-
restore_model_path='../../saved_best_checkpoint')
18+
net = HyperFace(True, tf_record_file_path='../../aflw_train_new.tfrecords',model_save_path='../checkpoint/',best_model_save_path='../best_checkpoint/',
19+
restore_model_path='../full_best_checkpoint/')
2220

2321
with tf.Session() as sess:
2422
print 'Building Graph...'

‎model.py

+40-37
Original file line numberDiff line numberDiff line change
@@ -14,7 +14,7 @@ def __init__(self,load_model,tf_record_file_path=None,model_save_path=None,best_
1414
self.img_width = 227
1515
self.channel = 3
1616

17-
self.num_epochs = 2
17+
self.num_epochs =10
1818

1919
# Hyperparameters 1,5,0.5,5,2
2020
self.weight_detect = 1
@@ -43,33 +43,34 @@ def build_network(self, sess):
4343

4444
self.X = tf.placeholder(tf.float32, [self.batch_size, self.img_height, self.img_width, self.channel], name='images')
4545
self.detection = tf.placeholder(tf.int32, [self.batch_size], name='detection')
46-
# self.landmarks = tf.placeholder(tf.float32, [self.batch_size, 42], name='landmarks')
47-
# self.visibility = tf.placeholder(tf.float32, [self.batch_size,21], name='visibility')
48-
# self.pose = tf.placeholder(tf.float32, [self.batch_size,3], name='pose')
49-
# self.gender = tf.placeholder(tf.int32, [self.batch_size], name='gender')
46+
self.landmarks = tf.placeholder(tf.float32, [self.batch_size, 42], name='landmarks')
47+
self.visibility = tf.placeholder(tf.float32, [self.batch_size,21], name='visibility')
48+
self.pose = tf.placeholder(tf.float32, [self.batch_size,3], name='pose')
49+
self.gender = tf.placeholder(tf.int32, [self.batch_size], name='gender')
5050

51-
net_output = self.network_det(self.X) # (out_detection, out_landmarks, out_visibility, out_pose, out_gender)
51+
net_output = self.network(self.X) # (out_detection, out_landmarks, out_visibility, out_pose, out_gender)
52+
self.test_model = net_output
53+
self.loss_detection = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=net_output[0], labels=tf.one_hot(self.detection, 2)))
5254

53-
self.loss_detection = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=net_output, labels=tf.one_hot(self.detection, 2)))
54-
55-
# detection_mask = tf.cast(tf.expand_dims(self.detection, axis=1),tf.float32)
55+
detection_mask = tf.cast(tf.expand_dims(self.detection, axis=1),tf.float32)
5656

57-
# visibility_mask = tf.reshape(tf.tile(tf.expand_dims(self.visibility, axis=2), [1,1,2]), [self.batch_size, -1])
58-
# self.loss_landmarks = tf.reduce_mean(tf.square(detection_mask*visibility_mask*(net_output[1] - self.landmarks)))
57+
visibility_mask = tf.reshape(tf.tile(tf.expand_dims(self.visibility, axis=2), [1,1,2]), [self.batch_size, -1])
58+
self.loss_landmarks = tf.reduce_mean(tf.square(detection_mask*visibility_mask*(net_output[1] - self.landmarks)))
5959

60-
# self.loss_visibility = tf.reduce_mean(tf.square(detection_mask*(net_output[2] - self.visibility)))
61-
# self.loss_pose = tf.reduce_mean(tf.square(detection_mask*(net_output[3] - self.pose)))
62-
# self.loss_gender = tf.reduce_mean(detection_mask*tf.nn.sigmoid_cross_entropy_with_logits(logits=net_output[4], labels=tf.one_hot(self.gender,2)))
60+
self.loss_visibility = tf.reduce_mean(tf.square(detection_mask*(net_output[2] - self.visibility)))
61+
self.loss_pose = tf.reduce_mean(tf.square(detection_mask*(net_output[3] - self.pose)))
62+
self.loss_gender = tf.reduce_mean(detection_mask*tf.nn.sigmoid_cross_entropy_with_logits(logits=net_output[4], labels=tf.one_hot(self.gender,2)))
6363

6464

65-
# self.loss = self.weight_detect*self.loss_detection + self.weight_landmarks*self.loss_landmarks \
66-
# + self.weight_visibility*self.loss_visibility + self.weight_pose*self.loss_pose \
67-
# + self.weight_gender*self.loss_gender
65+
self.loss = self.weight_detect*self.loss_detection + self.weight_landmarks*self.loss_landmarks \
66+
+ self.weight_visibility*self.loss_visibility + self.weight_pose*self.loss_pose \
67+
+ self.weight_gender*self.loss_gender
6868

69-
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.cast(tf.argmax(net_output,1),tf.int32),self.detection),tf.float32))
69+
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.cast(tf.argmax(net_output[0],1),tf.int32),self.detection),tf.float32))
7070

71-
self.loss = self.loss_detection
72-
self.optimizer = tf.train.AdamOptimizer().minimize(self.loss)
71+
#self.loss = self.loss_detection
72+
#self.optimizer = tf.train.AdamOptimizer(1e-7).minimize(self.loss)
73+
self.optimizer = tf.train.MomentumOptimizer(1e-3,0.9,use_nesterov=True).minimize(self.loss)
7374
self.saver = tf.train.Saver(max_to_keep=4, keep_checkpoint_every_n_hours=4)
7475
self.best_saver = tf.train.Saver(max_to_keep=10, keep_checkpoint_every_n_hours=4)
7576

@@ -87,21 +88,21 @@ def train(self):
8788
print "Initializing Model"
8889
self.sess.run(tf.group(tf.global_variables_initializer(),tf.local_variables_initializer()))
8990

90-
# self.load_det_weights(self.best_model_save_path+'weights_0.npy')
91-
91+
#self.load_det_weights(self.restore_model_path+'weights.npy')
92+
9293

9394
coord = tf.train.Coordinator()
9495
threads = tf.train.start_queue_runners(sess=self.sess,coord=coord)
9596

96-
writer = tf.summary.FileWriter('./logs', self.sess.graph)
97+
writer = tf.summary.FileWriter('../logs', self.sess.graph)
9798
loss_summ = tf.summary.scalar('loss', self.loss)
9899
img_summ = tf.summary.image('images', self.images, max_outputs=5)
99-
# label_summ = tf.summary.histogram('labels', self.detection)
100-
# detect_summ = tf.summary.scalar('det_loss', self.loss_detection)
101-
# landmarks_summ = tf.summary.scalar('landmarks_loss', self.loss_landmarks)
102-
# vis_summ = tf.summary.scalar('visibility_loss', self.loss_visibility)
103-
# pose_summ = tf.summary.scalar('pose_loss', self.loss_pose)
104-
# gender_summ = tf.summary.scalar('gender_loss', self.loss_gender)
100+
label_summ = tf.summary.histogram('labels', self.detection)
101+
detect_summ = tf.summary.scalar('det_loss', self.loss_detection)
102+
landmarks_summ = tf.summary.scalar('landmarks_loss', self.loss_landmarks)
103+
vis_summ = tf.summary.scalar('visibility_loss', self.loss_visibility)
104+
pose_summ = tf.summary.scalar('pose_loss', self.loss_pose)
105+
gender_summ = tf.summary.scalar('gender_loss', self.loss_gender)
105106

106107
summ_op = tf.summary.merge_all()
107108

@@ -111,24 +112,26 @@ def train(self):
111112
while not coord.should_stop():
112113
batch_imgs, batch_labels, batch_landmarks, batch_visibility, batch_pose, batch_gender = self.sess.run([self.images,self.labels,self.land, self.vis, self.po, self.gen])
113114
batch_imgs = (batch_imgs - 127.5) / 128.0
114-
# input_feed={self.X: batch_imgs, self.detection: batch_labels, self.landmarks: batch_landmarks, self.visibility: batch_visibility, self.pose: batch_pose, self.gender: np.squeeze(batch_gender)}
115-
input_feed={self.X: batch_imgs, self.detection: batch_labels}
115+
input_feed={self.X: batch_imgs, self.detection: batch_labels, self.landmarks: batch_landmarks, self.visibility: batch_visibility, self.pose: batch_pose, self.gender: np.squeeze(batch_gender)}
116+
#input_feed={self.X: batch_imgs, self.detection: batch_labels}
116117

117-
_,loss, summ, accuracy = self.sess.run([self.optimizer, self.loss, summ_op, self.accuracy], input_feed)
118+
_,model_op,loss,l_d,l_l,l_v,l_p,l_g, summ, accuracy = self.sess.run([self.optimizer,self.test_model,self.loss,self.loss_detection,
119+
self.loss_landmarks,self.loss_visibility,self.loss_pose,self.loss_gender, summ_op, self.accuracy], input_feed)
118120

119121
writer.add_summary(summ, counter)
120122

121123
if counter % self.save_after_steps == 0:
122-
self.saver.save(self.sess,self.model_save_path+'statefarm_model',global_step=int(counter),write_meta_graph=False)
124+
self.saver.save(self.sess,self.model_save_path+'hyperface_model',global_step=int(counter),write_meta_graph=False)
123125

124126

125127
if loss <= best_loss:
126128
best_loss = loss
127-
self.best_saver.save(self.sess,self.best_model_save_path+'statefarm_best_model',global_step=int(counter),write_meta_graph=False)
128-
self.save_weights(self.best_model_save_path)
129+
self.best_saver.save(self.sess,self.best_model_save_path+'hyperface_best_model',global_step=int(counter),write_meta_graph=False)
130+
#self.save_weights(self.best_model_save_path)
129131

130132
if counter % self.print_after_steps == 0:
131-
print "Iteration:{},Loss:{},Accuracy:{}".format(counter,loss,accuracy)
133+
print "Iteration:{},Total Loss:{},Detection loss:{},Landmark loss:{},Visbility Loss :{},Pose Loss:{},Gender Loss:{},Accuracy:{}".format(counter,loss,l_d,l_l,l_v,l_p,l_g,accuracy)
134+
132135
counter += 1
133136

134137
except tf.errors.OutOfRangeError:
@@ -213,7 +216,7 @@ def network(self,inputs,reuse=False):
213216
fc_gender = slim.fully_connected(fc_full, 512, scope='fc_gender1')
214217

215218
out_detection = slim.fully_connected(fc_detection, 2, scope='fc_detection2', activation_fn = None)
216-
out_landmarks = slim.fully_connected(fc_landmarks, 42, scope='fc_landmarks2', activation_fn = None)
219+
out_landmarks = slim.fully_connected(fc_landmarks, 42, scope='fc_landmarks2', activation_fn = None )
217220
out_visibility = slim.fully_connected(fc_visibility, 21, scope='fc_visibility2', activation_fn = None)
218221
out_pose = slim.fully_connected(fc_pose, 3, scope='fc_pose2', activation_fn = None)
219222
out_gender = slim.fully_connected(fc_gender, 2, scope='fc_gender2', activation_fn = None)

0 commit comments

Comments
 (0)