def main(argv=None):
if FLAGS.mode == 'test':
FLAGS.batch_size = 1
if FLAGS.mode == 'cmc':
FLAGS.batch_size = 1
learning_rate = tf.placeholder(tf.float32, name='learning_rate')
images = tf.placeholder(
tf.float32, [3, FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3],
name='images')
images_total = tf.placeholder(
tf.float32, [FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3],
name='images_total')
labels = tf.placeholder(tf.float32, [FLAGS.batch_size], name='labels')
is_train = tf.placeholder(tf.bool, name='is_train')
global_step = tf.Variable(0, name='global_step', trainable=False)
weight_decay = 0.0005
tarin_num_id = 0
val_num_id = 0
if FLAGS.mode == 'train':
tarin_num_id = cuhk03_dataset_label2.get_num_id(
FLAGS.data_dir, 'train')
print(
tarin_num_id,
' 11111111111111111111 1111111111111111'
)
elif FLAGS.mode == 'val':
val_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'val')
#images1, images2,images3 = preprocess(images, is_train)
# Create the model and an embedding head.
model = import_module('nets.' + 'mobilenet_v1_1_224')
head = import_module('heads.' + 'fc1024')
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
endpoints, body_prefix = model.endpoints(images_total, is_training=True)
with tf.name_scope('head'):
endpoints = head.head(endpoints, FLAGS.embedding_dim, is_training=True)
print 'model_output : ', endpoints['model_output'] # (bt,2048)
print 'global_pool : ', endpoints['global_pool'] # (bt,2048)
print 'resnet_v1_50/block4 : ', endpoints[
'Conv2d_12_pointwise'] # (bt,7,7,2048)
# see net.resnet_V1 line 258
print ' 1\n'
train_mode = tf.placeholder(tf.bool)
print('Build network')
feat = endpoints['Conv2d_12_pointwise'] # (bt,7,7,2048)
#feat = tf.convert_to_tensor(feat, dtype=tf.float32)
# global
feature = global_pooling(feat, weight_decay)
#loss_triplet,PP,NN = triplet_hard_loss(feature,FLAGS.ID_num,FLAGS.IMG_PER_ID)
loss_triplet, PP, NN = batch_hard_triplet_loss(labels, feature, 0.3)
loss = loss_triplet * FLAGS.global_rate
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
model_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
body_prefix)
#optimizer = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(loss)
#optimizer = tf.train.AdadeltaOptimizer(learning_rate)
#train = optimizer.minimize(loss, global_step=global_step)
# Update_ops are used to update batchnorm stats.
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
#train_op = optimizer.minimize(loss_mean, global_step=global_step)
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
train = optimizer.minimize(loss, global_step=global_step)
lr = FLAGS.learning_rate
#config=tf.ConfigProto(log_device_placement=True)
#config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
# GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.9
with tf.Session(config=config) as sess:
print '\n'
#print model_variables
print '\n'
#sess.run(tf.global_variables_initializer())
#saver = tf.train.Saver()
#checkpoint_saver = tf.train.Saver(max_to_keep=0)
checkpoint_saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
if ckpt and ckpt.model_checkpoint_path:
print('Restore model')
print ckpt.model_checkpoint_path
#saver.restore(sess, ckpt.model_checkpoint_path)
checkpoint_saver.restore(sess, ckpt.model_checkpoint_path)
#for first , training load imagenet
else:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(model_variables)
print FLAGS.initial_checkpoint
saver.restore(sess, FLAGS.initial_checkpoint)
if FLAGS.mode == 'train':
step = sess.run(global_step)
for i in xrange(step, FLAGS.max_steps + 1):
batch_images, batch_labels, batch_images_total = cuhk03_dataset_label2.read_data(
FLAGS.data_dir, 'train', tarin_num_id, IMAGE_WIDTH,
IMAGE_HEIGHT, FLAGS.batch_size, FLAGS.ID_num,
FLAGS.IMG_PER_ID)
feed_dict = {
learning_rate: lr,
is_train: True,
train_mode: True,
images_total: batch_images_total,
labels: batch_labels
}
_, train_loss = sess.run([train, loss], feed_dict=feed_dict)
print('Step: %d, Learning rate: %f, Train loss: %f ' %
(i, lr, train_loss))
gtoloss, gp, gn = sess.run([loss_triplet, PP, NN],
feed_dict=feed_dict)
print 'global hard: ', gtoloss
print 'global P: ', gp
print 'global N: ', gn
#lr = FLAGS.learning_rate / ((2) ** (i/160000)) * 0.1
lr = FLAGS.learning_rate * ((0.0001 * i + 1)**-0.75)
if i % 100 == 0:
#saver.save(sess, FLAGS.logs_dir + 'model.ckpt', i)
checkpoint_saver.save(sess, FLAGS.logs_dir + 'model.ckpt',
i)
elif FLAGS.mode == 'val':
total = 0.
for _ in xrange(10):
batch_images, batch_labels = cuhk03_dataset_label2.read_data(
FLAGS.data_dir, 'val', val_num_id, IMAGE_WIDTH,
IMAGE_HEIGHT, FLAGS.batch_size)
feed_dict = {
images: batch_images,
labels: batch_labels,
is_train: False
}
prediction = sess.run(inference, feed_dict=feed_dict)
prediction = np.argmax(prediction, axis=1)
label = np.argmax(batch_labels, axis=1)
for i in xrange(len(prediction)):
if prediction[i] == label[i]:
total += 1
print('Accuracy: %f' % (total / (FLAGS.batch_size * 10)))
'''
for i in xrange(len(prediction)):
print('Prediction: %s, Label: %s' % (prediction[i] == 0, labels[i] == 0))
image1 = cv2.cvtColor(batch_images[0][i], cv2.COLOR_RGB2BGR)
image2 = cv2.cvtColor(batch_images[1][i], cv2.COLOR_RGB2BGR)
image = np.concatenate((image1, image2), axis=1)
cv2.imshow('image', image)
key = cv2.waitKey(0)
if key == 1048603: # ESC key
break
'''
elif FLAGS.mode == 'cmc':
do_times = 1
cmc_sum = np.zeros((100, 100), dtype='f')
for times in xrange(do_times):
path = 'data'
set = 'train'
cmc_array = np.ones((100, 100), dtype='f')
batch_images = []
batch_labels = []
index_gallery_array = np.ones((1, 100), dtype='f')
gallery_bool = True
probe_bool = True
for j in xrange(100):
id_probe = j
for i in xrange(100):
batch_images = []
batch_labels = []
filepath = ''
#filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, i, index_gallery)
#filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, id_probe, index_probe)
if gallery_bool == True:
while True:
index_gallery = int(random.random() * 10)
index_gallery_array[0, i] = index_gallery
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (
path, set, i, index_gallery)
if not os.path.exists(filepath_gallery):
continue
break
if i == 99:
gallery_bool = False
if gallery_bool == False:
index_gallery = index_gallery_array[0, i]
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (
path, set, i, index_gallery)
if probe_bool == True:
while True:
index_probe = int(random.random() * 10)
filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (
path, set, id_probe, index_probe)
if not os.path.exists(filepath_probe):
continue
if index_gallery_array[
0, id_probe] == index_probe:
continue
probe_bool = False
break
if i == 99:
probe_bool = True
'''
while True:
index_probe = int(random.random() * 10)
filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, id_probe, index_probe)
if not os.path.exists(filepath_gallery):
continue
if index_gallery_array[1,id_probe] == index_probe:
continue
break
'''
#filepath_gallery = 'data/labeled/val/0000_01.jpg'
#filepath_probe = 'data/labeled/val/0000_02.jpg'
image1 = cv2.imread(filepath_gallery)
image1 = cv2.resize(image1,
(IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
image1 = np.reshape(
image1,
(1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
image2 = cv2.imread(filepath_probe)
image2 = cv2.resize(image2,
(IMAGE_WIDTH, IMAGE_HEIGHT))
image2 = cv2.cvtColor(image2, cv2.COLOR_BGR2RGB)
image2 = np.reshape(
image2,
(1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
test_images = np.array([image1, image2, image2])
#print (filepath_gallery)
#print (filepath_probe)
#print ('1111111111111111111111')
if i == j:
batch_labels = [1., 0.]
if i != j:
batch_labels = [0., 1.]
batch_labels = np.array(batch_labels)
print('test img :', test_images.shape)
feed_dict = {images: test_images, is_train: False}
prediction = sess.run(DD, feed_dict=feed_dict)
#print (prediction, prediction[0][1])
print(filepath_gallery, filepath_probe)
#print(bool(not np.argmax(prediction[0])))
print(prediction)
cmc_array[j, i] = prediction
#print(i,j)
#prediction = sess.run(inference, feed_dict=feed_dict)
#prediction = np.argmax(prediction, axis=1)
#label = np.argmax(batch_labels, axis=1)
cmc_score = cmc.cmc(cmc_array)
cmc_sum = cmc_score + cmc_sum
print(cmc_score)
cmc_sum = cmc_sum / do_times
print(cmc_sum)
print('final cmc')
elif FLAGS.mode == 'test':
image1 = cv2.imread(FLAGS.image1)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
image1 = np.reshape(
image1, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
image2 = cv2.imread(FLAGS.image2)
image2 = cv2.resize(image2, (IMAGE_WIDTH, IMAGE_HEIGHT))
image2 = cv2.cvtColor(image2, cv2.COLOR_BGR2RGB)
image2 = np.reshape(
image2, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
test_images = np.array([image1, image2, image2])
feed_dict = {
images: test_images,
is_train: False,
droup_is_training: False
}
#prediction, prediction2 = sess.run([DD,DD2], feed_dict=feed_dict)
prediction = sess.run([inference], feed_dict=feed_dict)
prediction = np.array(prediction)
print prediction.shape
print(np.argmax(prediction[0]) + 1)
def main(argv=None):
if FLAGS.mode == 'test':
FLAGS.batch_size = 1
if FLAGS.mode == 'cmc':
FLAGS.batch_size = 1
if FLAGS.mode == 'top1':
FLAGS.batch_size = 100
learning_rate = tf.placeholder(tf.float32, name='learning_rate')
images = tf.placeholder(
tf.float32, [3, FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3],
name='images')
images_total = tf.placeholder(
tf.float32, [FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3],
name='images_total')
images_one = tf.placeholder(tf.float32, [1, IMAGE_HEIGHT, IMAGE_WIDTH, 3],
name='images_one')
is_train = tf.placeholder(tf.bool, name='is_train')
global_step = tf.Variable(0, name='global_step', trainable=False)
weight_decay = 0.0005
tarin_num_id = 0
val_num_id = 0
if FLAGS.mode == 'train':
tarin_num_id = cuhk03_dataset_label2.get_num_id(
FLAGS.data_dir, 'train')
print(
tarin_num_id,
' 11111111111111111111 1111111111111111'
)
elif FLAGS.mode == 'val':
val_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'val')
images1, images2, images3 = preprocess(images, is_train)
img_combine = tf.concat([images1, images2, images3], 0)
train_mode = tf.placeholder(tf.bool)
# Create the model and an embedding head.
model = import_module('nets.' + 'resnet_v1_50')
head = import_module('heads.' + 'fc1024')
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
endpoints, body_prefix = model.endpoints(images_total, is_training=False)
feat = endpoints['resnet_v1_50/block4'] # (bt,7,7,2048)
print('Build network')
# global
anchor_feature = global_pooling(feat, weight_decay)
lr = FLAGS.learning_rate
#config=tf.ConfigProto(log_device_placement=True)
#config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
# GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.55
with tf.Session(config=config) as sess:
checkpoint_saver = tf.train.Saver(max_to_keep=0)
ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
if ckpt and ckpt.model_checkpoint_path:
print('Restore model')
print ckpt.model_checkpoint_path
#saver.restore(sess, ckpt.model_checkpoint_path)
checkpoint_saver.restore(sess, ckpt.model_checkpoint_path)
#for first , training load imagenet
else:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(model_variables)
print FLAGS.initial_checkpoint
saver.restore(sess, FLAGS.initial_checkpoint)
if FLAGS.mode == 'train':
step = sess.run(global_step)
for i in xrange(step, FLAGS.max_steps + 1):
batch_images, batch_labels, batch_images_total = cuhk03_dataset_label2.read_data(
FLAGS.data_dir, 'train', tarin_num_id, IMAGE_WIDTH,
IMAGE_HEIGHT, FLAGS.batch_size, FLAGS.ID_num,
FLAGS.IMG_PER_ID)
feed_dict = {
learning_rate: lr,
is_train: True,
train_mode: True,
images_total: batch_images_total
}
_, train_loss = sess.run([train, loss], feed_dict=feed_dict)
print('Step: %d, Learning rate: %f, Train loss: %f ' %
(i, lr, train_loss))
h, p, l = sess.run([NN, PP, loss], feed_dict=feed_dict)
print 'n:', h
print 'p:', p
print 'hard loss', l
lr = FLAGS.learning_rate * ((0.0001 * i + 1)**-0.75)
if i % 100 == 0:
saver.save(sess, FLAGS.logs_dir + 'model.ckpt', i)
elif FLAGS.mode == 'top1':
path = 'data_eye'
set = 'val'
cmc_sum = np.zeros((100, 100), dtype='f')
cmc_total = []
do_times = 20
for times in xrange(do_times):
query_feature = []
test_feature = []
for i in range(100):
while True:
index_gallery = int(random.random() * 10)
index_temp = index_gallery
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (
path, set, i, index_gallery)
if not os.path.exists(filepath_gallery):
continue
break
image1 = cv2.imread(filepath_gallery)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
query_feature.append(image1)
while True:
index_gallery = int(random.random() * 10)
if index_temp == index_gallery:
continue
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (
path, set, i, index_gallery)
if not os.path.exists(filepath_gallery):
continue
break
image1 = cv2.imread(filepath_gallery)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
test_feature.append(image1)
#print filepath_gallery,'\n'
query_feature = np.array(query_feature)
test_feature = np.array(test_feature)
feed_dict = {images_total: query_feature, is_train: False}
q_feat = sess.run(anchor_feature, feed_dict=feed_dict)
feed_dict = {images_total: test_feature, is_train: False}
test_feat = sess.run(anchor_feature, feed_dict=feed_dict)
cmc_array = []
tf_q_feat = tf.constant(q_feat)
tf_test_feat = tf.constant(test_feat)
h = tf.placeholder(tf.int32)
pick = tf_q_feat[h]
tf_q_feat = tf.reshape(pick, [1, 2048])
feat1 = tf.tile(tf_q_feat, [100, 1])
f = tf.square(tf.subtract(feat1, tf_test_feat))
d = tf.sqrt(tf.reduce_sum(f, 1)) # What about the axis ???
print d, 'f\n'
for t in range(100):
feed_dict = {h: t}
D = sess.run(d, feed_dict=feed_dict)
cmc_array.append(D)
cmc_array = np.array(cmc_array)
cmc_score = cmc.cmc(cmc_array)
cmc_sum = cmc_score + cmc_sum
cmc_total.append(cmc_score)
#top1=single_query(q_feat,test_feat,labels,labels,test_num=10)
print cmc_score
cmc_sum = cmc_sum / do_times
print(cmc_sum)
print('final cmc')
print('\n')
print cmc_total
def main(argv=None):
if FLAGS.mode == 'test':
FLAGS.batch_size = 1
if FLAGS.mode == 'cmc':
FLAGS.batch_size = 1
learning_rate = tf.placeholder(tf.float32, name='learning_rate')
images = tf.placeholder(tf.float32, [3, FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images')
images_total = tf.placeholder(tf.float32, [FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images_total')
labels = tf.placeholder(tf.float32, [FLAGS.batch_size], name='labels')
label_var = tf.placeholder(tf.int64, (None,))
is_train = tf.placeholder(tf.bool, name='is_train')
#global_step = tf.Variable(0, name='global_step', trainable=False)
weight_decay = 0.0005
tarin_num_id = 0
val_num_id = 0
if FLAGS.mode == 'train':
tarin_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'train')
print(tarin_num_id, ' 11111111111111111111 1111111111111111')
elif FLAGS.mode == 'val':
val_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'val')
#images1, images2,images3 = preprocess(images, is_train)
img_temp = []
for t in range(FLAGS.batch_size):
images_total_ex = tf.image.random_flip_left_right(images_total[t])
images_total_ex = tf.image.random_brightness(images_total_ex, max_delta=32. / 255.)
img_temp.append(images_total_ex)
images_total = tf.cast(img_temp,tf.float32)
network_factory = net.create_network_factory(
is_training=True, num_classes=1500 + 1,
add_logits= "cosine-softmax")
feature_var, logit_var = network_factory(images_total)
_create_loss(feature_var, logit_var, label_var, mode="cosine-softmax")
'''
trainable_scopes=None
if trainable_scopes is None:
variables_to_train = tf.trainable_variables()
else:
variables_to_train = []
for scope in trainable_scopes:
variables = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope)
variables_to_train.extend(variables)
'''
global_step = tf.train.get_or_create_global_step()
loss_var = tf.losses.get_total_loss()
'''
train_op = slim.learning.create_train_op(
loss_var, tf.train.AdamOptimizer(learning_rate=learning_rate),
global_step, summarize_gradients=False,
variables_to_train=variables_to_train)
'''
tf.summary.scalar("total_loss", loss_var)
tf.summary.scalar("learning_rate", learning_rate)
regularization_var = tf.reduce_sum(tf.losses.get_regularization_loss())
tf.summary.scalar("weight_loss", regularization_var)
#optimizer = tf.train.AdadeltaOptimizer(learning_rate)
#train = optimizer.minimize(loss, global_step=global_step)
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
train = optimizer.minimize(loss_var, global_step=global_step)
lr = FLAGS.learning_rate
#config=tf.ConfigProto(log_device_placement=True)
#config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
# GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
if ckpt and ckpt.model_checkpoint_path:
print('Restore model')
print ckpt.model_checkpoint_path
saver.restore(sess, ckpt.model_checkpoint_path)
if FLAGS.mode == 'train':
step = sess.run(global_step)
for i in xrange(step, FLAGS.max_steps + 1):
batch_images, batch_labels, batch_images_total = cuhk03_dataset_label2.read_data(FLAGS.data_dir, 'train', tarin_num_id,
IMAGE_WIDTH, IMAGE_HEIGHT, FLAGS.batch_size,FLAGS.ID_num,FLAGS.IMG_PER_ID)
feed_dict = {learning_rate: lr, is_train: True , images_total: batch_images_total, labels: batch_labels, label_var : batch_labels}
start_time = time.time()
#_,train_loss = sess.run([train,loss], feed_dict=feed_dict)
_,train_loss = sess.run([train,loss_var], feed_dict=feed_dict)
duration = time.time() - start_time
train_loss = np.array(train_loss)
print duration,'duration'
print('Step: %d, Learning rate: %f, Train loss: %f ' % (i, lr, train_loss))
#lr = FLAGS.learning_rate * ((0.1) ** (i/160000))
lr = FLAGS.learning_rate * ((0.0001 * i + 1) ** -0.75)
if i % 1000 == 0:
saver.save(sess, FLAGS.logs_dir + 'model.ckpt', i)
elif FLAGS.mode == 'val':
total = 0.
for _ in xrange(10):
batch_images, batch_labels = cuhk03_dataset_label2.read_data(FLAGS.data_dir, 'val', val_num_id,
IMAGE_WIDTH, IMAGE_HEIGHT, FLAGS.batch_size)
feed_dict = {images: batch_images, labels: batch_labels, is_train: False}
prediction = sess.run(inference, feed_dict=feed_dict)
prediction = np.argmax(prediction, axis=1)
label = np.argmax(batch_labels, axis=1)
for i in xrange(len(prediction)):
if prediction[i] == label[i]:
total += 1
print('Accuracy: %f' % (total / (FLAGS.batch_size * 10)))
'''
for i in xrange(len(prediction)):
print('Prediction: %s, Label: %s' % (prediction[i] == 0, labels[i] == 0))
image1 = cv2.cvtColor(batch_images[0][i], cv2.COLOR_RGB2BGR)
image2 = cv2.cvtColor(batch_images[1][i], cv2.COLOR_RGB2BGR)
image = np.concatenate((image1, image2), axis=1)
cv2.imshow('image', image)
key = cv2.waitKey(0)
if key == 1048603: # ESC key
break
'''
elif FLAGS.mode == 'test':
image1 = cv2.imread(FLAGS.image1)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
image1 = np.reshape(image1, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
image2 = cv2.imread(FLAGS.image2)
image2 = cv2.resize(image2, (IMAGE_WIDTH, IMAGE_HEIGHT))
image2 = cv2.cvtColor(image2, cv2.COLOR_BGR2RGB)
image2 = np.reshape(image2, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
test_images = np.array([image1, image2,image2])
feed_dict = {images: test_images, is_train: False, droup_is_training: False}
#prediction, prediction2 = sess.run([DD,DD2], feed_dict=feed_dict)
prediction = sess.run([inference], feed_dict=feed_dict)
prediction = np.array(prediction)
print prediction.shape
print( np.argmax(prediction[0])+1)
def main(argv=None):
if FLAGS.mode == 'test':
FLAGS.batch_size = 1
if FLAGS.mode == 'cmc':
FLAGS.batch_size = 1
learning_rate = tf.placeholder(tf.float32, name='learning_rate')
#images = tf.placeholder(tf.float32, [2, FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images')
images = tf.placeholder(tf.float32, [3, FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images')
images_total = tf.placeholder(tf.float32, [FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images_total')
labels = tf.placeholder(tf.float32, [FLAGS.batch_size], name='labels')
#labels_neg = tf.placeholder(tf.float32, [FLAGS.batch_size, 743], name='labels')
#total
#labels = tf.placeholder(tf.float32, [FLAGS.batch_size, 847], name='labels')
#labels_neg = tf.placeholder(tf.float32, [FLAGS.batch_size, 847], name='labels')
#eye
#labels = tf.placeholder(tf.float32, [FLAGS.batch_size, 104], name='labels')
#labels_neg = tf.placeholder(tf.float32, [FLAGS.batch_size, 104], name='labels')
is_train = tf.placeholder(tf.bool, name='is_train')
global_step = tf.Variable(0, name='global_step', trainable=False)
weight_decay = 0.0005
tarin_num_id = 0
val_num_id = 0
if FLAGS.mode == 'train':
tarin_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'train')
print(tarin_num_id, ' 11111111111111111111 1111111111111111')
elif FLAGS.mode == 'val':
val_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'val')
# Create the model and an embedding head.
model = import_module('nets.' + 'resnet_v1_50')
head = import_module('heads.' + 'fc1024')
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
endpoints, body_prefix = model.endpoints(images_total, is_training=True)
with tf.name_scope('head'):
endpoints = head.head(endpoints, FLAGS.embedding_dim, is_training=True)
'''
print endpoints['model_output'] # (bt,2048)
print endpoints['global_pool'] # (bt,2048)
print endpoints['resnet_v1_50/block4']# (bt,7,7,2048)
'''
# Create the model and an embedding head.
model2 = import_module('nets.' + 'resnet_v1_101')
endpoints2, body_prefix2 = model2.endpoints(images_total, is_training=True)
train_mode = tf.placeholder(tf.bool)
print('Build network')
feat = endpoints['resnet_v1_50/block4']# (bt,7,7,2048)
feat2 = endpoints2['resnet_v1_101/block4']# (bt,7,7,2048)
#feat = tf.convert_to_tensor(feat, dtype=tf.float32)
# global
feature,feature2 = global_pooling(feat,feat2,weight_decay)
loss_triplet ,PP,NN = batch_hard_triplet_loss(labels,feature,0.3)
_,dis_matrix1 = triplet_hard_loss(feature,FLAGS.ID_num,FLAGS.IMG_PER_ID)
_,dis_matrix2 = triplet_hard_loss(feature2,FLAGS.ID_num,FLAGS.IMG_PER_ID)
mul_loss = multual_loss(dis_matrix1,dis_matrix2)
local_anchor_feature, local_anchor_feature2 = local_pooling(feat,feat2,weight_decay)
local_loss_triplet ,local_pos_loss, local_neg_loss = local_triplet_hard_loss(local_anchor_feature,FLAGS.ID_num,FLAGS.IMG_PER_ID)
loss_triplet2 ,PP2,NN2 = batch_hard_triplet_loss(labels,feature2,0.3)
local_loss_triplet2 ,local_pos_loss2, local_neg_loss2 = local_triplet_hard_loss(local_anchor_feature2,FLAGS.ID_num,FLAGS.IMG_PER_ID)
s1 = fully_connected_class(feature,feature_dim=2048,num_classes=1000)#tarin_num_id
cross_entropy_var = slim.losses.sparse_softmax_cross_entropy(s1, tf.cast(labels, tf.int64))
loss_softmax = cross_entropy_var
#loss_softmax = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels_softmax, logits=s1))
inference = tf.nn.softmax(s1)
s2 = fully_connected_class2(feature2,feature_dim=2048,num_classes=1000)
cross_entropy_var2 = slim.losses.sparse_softmax_cross_entropy(s2, tf.cast(labels, tf.int64))
loss_softmax2 = cross_entropy_var2
#loss_softmax2 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels_softmax, logits=s2))
inference2 = tf.nn.softmax(s2)
multual_softmax1 = kl_loss_compute(s1, s2)
multual_softmax2 = kl_loss_compute(s2, s1)
P1= tf.reduce_mean(PP)
P2= tf.reduce_mean(PP2)
N1= tf.reduce_mean(NN)
N2= tf.reduce_mean(NN2)
LP1= tf.reduce_mean(local_pos_loss)
LN1= tf.reduce_mean(local_neg_loss)
'''
# global
feature2 = global_pooling(feat2,weight_decay)
#loss_triplet,PP,NN = triplet_hard_loss(feature,FLAGS.ID_num,FLAGS.IMG_PER_ID)
loss_triplet2 ,PP2,NN2 = batch_hard_triplet_loss(labels,feature2,0.3)
#local
local_anchor_feature2 = local_pooling(feat2,weight_decay)
local_loss_triplet2 ,local_pos_loss2, local_neg_loss2 = local_triplet_hard_loss(local_anchor_feature2,FLAGS.ID_num,FLAGS.IMG_PER_ID)
'''
loss = local_loss_triplet*FLAGS.local_rate + loss_triplet*FLAGS.global_rate + mul_loss + loss_softmax + multual_softmax1
#DD = compute_euclidean_distance(anchor_feature,positive_feature)
loss2 = local_loss_triplet2*FLAGS.local_rate + loss_triplet2*FLAGS.global_rate + mul_loss + loss_softmax2 + multual_softmax2
if FLAGS.mode == 'val' or FLAGS.mode == 'cmc' or FLAGS.mode == 'test':
loss ,pos_loss, neg_loss = triplet_loss(anchor_feature, positive_feature, negative_feature, 0.3)
print ' ERROR ERROR '
None
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
model_variables = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES, body_prefix)
model_variables2 = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES, body_prefix2)
# Update_ops are used to update batchnorm stats.
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
train = optimizer.minimize(loss, global_step=global_step)
optimizer2 = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
train2 = optimizer2.minimize(loss2, global_step=global_step)
tf.summary.scalar("total_loss 1", loss)
tf.summary.scalar("total_loss 2", loss2)
tf.summary.scalar("learning_rate", learning_rate)
regularization_var = tf.reduce_sum(tf.losses.get_regularization_loss())
tf.summary.scalar("weight_loss", regularization_var)
lr = FLAGS.learning_rate
#config=tf.ConfigProto(log_device_placement=True)
#config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
# GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.95
with tf.Session(config=config) as sess:
merged = tf.summary.merge_all()
writer = tf.summary.FileWriter("TensorBoard_1x1_a_1x7/", graph = sess.graph)
#sess.run(tf.global_variables_initializer())
#saver = tf.train.Saver()
#checkpoint_saver = tf.train.Saver(max_to_keep=0)
checkpoint_saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
if ckpt and ckpt.model_checkpoint_path:
print('Restore model')
print ckpt.model_checkpoint_path
#saver.restore(sess, ckpt.model_checkpoint_path)
checkpoint_saver.restore(sess, ckpt.model_checkpoint_path)
#for first , training load imagenet
else:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(model_variables)
print FLAGS.initial_checkpoint
saver.restore(sess, FLAGS.initial_checkpoint)
saver2 = tf.train.Saver(model_variables2)
print FLAGS.initial_checkpoint2
saver2.restore(sess, FLAGS.initial_checkpoint2)
if FLAGS.mode == 'train':
step = sess.run(global_step)
for i in xrange(step, FLAGS.max_steps + 1):
batch_images, batch_labels, batch_images_total = cuhk03_dataset_label2.read_data(FLAGS.data_dir, 'train', tarin_num_id,
IMAGE_WIDTH, IMAGE_HEIGHT, FLAGS.batch_size,FLAGS.ID_num,FLAGS.IMG_PER_ID)
#feed_dict = {learning_rate: lr, is_train: True , labels: batch_labels, droup_is_training: False, train_mode: True, images_total: batch_images_total} #no label images: batch_images,
feed_dict = {learning_rate: lr, is_train: True , train_mode: True, images_total: batch_images_total, labels: batch_labels}
start = time.time()
_,_,train_loss,train_loss2 = sess.run([train,train2,loss,loss2 ], feed_dict=feed_dict)
print('Step: %d, Learning rate: %f, Train loss: %f , Train loss2: %f' % (i, lr, train_loss,train_loss2))
gtoloss,gp,gn = sess.run([loss_triplet,P1,N1], feed_dict=feed_dict)
print 'global hard: ',gtoloss
print 'global P1: ',gp
print 'global N1: ',gn
toloss,p,n = sess.run([local_loss_triplet,LP1,LN1], feed_dict=feed_dict)
print 'local hard: ',toloss
print 'local P: ',p
print 'local N: ',n
mul,p2,n2 = sess.run([mul_loss,loss_triplet2,local_loss_triplet2], feed_dict=feed_dict)
print 'mul loss: ',mul
print 'loss_triplet2: ',p2
print 'local_loss_triplet2: ',n2
end = time.time()
elapsed = end - start
print "Time taken: ", elapsed, "seconds."
#lr = FLAGS.learning_rate / ((2) ** (i/160000)) * 0.1
lr = FLAGS.learning_rate * ((0.0001 * i + 1) ** -0.75)
if i % 100 == 0:
checkpoint_saver.save(sess,FLAGS.logs_dir + 'model.ckpt', i)
if i % 20 == 0:
result = sess.run(merged, feed_dict=feed_dict)
writer.add_summary(result, i)
elif FLAGS.mode == 'val':
total = 0.
for _ in xrange(10):
batch_images, batch_labels = cuhk03_dataset_label2.read_data(FLAGS.data_dir, 'val', val_num_id,
IMAGE_WIDTH, IMAGE_HEIGHT, FLAGS.batch_size)
feed_dict = {images: batch_images, labels: batch_labels, is_train: False}
prediction = sess.run(inference, feed_dict=feed_dict)
prediction = np.argmax(prediction, axis=1)
label = np.argmax(batch_labels, axis=1)
for i in xrange(len(prediction)):
if prediction[i] == label[i]:
total += 1
print('Accuracy: %f' % (total / (FLAGS.batch_size * 10)))
'''
for i in xrange(len(prediction)):
print('Prediction: %s, Label: %s' % (prediction[i] == 0, labels[i] == 0))
image1 = cv2.cvtColor(batch_images[0][i], cv2.COLOR_RGB2BGR)
image2 = cv2.cvtColor(batch_images[1][i], cv2.COLOR_RGB2BGR)
image = np.concatenate((image1, image2), axis=1)
cv2.imshow('image', image)
key = cv2.waitKey(0)
if key == 1048603: # ESC key
break
'''
elif FLAGS.mode == 'cmc':
do_times = 1
cmc_sum=np.zeros((100, 100), dtype='f')
for times in xrange(do_times):
path = 'data'
set = 'val'
cmc_array=np.ones((100, 100), dtype='f')
batch_images = []
batch_labels = []
index_gallery_array=np.ones((1, 100), dtype='f')
gallery_bool = True
probe_bool = True
for j in xrange(100):
id_probe = j
for i in xrange(100):
batch_images = []
batch_labels = []
filepath = ''
#filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, i, index_gallery)
#filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, id_probe, index_probe)
if gallery_bool == True:
while True:
index_gallery = int(random.random() * 10)
index_gallery_array[0,i] = index_gallery
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, i, index_gallery)
if not os.path.exists(filepath_gallery):
continue
break
if i ==99:
gallery_bool = False
if gallery_bool == False:
index_gallery = index_gallery_array[0,i]
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, i, index_gallery)
if probe_bool == True:
while True:
index_probe = int(random.random() * 10)
filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, id_probe, index_probe)
if not os.path.exists(filepath_probe):
continue
if index_gallery_array[0,id_probe] == index_probe:
continue
probe_bool = False
break
if i ==99:
probe_bool = True
'''
while True:
index_probe = int(random.random() * 10)
filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, id_probe, index_probe)
if not os.path.exists(filepath_gallery):
continue
if index_gallery_array[1,id_probe] == index_probe:
continue
break
'''
#filepath_gallery = 'data/labeled/val/0000_01.jpg'
#filepath_probe = 'data/labeled/val/0000_02.jpg'
image1 = cv2.imread(filepath_gallery)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
image1 = np.reshape(image1, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
image2 = cv2.imread(filepath_probe)
image2 = cv2.resize(image2, (IMAGE_WIDTH, IMAGE_HEIGHT))
image2 = cv2.cvtColor(image2, cv2.COLOR_BGR2RGB)
image2 = np.reshape(image2, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
test_images = np.array([image1, image2, image2])
#print (filepath_gallery)
#print (filepath_probe)
#print ('1111111111111111111111')
if i == j:
batch_labels = [1., 0.]
if i != j:
batch_labels = [0., 1.]
batch_labels = np.array(batch_labels)
print('test img :',test_images.shape)
feed_dict = {images: test_images, is_train: False}
prediction = sess.run(DD, feed_dict=feed_dict)
#print (prediction, prediction[0][1])
print (filepath_gallery,filepath_probe)
#print(bool(not np.argmax(prediction[0])))
print (prediction)
cmc_array[j,i] = prediction
#print(i,j)
#prediction = sess.run(inference, feed_dict=feed_dict)
#prediction = np.argmax(prediction, axis=1)
#label = np.argmax(batch_labels, axis=1)
cmc_score = cmc.cmc(cmc_array)
cmc_sum = cmc_score + cmc_sum
print(cmc_score)
cmc_sum = cmc_sum/do_times
print(cmc_sum)
print('final cmc')
elif FLAGS.mode == 'test':
image1 = cv2.imread(FLAGS.image1)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
image1 = np.reshape(image1, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
image2 = cv2.imread(FLAGS.image2)
image2 = cv2.resize(image2, (IMAGE_WIDTH, IMAGE_HEIGHT))
image2 = cv2.cvtColor(image2, cv2.COLOR_BGR2RGB)
image2 = np.reshape(image2, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
test_images = np.array([image1, image2,image2])
feed_dict = {images: test_images, is_train: False, droup_is_training: False}
#prediction, prediction2 = sess.run([DD,DD2], feed_dict=feed_dict)
prediction = sess.run([inference], feed_dict=feed_dict)
prediction = np.array(prediction)
print prediction.shape
print( np.argmax(prediction[0])+1)
def main(argv=None):
if FLAGS.mode == 'test':
FLAGS.batch_size = 1
if FLAGS.mode == 'cmc':
FLAGS.batch_size = 1
learning_rate = tf.placeholder(tf.float32, name='learning_rate')
images = tf.placeholder(tf.float32, [3, FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images')
images_total = tf.placeholder(tf.float32, [FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images_total')
labels = tf.placeholder(tf.float32, [FLAGS.batch_size], name='labels')
is_train = tf.placeholder(tf.bool, name='is_train')
global_step = tf.Variable(0, name='global_step', trainable=False)
weight_decay = 0.0005
tarin_num_id = 0
val_num_id = 0
if FLAGS.mode == 'train':
tarin_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'train')
print(tarin_num_id, ' 11111111111111111111 1111111111111111')
elif FLAGS.mode == 'val':
val_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'val')
# Create the model and an embedding head.
model = import_module('nets.' + 'resnet_v1_50')
head = import_module('heads.' + 'fc1024')
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
endpoints, body_prefix = model.endpoints(images_total, is_training=True)
with tf.name_scope('head'):
endpoints = head.head(endpoints, FLAGS.embedding_dim, is_training=True)
'''
print endpoints['model_output'] # (bt,2048)
print endpoints['global_pool'] # (bt,2048)
print endpoints['resnet_v1_50/block4']# (bt,7,7,2048)
print ' 1\n'
'''
train_mode = tf.placeholder(tf.bool)
print('Build network')
feat = endpoints['resnet_v1_50/block4']# (bt,7,7,2048)
#feat = tf.convert_to_tensor(feat, dtype=tf.float32)
feat_1x1 = tf.layers.conv2d(feat, 2048, [1, 1],padding='valid',
kernel_regularizer=tf.contrib.layers.l2_regularizer(weight_decay), reuse=None, name='conv1x1')
feature = part_attend(feat_1x1,weight_decay)
#loss_triplet,PP,NN = triplet_hard_loss(feature,FLAGS.ID_num,FLAGS.IMG_PER_ID)
loss_triplet ,PP,NN = batch_hard_triplet_loss(labels,feature,0.3)
loss = loss_triplet*FLAGS.global_rate
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
model_variables = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES, body_prefix)
#optimizer = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(loss)
#optimizer = tf.train.AdadeltaOptimizer(learning_rate)
#train = optimizer.minimize(loss, global_step=global_step)
# Update_ops are used to update batchnorm stats.
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
#train_op = optimizer.minimize(loss_mean, global_step=global_step)
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
train = optimizer.minimize(loss, global_step=global_step)
lr = FLAGS.learning_rate
#config=tf.ConfigProto(log_device_placement=True)
#config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
# GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.9
with tf.Session(config=config) as sess:
print '\n'
#print model_variables
print '\n'
#sess.run(tf.global_variables_initializer())
#saver = tf.train.Saver()
#checkpoint_saver = tf.train.Saver(max_to_keep=0)
checkpoint_saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
if ckpt and ckpt.model_checkpoint_path:
print('Restore model')
print ckpt.model_checkpoint_path
#saver.restore(sess, ckpt.model_checkpoint_path)
checkpoint_saver.restore(sess, ckpt.model_checkpoint_path)
#for first , training load imagenet
else:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(model_variables)
print FLAGS.initial_checkpoint
saver.restore(sess, FLAGS.initial_checkpoint)
if FLAGS.mode == 'train':
step = sess.run(global_step)
for i in xrange(step, FLAGS.max_steps + 1):
batch_images, batch_labels, batch_images_total = cuhk03_dataset_label2.read_data(FLAGS.data_dir, 'train', tarin_num_id,
IMAGE_WIDTH, IMAGE_HEIGHT, FLAGS.batch_size,FLAGS.ID_num,FLAGS.IMG_PER_ID)
feed_dict = {learning_rate: lr, is_train: True , train_mode: True, images_total: batch_images_total, labels: batch_labels}
_,train_loss = sess.run([train,loss], feed_dict=feed_dict)
print('Step: %d, Learning rate: %f, Train loss: %f ' % (i, lr, train_loss))
gtoloss,gp,gn = sess.run([loss_triplet,PP,NN], feed_dict=feed_dict)
print 'global hard: ',gtoloss
print 'global P: ',gp
print 'global N: ',gn
#lr = FLAGS.learning_rate / ((2) ** (i/160000)) * 0.1
lr = FLAGS.learning_rate * ((0.0001 * i + 1) ** -0.75)
if i % 100 == 0:
#saver.save(sess, FLAGS.logs_dir + 'model.ckpt', i)
# test save
#vgg.save_npy(sess, './big.npy')
checkpoint_saver.save(sess,FLAGS.logs_dir + 'model.ckpt', i)
elif FLAGS.mode == 'val':
total = 0.
for _ in xrange(10):
batch_images, batch_labels = cuhk03_dataset_label2.read_data(FLAGS.data_dir, 'val', val_num_id,
IMAGE_WIDTH, IMAGE_HEIGHT, FLAGS.batch_size)
feed_dict = {images: batch_images, labels: batch_labels, is_train: False}
prediction = sess.run(inference, feed_dict=feed_dict)
prediction = np.argmax(prediction, axis=1)
label = np.argmax(batch_labels, axis=1)
for i in xrange(len(prediction)):
if prediction[i] == label[i]:
total += 1
print('Accuracy: %f' % (total / (FLAGS.batch_size * 10)))
'''
for i in xrange(len(prediction)):
print('Prediction: %s, Label: %s' % (prediction[i] == 0, labels[i] == 0))
image1 = cv2.cvtColor(batch_images[0][i], cv2.COLOR_RGB2BGR)
image2 = cv2.cvtColor(batch_images[1][i], cv2.COLOR_RGB2BGR)
image = np.concatenate((image1, image2), axis=1)
cv2.imshow('image', image)
key = cv2.waitKey(0)
if key == 1048603: # ESC key
break
'''
elif FLAGS.mode == 'test':
image1 = cv2.imread(FLAGS.image1)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
image1 = np.reshape(image1, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
image2 = cv2.imread(FLAGS.image2)
image2 = cv2.resize(image2, (IMAGE_WIDTH, IMAGE_HEIGHT))
image2 = cv2.cvtColor(image2, cv2.COLOR_BGR2RGB)
image2 = np.reshape(image2, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
test_images = np.array([image1, image2,image2])
feed_dict = {images: test_images, is_train: False, droup_is_training: False}
#prediction, prediction2 = sess.run([DD,DD2], feed_dict=feed_dict)
prediction = sess.run([inference], feed_dict=feed_dict)
prediction = np.array(prediction)
print prediction.shape
print( np.argmax(prediction[0])+1)
def main(argv=None):
if FLAGS.mode == 'test':
FLAGS.batch_size = 1
if FLAGS.mode == 'cmc':
FLAGS.batch_size = 1
learning_rate = tf.placeholder(tf.float32, name='learning_rate')
#images = tf.placeholder(tf.float32, [2, FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images')
images = tf.placeholder(
tf.float32, [3, FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3],
name='images')
images_total = tf.placeholder(
tf.float32, [FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3],
name='images_total')
labels = tf.placeholder(tf.float32, [FLAGS.batch_size], name='labels')
#labels_neg = tf.placeholder(tf.float32, [FLAGS.batch_size, 743], name='labels')
#total
#labels = tf.placeholder(tf.float32, [FLAGS.batch_size, 847], name='labels')
#labels_neg = tf.placeholder(tf.float32, [FLAGS.batch_size, 847], name='labels')
#eye
#labels = tf.placeholder(tf.float32, [FLAGS.batch_size, 104], name='labels')
#labels_neg = tf.placeholder(tf.float32, [FLAGS.batch_size, 104], name='labels')
is_train = tf.placeholder(tf.bool, name='is_train')
global_step = tf.Variable(0, name='global_step', trainable=False)
weight_decay = 0.0005
tarin_num_id = 0
val_num_id = 0
if FLAGS.mode == 'train':
tarin_num_id = cuhk03_dataset_label2.get_num_id(
FLAGS.data_dir, 'train')
print(
tarin_num_id,
' 11111111111111111111 1111111111111111'
)
elif FLAGS.mode == 'val':
val_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'val')
# Create the model and an embedding head.
model = import_module('nets.' + 'resnet_v1_50')
head = import_module('heads.' + 'fc1024')
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
endpoints, body_prefix = model.endpoints(images_total, is_training=True)
with tf.name_scope('head'):
endpoints = head.head(endpoints, FLAGS.embedding_dim, is_training=True)
'''
print endpoints['model_output'] # (bt,2048)
print endpoints['global_pool'] # (bt,2048)
print endpoints['resnet_v1_50/block4']# (bt,7,7,2048)
'''
# Create the model and an embedding head.
model2 = import_module('nets.' + 'resnet_v1_101')
endpoints2, body_prefix2 = model2.endpoints(images_total, is_training=True)
train_mode = tf.placeholder(tf.bool)
print('Build network')
feat = endpoints['resnet_v1_50/block4'] # (bt,7,7,2048)
feat2 = endpoints2['resnet_v1_101/block4'] # (bt,7,7,2048)
#feat = tf.convert_to_tensor(feat, dtype=tf.float32)
# global
feature, feature2 = global_pooling(feat, feat2, weight_decay)
loss_triplet, PP, NN = batch_hard_triplet_loss(labels, feature, 0.3)
_, dis_matrix1 = triplet_hard_loss(feature, FLAGS.ID_num, FLAGS.IMG_PER_ID)
_, dis_matrix2 = triplet_hard_loss(feature2, FLAGS.ID_num,
FLAGS.IMG_PER_ID)
mul_loss = multual_loss(dis_matrix1, dis_matrix2)
local_anchor_feature, local_anchor_feature2 = local_pooling(
feat, feat2, weight_decay)
local_loss_triplet, local_pos_loss, local_neg_loss = local_triplet_hard_loss(
local_anchor_feature, FLAGS.ID_num, FLAGS.IMG_PER_ID)
loss_triplet2, PP2, NN2 = batch_hard_triplet_loss(labels, feature2, 0.3)
local_loss_triplet2, local_pos_loss2, local_neg_loss2 = local_triplet_hard_loss(
local_anchor_feature2, FLAGS.ID_num, FLAGS.IMG_PER_ID)
s1 = fully_connected_class(feature, feature_dim=2048,
num_classes=1000) #tarin_num_id
cross_entropy_var = slim.losses.sparse_softmax_cross_entropy(
s1, tf.cast(labels, tf.int64))
loss_softmax = cross_entropy_var
#loss_softmax = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels_softmax, logits=s1))
inference = tf.nn.softmax(s1)
s2 = fully_connected_class2(feature2, feature_dim=2048, num_classes=1000)
cross_entropy_var2 = slim.losses.sparse_softmax_cross_entropy(
s2, tf.cast(labels, tf.int64))
loss_softmax2 = cross_entropy_var2
#loss_softmax2 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels_softmax, logits=s2))
inference2 = tf.nn.softmax(s2)
multual_softmax1 = kl_loss_compute(s1, s2)
multual_softmax2 = kl_loss_compute(s2, s1)
P1 = tf.reduce_mean(PP)
P2 = tf.reduce_mean(PP2)
N1 = tf.reduce_mean(NN)
N2 = tf.reduce_mean(NN2)
LP1 = tf.reduce_mean(local_pos_loss)
LN1 = tf.reduce_mean(local_neg_loss)
'''
# global
feature2 = global_pooling(feat2,weight_decay)
#loss_triplet,PP,NN = triplet_hard_loss(feature,FLAGS.ID_num,FLAGS.IMG_PER_ID)
loss_triplet2 ,PP2,NN2 = batch_hard_triplet_loss(labels,feature2,0.3)
#local
local_anchor_feature2 = local_pooling(feat2,weight_decay)
local_loss_triplet2 ,local_pos_loss2, local_neg_loss2 = local_triplet_hard_loss(local_anchor_feature2,FLAGS.ID_num,FLAGS.IMG_PER_ID)
'''
loss = local_loss_triplet * FLAGS.local_rate + loss_triplet * FLAGS.global_rate + mul_loss + loss_softmax + multual_softmax1
#DD = compute_euclidean_distance(anchor_feature,positive_feature)
loss2 = local_loss_triplet2 * FLAGS.local_rate + loss_triplet2 * FLAGS.global_rate + mul_loss + loss_softmax2 + multual_softmax2
if FLAGS.mode == 'val' or FLAGS.mode == 'cmc' or FLAGS.mode == 'test':
loss, pos_loss, neg_loss = triplet_loss(anchor_feature,
positive_feature,
negative_feature, 0.3)
print ' ERROR ERROR '
None
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
model_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
body_prefix)
model_variables2 = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
body_prefix2)
# Update_ops are used to update batchnorm stats.
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
train = optimizer.minimize(loss, global_step=global_step)
optimizer2 = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
train2 = optimizer2.minimize(loss2, global_step=global_step)
tf.summary.scalar("total_loss 1", loss)
tf.summary.scalar("total_loss 2", loss2)
tf.summary.scalar("learning_rate", learning_rate)
regularization_var = tf.reduce_sum(tf.losses.get_regularization_loss())
tf.summary.scalar("weight_loss", regularization_var)
lr = FLAGS.learning_rate
#config=tf.ConfigProto(log_device_placement=True)
#config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
# GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.95
with tf.Session(config=config) as sess:
merged = tf.summary.merge_all()
writer = tf.summary.FileWriter("TensorBoard_1x1_a_1x7/",
graph=sess.graph)
#sess.run(tf.global_variables_initializer())
#saver = tf.train.Saver()
#checkpoint_saver = tf.train.Saver(max_to_keep=0)
checkpoint_saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
if ckpt and ckpt.model_checkpoint_path:
print('Restore model')
print ckpt.model_checkpoint_path
#saver.restore(sess, ckpt.model_checkpoint_path)
checkpoint_saver.restore(sess, ckpt.model_checkpoint_path)
#for first , training load imagenet
else:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(model_variables)
print FLAGS.initial_checkpoint
saver.restore(sess, FLAGS.initial_checkpoint)
saver2 = tf.train.Saver(model_variables2)
print FLAGS.initial_checkpoint2
saver2.restore(sess, FLAGS.initial_checkpoint2)
if FLAGS.mode == 'train':
step = sess.run(global_step)
for i in xrange(step, FLAGS.max_steps + 1):
batch_images, batch_labels, batch_images_total = cuhk03_dataset_label2.read_data(
FLAGS.data_dir, 'train', tarin_num_id, IMAGE_WIDTH,
IMAGE_HEIGHT, FLAGS.batch_size, FLAGS.ID_num,
FLAGS.IMG_PER_ID)
#feed_dict = {learning_rate: lr, is_train: True , labels: batch_labels, droup_is_training: False, train_mode: True, images_total: batch_images_total} #no label images: batch_images,
feed_dict = {
learning_rate: lr,
is_train: True,
train_mode: True,
images_total: batch_images_total,
labels: batch_labels
}
start = time.time()
_, _, train_loss, train_loss2 = sess.run(
[train, train2, loss, loss2], feed_dict=feed_dict)
print(
'Step: %d, Learning rate: %f, Train loss: %f , Train loss2: %f'
% (i, lr, train_loss, train_loss2))
gtoloss, gp, gn = sess.run([loss_triplet, P1, N1],
feed_dict=feed_dict)
print 'global hard: ', gtoloss
print 'global P1: ', gp
print 'global N1: ', gn
toloss, p, n = sess.run([local_loss_triplet, LP1, LN1],
feed_dict=feed_dict)
print 'local hard: ', toloss
print 'local P: ', p
print 'local N: ', n
mul, p2, n2 = sess.run(
[mul_loss, loss_triplet2, local_loss_triplet2],
feed_dict=feed_dict)
print 'mul loss: ', mul
print 'loss_triplet2: ', p2
print 'local_loss_triplet2: ', n2
end = time.time()
elapsed = end - start
print "Time taken: ", elapsed, "seconds."
#lr = FLAGS.learning_rate / ((2) ** (i/160000)) * 0.1
lr = FLAGS.learning_rate * ((0.0001 * i + 1)**-0.75)
if i % 100 == 0:
checkpoint_saver.save(sess, FLAGS.logs_dir + 'model.ckpt',
i)
if i % 20 == 0:
result = sess.run(merged, feed_dict=feed_dict)
writer.add_summary(result, i)
def main(argv=None):
if FLAGS.mode == 'test':
FLAGS.batch_size = 1
if FLAGS.mode == 'cmc':
FLAGS.batch_size = 1
learning_rate = tf.placeholder(tf.float32, name='learning_rate')
images = tf.placeholder(
tf.float32, [3, FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3],
name='images')
images_total = tf.placeholder(
tf.float32, [FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3],
name='images_total')
is_train = tf.placeholder(tf.bool, name='is_train')
global_step = tf.Variable(0, name='global_step', trainable=False)
weight_decay = 0.0005
tarin_num_id = 0
val_num_id = 0
if FLAGS.mode == 'train':
tarin_num_id = cuhk03_dataset_label2.get_num_id(
FLAGS.data_dir, 'train')
print(
tarin_num_id,
' 11111111111111111111 1111111111111111'
)
elif FLAGS.mode == 'val':
val_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'val')
images1, images2, images3 = preprocess(images, is_train)
img_combine = tf.concat([images1, images2, images3], 0)
train_mode = tf.placeholder(tf.bool)
'''
vgg = vgg19.Vgg19('./vgg19.npy') # fix !!!!!!!!!!!!!!!!!!!!
print images_total,'images_total'
vgg.build(img_combine, train_mode)
feat = vgg.pool5
print feat,'1'
feat1 ,feat2 ,feat3 = tf.split(feat, [FLAGS.batch_size, FLAGS.batch_size,FLAGS.batch_size])
'''
# Create the model and an embedding head.
model = import_module('nets.' + 'resnet_v1_50')
head = import_module('heads.' + 'fc1024')
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
endpoints, body_prefix = model.endpoints(img_combine, is_training=False)
feat = endpoints['resnet_v1_50/block4'] # (bt,7,7,2048)
feat1, feat2, feat3 = tf.split(
feat, [FLAGS.batch_size, FLAGS.batch_size, FLAGS.batch_size])
print('Build network')
# global
anchor_feature, positive_feature, negative_feature = global_pooling(
feat1, feat2, feat3, weight_decay)
#feature = global_pooling(feat,weight_decay)
#loss_triplet ,pos_loss, neg_loss = triplet_loss(feature[:4], feature[4:8], feature[8:12], 0.3 )
#loss_triplet,NN,PP = triplet_hard_loss(feature,FLAGS.ID_num,FLAGS.IMG_PER_ID)
#loss = loss_triplet
DD = compute_euclidean_distance(anchor_feature, positive_feature)
#optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
#train = optimizer.minimize(loss, global_step=global_step)
#train_class = optimizer.minimize(loss_class, global_step=global_step)
lr = FLAGS.learning_rate
#config=tf.ConfigProto(log_device_placement=True)
#config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
# GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.35
with tf.Session(config=config) as sess:
checkpoint_saver = tf.train.Saver(max_to_keep=0)
ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
if ckpt and ckpt.model_checkpoint_path:
print('Restore model')
print ckpt.model_checkpoint_path
#saver.restore(sess, ckpt.model_checkpoint_path)
checkpoint_saver.restore(sess, ckpt.model_checkpoint_path)
#for first , training load imagenet
else:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(model_variables)
print FLAGS.initial_checkpoint
saver.restore(sess, FLAGS.initial_checkpoint)
if FLAGS.mode == 'train':
step = sess.run(global_step)
for i in xrange(step, FLAGS.max_steps + 1):
batch_images, batch_labels, batch_images_total = cuhk03_dataset_label2.read_data(
FLAGS.data_dir, 'train', tarin_num_id, IMAGE_WIDTH,
IMAGE_HEIGHT, FLAGS.batch_size, FLAGS.ID_num,
FLAGS.IMG_PER_ID)
#feed_dict = {learning_rate: lr, is_train: True , labels: batch_labels, droup_is_training: False, train_mode: True, images_total: batch_images_total} #no label images: batch_images,
feed_dict = {
learning_rate: lr,
is_train: True,
train_mode: True,
images_total: batch_images_total
}
#sess.run([train], feed_dict=feed_dict)
#train_loss,posi_loss,nega_loss = sess.run([loss,positives,negatives], feed_dict=feed_dict)
#train_loss,posi_loss,nega_loss = sess.run([loss,positives,negatives], feed_dict=feed_dict)
#print('Step: %d, Learning rate: %f, Train loss: %f , posi:%f, nega:%f' % (i, lr, train_loss,posi_loss,nega_loss))
_, train_loss = sess.run([train, loss], feed_dict=feed_dict)
print('Step: %d, Learning rate: %f, Train loss: %f ' %
(i, lr, train_loss))
h, p, l = sess.run([NN, PP, loss], feed_dict=feed_dict)
print 'n:', h
print 'p:', p
print 'hard loss', l
#train_loss,pos , neg , soft , tri = sess.run([loss,pos_loss,neg_loss,loss_softmax,loss_triplet], feed_dict=feed_dict)
#print 'triplet : ',tri
#print("pos_loss : ",pos)
#print("neg_loss : ",neg)
#print 'softmax : ',soft
'''
local_loss ,local_p , local_n = sess.run([local_loss_triplet,local_pos_loss,local_neg_loss], feed_dict=feed_dict)
print 'local_loss : ',local_loss
print 'local p : ',local_p
print 'local n : ',local_n
print ' '
a, aa = sess.run([local_pos_dis,local_neg_dis], feed_dict=feed_dict)
print 'a',a
print ' '
print 'aa',aa
'''
lr = FLAGS.learning_rate * ((0.0001 * i + 1)**-0.75)
if i % 100 == 0:
saver.save(sess, FLAGS.logs_dir + 'model.ckpt', i)
# test save
#vgg.save_npy(sess, './big.npy')
elif FLAGS.mode == 'val':
total = 0.
for _ in xrange(10):
batch_images, batch_labels = cuhk03_dataset_label2.read_data(
FLAGS.data_dir, 'val', val_num_id, IMAGE_WIDTH,
IMAGE_HEIGHT, FLAGS.batch_size)
feed_dict = {
images: batch_images,
labels: batch_labels,
is_train: False
}
prediction = sess.run(inference, feed_dict=feed_dict)
prediction = np.argmax(prediction, axis=1)
label = np.argmax(batch_labels, axis=1)
for i in xrange(len(prediction)):
if prediction[i] == label[i]:
total += 1
print('Accuracy: %f' % (total / (FLAGS.batch_size * 10)))
'''
for i in xrange(len(prediction)):
print('Prediction: %s, Label: %s' % (prediction[i] == 0, labels[i] == 0))
image1 = cv2.cvtColor(batch_images[0][i], cv2.COLOR_RGB2BGR)
image2 = cv2.cvtColor(batch_images[1][i], cv2.COLOR_RGB2BGR)
image = np.concatenate((image1, image2), axis=1)
cv2.imshow('image', image)
key = cv2.waitKey(0)
if key == 1048603: # ESC key
break
'''
elif FLAGS.mode == 'cmc':
cmc_total = []
do_times = 20
cmc_sum = np.zeros((100, 100), dtype='f')
for times in xrange(do_times):
path = 'data'
set = 'val'
cmc_array = np.ones((100, 100), dtype='f')
batch_images = []
batch_labels = []
index_gallery_array = np.ones((1, 100), dtype='f')
gallery_bool = True
probe_bool = True
for j in xrange(100):
id_probe = j
for i in xrange(100):
batch_images = []
batch_labels = []
filepath = ''
#filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, i, index_gallery)
#filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, id_probe, index_probe)
if gallery_bool == True:
while True:
index_gallery = int(random.random() * 10)
index_gallery_array[0, i] = index_gallery
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (
path, set, i, index_gallery)
if not os.path.exists(filepath_gallery):
continue
break
if i == 99:
gallery_bool = False
if gallery_bool == False:
index_gallery = index_gallery_array[0, i]
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (
path, set, i, index_gallery)
if probe_bool == True:
while True:
index_probe = int(random.random() * 10)
filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (
path, set, id_probe, index_probe)
if not os.path.exists(filepath_probe):
continue
if index_gallery_array[
0, id_probe] == index_probe:
continue
probe_bool = False
break
if i == 99:
probe_bool = True
'''
while True:
index_probe = int(random.random() * 10)
filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, id_probe, index_probe)
if not os.path.exists(filepath_gallery):
continue
if index_gallery_array[1,id_probe] == index_probe:
continue
break
'''
#filepath_gallery = 'data/labeled/val/0000_01.jpg'
#filepath_probe = 'data/labeled/val/0000_02.jpg'
image1 = cv2.imread(filepath_gallery)
image1 = cv2.resize(image1,
(IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
image1 = np.reshape(
image1,
(1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
image2 = cv2.imread(filepath_probe)
image2 = cv2.resize(image2,
(IMAGE_WIDTH, IMAGE_HEIGHT))
image2 = cv2.cvtColor(image2, cv2.COLOR_BGR2RGB)
image2 = np.reshape(
image2,
(1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
test_images = np.array([image1, image2, image2])
#print (filepath_gallery)
#print (filepath_probe)
#print ('1111111111111111111111')
if i == j:
batch_labels = [1., 0.]
if i != j:
batch_labels = [0., 1.]
batch_labels = np.array(batch_labels)
print('test img :', test_images.shape)
feed_dict = {images: test_images, is_train: False}
prediction = sess.run(DD, feed_dict=feed_dict)
#print (prediction, prediction[0][1])
print(filepath_gallery, filepath_probe)
#print(bool(not np.argmax(prediction[0])))
print(prediction)
cmc_array[j, i] = prediction
#print(i,j)
#prediction = sess.run(inference, feed_dict=feed_dict)
#prediction = np.argmax(prediction, axis=1)
#label = np.argmax(batch_labels, axis=1)
cmc_score = cmc.cmc(cmc_array)
cmc_sum = cmc_score + cmc_sum
cmc_total.append(cmc_score)
print(cmc_score)
cmc_sum = cmc_sum / do_times
print(cmc_sum)
print('final cmc')
print('\n')
print cmc_total
elif FLAGS.mode == 'test':
image1 = cv2.imread(FLAGS.image1)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
image1 = np.reshape(
image1, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
image2 = cv2.imread(FLAGS.image2)
image2 = cv2.resize(image2, (IMAGE_WIDTH, IMAGE_HEIGHT))
image2 = cv2.cvtColor(image2, cv2.COLOR_BGR2RGB)
image2 = np.reshape(
image2, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
test_images = np.array([image1, image2, image2])
feed_dict = {
images: test_images,
is_train: False,
droup_is_training: False
}
#prediction, prediction2 = sess.run([DD,DD2], feed_dict=feed_dict)
prediction = sess.run([inference], feed_dict=feed_dict)
prediction = np.array(prediction)
print prediction.shape
print(np.argmax(prediction[0]) + 1)
def main(argv=None):
if FLAGS.mode == 'test':
FLAGS.batch_size = 1
if FLAGS.mode == 'cmc':
FLAGS.batch_size = 1
if FLAGS.mode == 'top1':
FLAGS.batch_size = 100
learning_rate = tf.placeholder(tf.float32, name='learning_rate')
images = tf.placeholder(tf.float32, [3, FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images')
images_total = tf.placeholder(tf.float32, [FLAGS.batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images_total')
images_one = tf.placeholder(tf.float32, [1, IMAGE_HEIGHT, IMAGE_WIDTH, 3], name='images_one')
is_train = tf.placeholder(tf.bool, name='is_train')
global_step = tf.Variable(0, name='global_step', trainable=False)
weight_decay = 0.0005
tarin_num_id = 0
val_num_id = 0
if FLAGS.mode == 'train':
tarin_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'train')
print(tarin_num_id, ' 11111111111111111111 1111111111111111')
elif FLAGS.mode == 'val':
val_num_id = cuhk03_dataset_label2.get_num_id(FLAGS.data_dir, 'val')
images1, images2,images3 = preprocess(images, is_train)
img_combine = tf.concat([images1, images2,images3], 0)
train_mode = tf.placeholder(tf.bool)
# Create the model and an embedding head.
model = import_module('nets.' + 'resnet_v1_50')
head = import_module('heads.' + 'fc1024')
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
endpoints, body_prefix = model.endpoints(images_total, is_training=False)
feat = endpoints['resnet_v1_50/block4']# (bt,7,7,2048)
#feat1 ,feat2 ,feat3 = tf.split(feat, [FLAGS.batch_size, FLAGS.batch_size,FLAGS.batch_size])
print('Build network')
feat_1x1 = tf.layers.conv2d(feat, 2048, [1, 1],padding='valid',
kernel_regularizer=tf.contrib.layers.l2_regularizer(weight_decay), reuse=None, name='conv1x1')
anchor_feature = part_attend(feat_1x1, weight_decay)
lr = FLAGS.learning_rate
#config=tf.ConfigProto(log_device_placement=True)
#config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
# GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.55
with tf.Session(config=config) as sess:
checkpoint_saver = tf.train.Saver(max_to_keep=0)
ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
if ckpt and ckpt.model_checkpoint_path:
print('Restore model')
print ckpt.model_checkpoint_path
#saver.restore(sess, ckpt.model_checkpoint_path)
checkpoint_saver.restore(sess, ckpt.model_checkpoint_path)
#for first , training load imagenet
else:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(model_variables)
print FLAGS.initial_checkpoint
saver.restore(sess, FLAGS.initial_checkpoint)
if FLAGS.mode == 'train':
step = sess.run(global_step)
for i in xrange(step, FLAGS.max_steps + 1):
batch_images, batch_labels, batch_images_total = cuhk03_dataset_label2.read_data(FLAGS.data_dir, 'train', tarin_num_id,
IMAGE_WIDTH, IMAGE_HEIGHT, FLAGS.batch_size,FLAGS.ID_num,FLAGS.IMG_PER_ID)
_,train_loss = sess.run([train,loss], feed_dict=feed_dict)
print('Step: %d, Learning rate: %f, Train loss: %f ' % (i, lr, train_loss))
h,p,l = sess.run([NN,PP,loss], feed_dict=feed_dict)
print 'n:',h
print 'p:',p
print 'hard loss',l
lr = FLAGS.learning_rate * ((0.0001 * i + 1) ** -0.75)
if i % 100 == 0:
saver.save(sess, FLAGS.logs_dir + 'model.ckpt', i)
# test save
#vgg.save_npy(sess, './big.npy')
elif FLAGS.mode == 'val':
total = 0.
for _ in xrange(10):
batch_images, batch_labels = cuhk03_dataset_label2.read_data(FLAGS.data_dir, 'val', val_num_id,
IMAGE_WIDTH, IMAGE_HEIGHT, FLAGS.batch_size)
feed_dict = {images: batch_images, labels: batch_labels, is_train: False}
prediction = sess.run(inference, feed_dict=feed_dict)
prediction = np.argmax(prediction, axis=1)
label = np.argmax(batch_labels, axis=1)
for i in xrange(len(prediction)):
if prediction[i] == label[i]:
total += 1
print('Accuracy: %f' % (total / (FLAGS.batch_size * 10)))
'''
for i in xrange(len(prediction)):
print('Prediction: %s, Label: %s' % (prediction[i] == 0, labels[i] == 0))
image1 = cv2.cvtColor(batch_images[0][i], cv2.COLOR_RGB2BGR)
image2 = cv2.cvtColor(batch_images[1][i], cv2.COLOR_RGB2BGR)
image = np.concatenate((image1, image2), axis=1)
cv2.imshow('image', image)
key = cv2.waitKey(0)
if key == 1048603: # ESC key
break
'''
elif FLAGS.mode == 'cmc':
cmc_total=[]
do_times = 20
cmc_sum=np.zeros((100, 100), dtype='f')
for times in xrange(do_times):
path = 'data'
set = 'val'
cmc_array=np.ones((100, 100), dtype='f')
batch_images = []
batch_labels = []
index_gallery_array=np.ones((1, 100), dtype='f')
gallery_bool = True
probe_bool = True
for j in xrange(100):
id_probe = j
for i in xrange(100):
batch_images = []
batch_labels = []
filepath = ''
#filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, i, index_gallery)
#filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, id_probe, index_probe)
if gallery_bool == True:
while True:
index_gallery = int(random.random() * 10)
index_gallery_array[0,i] = index_gallery
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, i, index_gallery)
if not os.path.exists(filepath_gallery):
continue
break
if i ==99:
gallery_bool = False
if gallery_bool == False:
index_gallery = index_gallery_array[0,i]
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, i, index_gallery)
if probe_bool == True:
while True:
index_probe = int(random.random() * 10)
filepath_probe = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, id_probe, index_probe)
if not os.path.exists(filepath_probe):
continue
if index_gallery_array[0,id_probe] == index_probe:
continue
probe_bool = False
break
if i ==99:
probe_bool = True
image1 = cv2.imread(filepath_gallery)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
image1 = np.reshape(image1, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
image2 = cv2.imread(filepath_probe)
image2 = cv2.resize(image2, (IMAGE_WIDTH, IMAGE_HEIGHT))
image2 = cv2.cvtColor(image2, cv2.COLOR_BGR2RGB)
image2 = np.reshape(image2, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
test_images = np.array([image1, image2, image2])
if i == j:
batch_labels = [1., 0.]
if i != j:
batch_labels = [0., 1.]
batch_labels = np.array(batch_labels)
print('test img :',test_images.shape)
feed_dict = {images: test_images, is_train: False}
prediction = sess.run(DD, feed_dict=feed_dict)
print (filepath_gallery,filepath_probe)
print (prediction)
cmc_array[j,i] = prediction
cmc_score = cmc.cmc(cmc_array)
cmc_sum = cmc_score + cmc_sum
cmc_total.append(cmc_score)
print(cmc_score)
cmc_sum = cmc_sum/do_times
print(cmc_sum)
print('final cmc')
print ('\n')
print cmc_total
elif FLAGS.mode == 'top1':
path = 'data'
set = 'val'
cmc_sum=np.zeros((100, 100), dtype='f')
cmc_total = []
do_times = 20
for times in xrange(do_times):
query_feature = []
test_feature = []
for i in range(100):
while True:
index_gallery = int(random.random() * 10)
index_temp = index_gallery
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, i, index_gallery)
if not os.path.exists(filepath_gallery):
continue
break
image1 = cv2.imread(filepath_gallery)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
query_feature.append(image1)
while True:
index_gallery = int(random.random() * 10)
if index_temp == index_gallery:
continue
filepath_gallery = '%s/labeled/%s/%04d_%02d.jpg' % (path, set, i, index_gallery)
if not os.path.exists(filepath_gallery):
continue
break
image1 = cv2.imread(filepath_gallery)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
test_feature.append(image1)
#print filepath_gallery,'\n'
query_feature = np.array(query_feature)
test_feature = np.array(test_feature)
feed_dict = {images_total: query_feature, is_train: False}
q_feat = sess.run(anchor_feature, feed_dict=feed_dict)
feed_dict = {images_total: test_feature, is_train: False}
test_feat = sess.run(anchor_feature, feed_dict=feed_dict)
cmc_array = []
tf_q_feat = tf.constant(q_feat)
tf_test_feat = tf.constant(test_feat)
h = tf.placeholder(tf.int32)
pick = tf_q_feat[h]
tf_q_feat = tf.reshape(pick,[1,2048])
feat1 = tf.tile(tf_q_feat,[100,1])
f = tf.square(tf.subtract(feat1 , tf_test_feat))
d = tf.sqrt(tf.reduce_sum(f,1)) # What about the axis ???
for t in range(100):
feed_dict = {h: t}
D = sess.run(d,feed_dict=feed_dict)
cmc_array.append(D)
cmc_array = np.array(cmc_array)
cmc_score = cmc.cmc(cmc_array)
cmc_sum = cmc_score + cmc_sum
cmc_total.append(cmc_score)
#top1=single_query(q_feat,test_feat,labels,labels,test_num=10)
print cmc_score
cmc_sum = cmc_sum/do_times
print(cmc_sum)
print('final cmc')
print ('\n')
print cmc_total
elif FLAGS.mode == 'test':
image1 = cv2.imread(FLAGS.image1)
image1 = cv2.resize(image1, (IMAGE_WIDTH, IMAGE_HEIGHT))
image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
image1 = np.reshape(image1, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
image2 = cv2.imread(FLAGS.image2)
image2 = cv2.resize(image2, (IMAGE_WIDTH, IMAGE_HEIGHT))
image2 = cv2.cvtColor(image2, cv2.COLOR_BGR2RGB)
image2 = np.reshape(image2, (1, IMAGE_HEIGHT, IMAGE_WIDTH, 3)).astype(float)
test_images = np.array([image1, image2,image2])
feed_dict = {images: test_images, is_train: False, droup_is_training: False}
#prediction, prediction2 = sess.run([DD,DD2], feed_dict=feed_dict)
prediction = sess.run([inference], feed_dict=feed_dict)
prediction = np.array(prediction)
print prediction.shape
print( np.argmax(prediction[0])+1)