def inference(self):
with tf.variable_scope('conv1') as scope:
kernel = tf.get_variable(name='weight',
shape=[3, 3, 3, 32],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer1_conv/weights:0']))
biases = tf.get_variable(name='bias',
shape=[32],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer1_conv/biases:0']))
net = tf.nn.conv2d(self.input, filter=kernel, strides=[1, 1, 1, 1], padding='SAME')
net = tf.nn.bias_add(net, bias=biases)
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('pool1') as scope:
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print_activations(net)
with tf.variable_scope('conv2') as scope:
kernel = tf.get_variable(name='weight',
shape=[3, 3, 32, 64],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer2_conv/weights:0']))
biases = tf.get_variable(name='bias',
shape=[64],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer2_conv/biases:0']))
net = tf.nn.conv2d(net, filter=kernel, strides=[1, 1, 1, 1], padding='SAME')
net = tf.nn.bias_add(net, bias=biases)
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('pool2') as scope:
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print_activations(net)
with tf.variable_scope('conv3') as scope:
kernel = tf.get_variable(name='weight',
shape=[3, 3, 64, 128],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer3_conv/weights:0']))
biases = tf.get_variable(name='bias',
shape=[128],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer3_conv/biases:0']))
net = tf.nn.conv2d(net, filter=kernel, strides=[1, 1, 1, 1], padding='SAME')
net = tf.nn.bias_add(net, bias=biases)
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('pool3') as scope:
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print_activations(net)
with tf.variable_scope('conv4') as scope:
kernel = tf.get_variable(name='weight',
shape=[3, 3, 128, 256],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer4_conv/weights:0']))
biases = tf.get_variable(name='bias',
shape=[256],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer4_conv/biases:0']))
net = tf.nn.conv2d(net, filter=kernel, strides=[1, 1, 1, 1], padding='SAME')
net = tf.nn.bias_add(net, bias=biases)
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('pool4') as scope:
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print_activations(net)
with tf.variable_scope('conv5') as scope:
kernel = tf.get_variable(name='weight',
shape=[3, 3, 256, 256],
dtype=tf.float32, initializer=tf.truncated_normal_initializer(stddev=0.01))
biases = tf.get_variable(name='bias',
shape=[256],
dtype=tf.float32, initializer=tf.truncated_normal_initializer(stddev=0.01))
net = tf.nn.conv2d(net, filter=kernel, strides=[1, 1, 1, 1], padding='SAME')
net = tf.nn.bias_add(net, bias=biases)
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('pool5') as scope:
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print_activations(net)
with tf.variable_scope('fc1') as scope:
net = tf.reshape(net, [-1, 3 * 3 * 256])
weight = tf.get_variable(name='weight',
shape=[3 * 3 * 256, 512],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer5_fc/weights:0']))
biases = tf.get_variable(name='bias',
shape=[512],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer5_fc/biases:0']))
net = tf.matmul(net, weight) + biases
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('fc2') as scope:
weight = tf.get_variable(name='weight',
shape=[512, 256],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer6_fc/weights:0']))
biases = tf.get_variable(name='bias',
shape=[256],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer6_fc/biases:0']))
net = tf.matmul(net, weight) + biases
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('fc3') as scope:
weight = tf.get_variable(name='weight',
shape=[256, 28],
dtype=tf.float32, initializer=tf.truncated_normal_initializer(stddev=0.01))
biases = tf.get_variable(name='bias',
shape=[28],
dtype=tf.float32, initializer=tf.truncated_normal_initializer(stddev=0.01))
net = tf.matmul(net, weight) + biases
net = tf.nn.relu(net)
print_activations(net)
net = tf.reshape(net, [-1, 2, 2, 7])
print_activations(net)
self.predict = net
return net
class hugenet():
def __init__(self, variables):
self.pretrain_variables = variables
self.lambda_noobj = FLAGS.lambda_noobj
self.lambda_coord = FLAGS.lambda_coord
self.label = tf.placeholder(tf.float32, shape=[None, 2, 2, 7])
self.input = tf.placeholder(tf.float32, [None, 100, 100, 3])
def inference(self):
with tf.variable_scope('conv1') as scope:
kernel = tf.get_variable(name='weight',
shape=[3, 3, 3, 32],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer1_conv/weights:0']))
biases = tf.get_variable(name='bias',
shape=[32],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer1_conv/biases:0']))
net = tf.nn.conv2d(self.input, filter=kernel, strides=[1, 1, 1, 1], padding='SAME')
net = tf.nn.bias_add(net, bias=biases)
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('pool1') as scope:
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print_activations(net)
with tf.variable_scope('conv2') as scope:
kernel = tf.get_variable(name='weight',
shape=[3, 3, 32, 64],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer2_conv/weights:0']))
biases = tf.get_variable(name='bias',
shape=[64],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer2_conv/biases:0']))
net = tf.nn.conv2d(net, filter=kernel, strides=[1, 1, 1, 1], padding='SAME')
net = tf.nn.bias_add(net, bias=biases)
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('pool2') as scope:
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print_activations(net)
with tf.variable_scope('conv3') as scope:
kernel = tf.get_variable(name='weight',
shape=[3, 3, 64, 128],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer3_conv/weights:0']))
biases = tf.get_variable(name='bias',
shape=[128],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer3_conv/biases:0']))
net = tf.nn.conv2d(net, filter=kernel, strides=[1, 1, 1, 1], padding='SAME')
net = tf.nn.bias_add(net, bias=biases)
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('pool3') as scope:
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print_activations(net)
with tf.variable_scope('conv4') as scope:
kernel = tf.get_variable(name='weight',
shape=[3, 3, 128, 256],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer4_conv/weights:0']))
biases = tf.get_variable(name='bias',
shape=[256],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer4_conv/biases:0']))
net = tf.nn.conv2d(net, filter=kernel, strides=[1, 1, 1, 1], padding='SAME')
net = tf.nn.bias_add(net, bias=biases)
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('pool4') as scope:
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print_activations(net)
with tf.variable_scope('conv5') as scope:
kernel = tf.get_variable(name='weight',
shape=[3, 3, 256, 256],
dtype=tf.float32, initializer=tf.truncated_normal_initializer(stddev=0.01))
biases = tf.get_variable(name='bias',
shape=[256],
dtype=tf.float32, initializer=tf.truncated_normal_initializer(stddev=0.01))
net = tf.nn.conv2d(net, filter=kernel, strides=[1, 1, 1, 1], padding='SAME')
net = tf.nn.bias_add(net, bias=biases)
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('pool5') as scope:
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print_activations(net)
with tf.variable_scope('fc1') as scope:
net = tf.reshape(net, [-1, 3 * 3 * 256])
weight = tf.get_variable(name='weight',
shape=[3 * 3 * 256, 512],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer5_fc/weights:0']))
biases = tf.get_variable(name='bias',
shape=[512],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer5_fc/biases:0']))
net = tf.matmul(net, weight) + biases
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('fc2') as scope:
weight = tf.get_variable(name='weight',
shape=[512, 256],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer6_fc/weights:0']))
biases = tf.get_variable(name='bias',
shape=[256],
dtype=tf.float32, initializer=tf.constant_initializer(self.pretrain_variables['layer6_fc/biases:0']))
net = tf.matmul(net, weight) + biases
net = tf.nn.relu(net)
print_activations(net)
with tf.variable_scope('fc3') as scope:
weight = tf.get_variable(name='weight',
shape=[256, 28],
dtype=tf.float32, initializer=tf.truncated_normal_initializer(stddev=0.01))
biases = tf.get_variable(name='bias',
shape=[28],
dtype=tf.float32, initializer=tf.truncated_normal_initializer(stddev=0.01))
net = tf.matmul(net, weight) + biases
net = tf.nn.relu(net)
print_activations(net)
net = tf.reshape(net, [-1, 2, 2, 7])
print_activations(net)
self.predict = net
return net
def train(self, net):
pobj_pred = net[:, :, :, 0:1]
pobj_truth = self.label[:, :, :, 0:1]
print_activations(pobj_pred)
print_activations(pobj_truth)
box_pred = net[:, :, :, 1:5]
box_truth = self.label[:, :, :, 1:5]
print_activations(box_pred)
print_activations(box_truth)
pclass_pred = net[:, :, :, 5:]
pclass_truth = self.label[:, :, :, 5:]
print_activations(pclass_pred)
print_activations(pclass_truth)
IOU = cacl_iou(box1=box_pred, box2=box_truth)
print_activations(IOU)
object_mask = tf.cast(IOU >= 0.5, tf.float32) * pobj_truth
print_activations(object_mask)
noonject_mask = tf.ones_like(object_mask, dtype=tf.float32) - object_mask
print_activations(noonject_mask)
loss_obj = tf.reduce_sum(tf.square(pobj_truth - pobj_pred) * object_mask)
print_activations(loss_obj)
loss_noobj = tf.reduce_sum(tf.square(pobj_truth - pobj_pred) * noonject_mask)
print_activations(loss_noobj)
loss_coord = tf.reduce_sum(tf.reduce_sum(tf.square(box_pred[:, :, :,0:2]
- box_truth[:, :, :, 0:2]), axis=3, keep_dims=True) * object_mask)
+ tf.reduce_sum(tf.reduce_sum(tf.square(tf.sqrt(box_pred[:, :, :, 2:4])
- tf.sqrt(box_truth[:, :, :, 2:4])), axis=3, keep_dims=True) * object_mask)
print_activations(loss_coord)
print_activations(box_pred[:, :, :, 2:])
print_activations(box_pred[:, :, :, 0:2])
loss_class = tf.reduce_sum(tf.square(pclass_pred - pclass_truth) * pobj_truth)
print_activations(loss_class)
total_loss = loss_obj + self.lambda_noobj * loss_noobj + self.lambda_coord * loss_coord + loss_class
train_step = tf.train.AdamOptimizer(1e-2).minimize(total_loss / FLAGS.train_batch_size)
image_batch_raw, label_batch_raw, filename_batch_raw = car_input(is_random=True, is_train=True)
image_batch_test, label_batch_test, filename_batch_test = car_input(is_random=False, is_train=False)
train_IOU = tf.reduce_mean(IOU)
with tf.Session(config=config) as sess:
tf.global_variables_initializer().run()
tf.local_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
for i in range(int(FLAGS.num_epoches * FLAGS.train_set_size / FLAGS.train_batch_size)):
image_batch, label_batch = sess.run([image_batch_raw, label_batch_raw])
sess.run([train_step], feed_dict={self.input: image_batch,
self.label: label_batch})
if not i % 10:
predict, loss, train_iou, iou = sess.run([net, total_loss, train_IOU, IOU],
feed_dict={
self.input: image_batch,
self.label: label_batch})
print('step %d, loss: %.4f, train_IOU: %.4f ' % (i, loss, train_iou))
image_test, label_test = sess.run([image_batch_test, label_batch_test])
loss = sess.run(total_loss, feed_dict={self.input: image_test, self.label: label_test})
print(loss / FLAGS.test_batch_size)
coord.request_stop()
coord.join(threads)
def inference(self):
with slim.arg_scope([slim.conv2d],
kernel_size=3,
reuse=None,
trainable=True):
with slim.arg_scope([slim.max_pool2d],
kernel_size=2,
stride=2,
padding='VALID'):
net = slim.conv2d(self.input_image, 32, scope='layer1_conv')
print_activations(net)
net = slim.max_pool2d(net, scope='pool1')
print_activations(net)
net = slim.conv2d(net, 64, scope='layer2_conv')
print_activations(net)
net = slim.max_pool2d(net, scope='pool2')
print_activations(net)
net = slim.conv2d(net, 128, scope='layer3_conv')
print_activations(net)
net = slim.max_pool2d(net, scope='pool3')
print_activations(net)
net = slim.conv2d(net, 256, scope='layer4_conv')
print_activations(net)
net = slim.max_pool2d(net, scope='pool4')
print_activations(net)
net = tf.reshape(net, [-1, 3 * 3 * 256])
net = slim.fully_connected(net, 512, scope='layer5_fc')
print_activations(net)
net = slim.fully_connected(net, 256, scope='layer6_fc')
print_activations(net)
net=slim.fully_connected(net,self.num_class,activation_fn=None, scope='layer7_fc')
print_activations(net)
net = slim.softmax(net)
print_activations(net)
return net
def train(self, net):
pobj_pred = net[:, :, :, 0:1]
pobj_truth = self.label[:, :, :, 0:1]
print_activations(pobj_pred)
print_activations(pobj_truth)
box_pred = net[:, :, :, 1:5]
box_truth = self.label[:, :, :, 1:5]
print_activations(box_pred)
print_activations(box_truth)
pclass_pred = net[:, :, :, 5:]
pclass_truth = self.label[:, :, :, 5:]
print_activations(pclass_pred)
print_activations(pclass_truth)
IOU = cacl_iou(box1=box_pred, box2=box_truth)
print_activations(IOU)
object_mask = tf.cast(IOU >= 0.5, tf.float32) * pobj_truth
print_activations(object_mask)
noonject_mask = tf.ones_like(object_mask, dtype=tf.float32) - object_mask
print_activations(noonject_mask)
loss_obj = tf.reduce_sum(tf.square(pobj_truth - pobj_pred) * object_mask)
print_activations(loss_obj)
loss_noobj = tf.reduce_sum(tf.square(pobj_truth - pobj_pred) * noonject_mask)
print_activations(loss_noobj)
loss_coord = tf.reduce_sum(tf.reduce_sum(tf.square(box_pred[:, :, :,0:2]
- box_truth[:, :, :, 0:2]), axis=3, keep_dims=True) * object_mask)
+ tf.reduce_sum(tf.reduce_sum(tf.square(tf.sqrt(box_pred[:, :, :, 2:4])
- tf.sqrt(box_truth[:, :, :, 2:4])), axis=3, keep_dims=True) * object_mask)
def inference(images, classes, batch_size):
"""Build the AlexNet logs.
Args:
batch_size: batch size
images: Images Tensor
classes: Image classes
Returns:
out: prediction image label.
"""
# conv1
with tf.name_scope('conv1') as scope:
kernel = tf.Variable(tf.truncated_normal([11, 11, 3, 64]),
name='weights')
conv = tf.nn.conv2d(images, kernel, [1, 4, 4, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.1,
shape=[64],
dtype=utils.data_type()),
name='biases')
bias = tf.nn.bias_add(conv, biases)
conv1 = tf.nn.relu(bias, name=scope)
utils.print_activations(conv1)
# lrn1
with tf.name_scope('lrn1') as scope:
lrn1 = tf.nn.lrn(conv1,
alpha=1e-4,
beta=0.75,
depth_radius=2,
bias=2.0)
# pool1
pool1 = tf.nn.max_pool(lrn1,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='VALID',
name='pool1')
utils.print_activations(pool1)
# conv2
with tf.name_scope('conv2') as scope:
kernel = tf.Variable(tf.truncated_normal([5, 5, 64, 192]),
name='weights')
conv = tf.nn.conv2d(pool1, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.1,
shape=[192],
dtype=utils.data_type()),
name='biases')
bias = tf.nn.bias_add(conv, biases)
conv2 = tf.nn.relu(bias, name=scope)
utils.print_activations(conv2)
# lrn2
with tf.name_scope('lrn2') as scope:
lrn2 = tf.nn.lrn(conv2,
alpha=1e-4,
beta=0.75,
depth_radius=2,
bias=2.0)
# pool2
pool2 = tf.nn.max_pool(lrn2,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='VALID',
name='pool2')
utils.print_activations(pool2)
# conv3
with tf.name_scope('conv3') as scope:
kernel = tf.Variable(tf.truncated_normal([3, 3, 192, 384]),
name='weights')
conv = tf.nn.conv2d(pool2, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.1,
shape=[384],
dtype=utils.data_type()),
name='biases')
bias = tf.nn.bias_add(conv, biases)
conv3 = tf.nn.relu(bias, name=scope)
utils.print_activations(conv3)
# conv4
with tf.name_scope('conv4') as scope:
kernel = tf.Variable(tf.truncated_normal([3, 3, 384, 256]),
name='weights')
conv = tf.nn.conv2d(conv3, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.1,
shape=[256],
dtype=utils.data_type()),
name='biases')
bias = tf.nn.bias_add(conv, biases)
conv4 = tf.nn.relu(bias, name=scope)
utils.print_activations(conv4)
# conv5
with tf.name_scope('conv5') as scope:
kernel = tf.Variable(tf.truncated_normal([3, 3, 256, 256]),
name='weights')
conv = tf.nn.conv2d(conv4, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.1,
shape=[256],
dtype=utils.data_type()),
name='biases')
bias = tf.nn.bias_add(conv, biases)
conv5 = tf.nn.relu(bias, name=scope)
utils.print_activations(conv5)
# pool5
pool5 = tf.nn.max_pool(conv5,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='VALID',
name='pool5')
utils.print_activations(pool5)
# fully 1
with tf.name_scope('fc1') as scope:
reshape = tf.reshape(pool5, [batch_size, -1])
dim = reshape.get_shape()[1].value
kernel = tf.Variable(tf.truncated_normal([dim, 4096]), name='weights')
biases = tf.Variable(tf.constant(0.1,
shape=[4096],
dtype=utils.data_type()),
name='biases')
fc1 = tf.nn.relu(tf.matmul(reshape, kernel) + biases)
# fully 2
with tf.name_scope('fc2') as scope:
kernel = tf.Variable(tf.truncated_normal([4096, 4096]), name='weights')
biases = tf.Variable(tf.constant(0.1,
shape=[4096],
dtype=utils.data_type()),
name='biases')
fc2 = tf.nn.relu(tf.matmul(fc1, kernel) + biases)
# fully 3
with tf.name_scope('fc3') as scope:
kernel = tf.Variable(tf.truncated_normal([4096, classes]),
name='weights')
biases = tf.Variable(tf.constant(0.1,
shape=[classes],
dtype=utils.data_type()),
name='biases')
out = tf.add(tf.matmul(fc2, kernel), biases)
return out