def main(_):
with tf.device('/gpu:0'):
sess = tf.Session()
query_image = tf.placeholder(tf.float32, [None, 224, 224, 3],
name='query_image')
train_mode = tf.placeholder(tf.bool, name='train_mode')
model = netvlad.Netvlad(FLAGS.model_path)
model.build(query_image, train_mode)
print("number of total parameters in the model is %d\n" %
model.get_var_count())
sess.run(tf.global_variables_initializer())
print("evaluation begins!\n")
batch = np.zeros((11, 224, 224, 3))
for i in range(1, 12):
batch[i - 1, :, :, :] = eva_utils.load_image(("test/%s.JPG" % (i)))
descriptor = sess.run(model.vlad_output,
feed_dict={
'query_image:0': batch,
'train_mode:0': False
})
A = np.dot(descriptor, descriptor.transpose())
B = np.sum(descriptor**2, axis=1, keepdims=True)
C = np.sum(descriptor**2, axis=1)
print(B + C - 2 * A)
D = B + C - 2 * A
D[D < 0] = 0
L2_distance = np.sqrt(D)
print(L2_distance)
def main(_):
qList, dbList = eva_init.get_List(FLAGS.mat_path)
update_index_every = 600 / FLAGS.batch_size
if FLAGS.initH5:
eva_init.h5_initial(FLAGS.eva_h5File)
if FLAGS.computeDist:
eva_init.compute_dist(FLAGS.mat_path, FLAGS.eva_h5File)
if FLAGS.loadImage:
eva_init.multipro_load_image(FLAGS.data_dir, FLAGS.eva_h5File, qList, dbList)
with tf.device('/gpu:0'):
sess = tf.Session()
query_image = tf.placeholder(tf.float32,[None, 224, 224, 3], name = 'query_image')
train_mode = tf.placeholder(tf.bool, name = 'train_mode')
model = netvlad.Netvlad(FLAGS.model_path)
model.build(query_image, train_mode)
print("number of total parameters in the model is %d\n" % model.get_var_count())
sess.run(tf.global_variables_initializer())
print("evaluation begins!\n")
eva_utils.evaluate(sess, model, FLAGS.batch_size, FLAGS.eva_h5File, qList, dbList, FLAGS.numRecall)
def main(_):
config = tf.ConfigProto(allow_soft_placement = True)
with tf.device('/gpu:0'):
with tf.Session(config = config) as sess:
X = tf.placeholder(tf.float32, [None, 32, 32, 3], name = 'X')
Y = tf.placeholder(tf.int64, [None], name = 'Y')
if FLAGS.use_vlad:
model = netvlad_1.Netvlad(FLAGS.modelPath)
else:
model = netvlad.Netvlad(FLAGS.modelPath)
model.build(X)
print("number of total parameters in the model is %d\n" % model.get_var_count())
correct_prediction = tf.equal(tf.argmax(tf.nn.softmax(model.fc3), axis = 1), Y)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
sess.run(tf.global_variables_initializer())
Acc = 0.0
acc = 0.0
numBatch = 10000 / FLAGS.batch_size
print("evaluation begins!\n")
count = 0.0
for x, y in eva_utils.next_batch(FLAGS.batch_size, 'cifar-10-batches-py'):
count += 1
acc = sess.run(accuracy, feed_dict = {X: x, Y: y})
Acc += acc
if count % FLAGS.print_every == 0:
print("progress: %.4f current accuracy = %.6f total accuracy = %.6f\n" % (count / numBatch, acc, Acc / count))
def main(_):
config = tf.ConfigProto(allow_soft_placement=True)
with tf.device('/gpu:0'):
with tf.Session(config=config) as sess:
X = tf.placeholder(tf.float32, [FLAGS.batch_size, 32, 32, 3],
name='X')
Y = tf.placeholder(tf.int64, [FLAGS.batch_size], name='Y')
if FLAGS.use_vlad:
model = netvlad_1.Netvlad('vgg16.npy')
else:
model = netvlad.Netvlad(FLAGS.modelPath)
model.build(X)
print("number of total parameters in the model is %d\n" %
model.get_var_count())
if FLAGS.use_a_softmax:
print('using angular softmax')
loss = A_softmax.A_softmax(model.fc1, tf.one_hot(Y, depth=10),
model.fc3, FLAGS.m)
else:
print('not using angular softmax')
loss = tf.losses.softmax_cross_entropy(tf.one_hot(Y, depth=10),
model.fc3)
# loss = tf.reduce_mean(tf.reduce_sum(-tf.log(tf.nn.softmax(model.fc3)) * tf.one_hot(Y, depth = 10), axis = -1))
global_step = tf.Variable(0, trainable=False)
starter_learning_rate = FLAGS.lr
learning_rate = tf.train.exponential_decay(starter_learning_rate,
global_step,
100000,
0.96,
staircase=True)
train = tf.train.GradientDescentOptimizer(learning_rate).minimize(
loss)
# output1 = model.fc1
# train = tf.train.RMSPropOptimizer(FLAGS.lr).minimize(loss)
correct_prediction = tf.equal(
tf.argmax(tf.nn.softmax(model.fc3), axis=1), Y)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
sess.run(tf.global_variables_initializer())
train_loss = 0
numBatch = 50000 / FLAGS.batch_size
print("training begins!\n")
for i in range(FLAGS.numEpoch):
count = 0.0
for x, y in train_utils.next_batch(FLAGS.batch_size,
'cifar-10-batches-py'):
count += 1
# assert False
_, train_loss, acc = sess.run([train, loss, accuracy],
feed_dict={
X: x,
Y: y
})
# _, train_loss, acc, out3 = sess.run([train, loss, accuracy, output3], feed_dict = {X: x, Y: y})
if count % FLAGS.print_every == 0:
print(
"Epoch: %s progress: %.4f accuracy = %.4f training_loss = %.6f\n"
% (i, count / numBatch, acc, train_loss))
# print(out1[:10, 0])
# print(out2[:10])
# print(out3[:10, 0])
# print [list(output).count(j) for j in range(10)]
# assert False
if (i + 1) % FLAGS.save_every == 0:
model.save_npy(
sess, "%s/epoch_%d_loss_%.6f" %
(FLAGS.checkpoint_dir, i, train_loss))
def main(_):
qList, dbList = train_init.get_List(FLAGS.mat_path)
update_index_every = 1000 / FLAGS.batch_size
if FLAGS.initH5:
train_init.h5_initial(FLAGS.train_h5File)
if FLAGS.computeDist:
train_init.compute_dist(FLAGS.mat_path, FLAGS.train_h5File)
if FLAGS.initIndex:
train_init.index_initial(FLAGS.train_h5File, qList, dbList)
if FLAGS.loadImage:
train_init.multipro_load_image(FLAGS.data_dir, FLAGS.train_h5File,
qList, dbList)
with tf.device('/gpu:0'):
sess = tf.Session()
query_image = tf.placeholder(tf.float32, [None, 224, 224, 3],
name='query_image')
labels = tf.placeholder(tf.float32, [None, 32768, 60])
train_mode = tf.placeholder(tf.bool, name='train_mode')
model = netvlad.Netvlad('vgg16.npy')
model.build(query_image, train_mode)
print("number of total parameters in the model is %d\n" %
model.get_var_count())
loss, mindistP, mindistN = triplet_loss(model.vlad_output, labels, 0.2)
train = tf.train.RMSPropOptimizer(FLAGS.lr).minimize(loss)
sess.run(tf.global_variables_initializer())
train_loss = 0
count = 0
print("training begins!\n")
for i in range(FLAGS.numEpoch):
for x, y, z in train_utils.next_batch(sess, model,
FLAGS.batch_size,
FLAGS.train_h5File,
FLAGS.randomStartIdx, qList,
dbList):
count = count + 1
if count >= update_index_every:
count = 0
train_utils.index_update(sess, model,
FLAGS.batch_size * 30,
FLAGS.train_h5File, qList, dbList)
_, train_loss, minP, minN = sess.run(
[train, loss, mindistP, mindistN],
feed_dict={
query_image: x,
labels: y,
train_mode: True
})
if count % FLAGS.print_every == 0:
print(
"Epoch: %d progress: %.4f%% training_loss = %.6f\n"
% (i, z, train_loss))
print("Minimum L2 distance of P and N is %s %s\n" %
(minP, minN))
if (i + 1) % FLAGS.save_every == 0:
model.save_npy(
sess, "%s/netvlad_epoch_%d_loss_%.6f" %
(FLAGS.checkpoint_dir, i, train_loss))
update_index_every *= 2