def testCreateFCWithoutWD(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
ops.fc(inputs, 32, weight_decay=0)
self.assertEquals(
tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES), [])
def testCreateFCWithoutWD(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
ops.fc(inputs, 32, weight_decay=0)
self.assertEqual(
tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES), [])
def testNonReuseVars(self):
height, width = 3, 3
inputs = tf.random_uniform((5, height * width * 3), seed=1)
with self.test_session():
ops.fc(inputs, 32)
self.assertEqual(len(variables.get_variables('FC')), 2)
ops.fc(inputs, 32)
self.assertEqual(len(variables.get_variables('FC')), 4)
def testReuseVars(self):
height, width = 3, 3
inputs = tf.random_uniform((5, height * width * 3), seed=1)
with self.test_session():
ops.fc(inputs, 32, scope='fc1')
self.assertEquals(len(variables.get_variables('fc1')), 2)
ops.fc(inputs, 32, scope='fc1', reuse=True)
self.assertEquals(len(variables.get_variables('fc1')), 2)
def testNonReuseVars(self):
height, width = 3, 3
inputs = tf.random_uniform((5, height * width * 3), seed=1)
with self.test_session():
ops.fc(inputs, 32)
self.assertEquals(len(variables.get_variables('FC')), 2)
ops.fc(inputs, 32)
self.assertEquals(len(variables.get_variables('FC')), 4)
def testReuseVars(self):
height, width = 3, 3
inputs = tf.random_uniform((5, height * width * 3), seed=1)
with self.test_session():
ops.fc(inputs, 32, scope='fc1')
self.assertEqual(len(variables.get_variables('fc1')), 2)
ops.fc(inputs, 32, scope='fc1', reuse=True)
self.assertEqual(len(variables.get_variables('fc1')), 2)
def lenet(inputs,
dropout_keep_prob=1.0,
num_classes=10,
is_training=True,
restore_logits=True,
weight_decay=0.0005,
seed=1,
scope=''):
"""LeNet in Caffe https://github.com/BVLC/caffe/blob/master/examples/mnist/lenet_train_test.prototxt
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: Optional scope for name_scope.
Returns:
a list containing 'logits', 'aux_logits' Tensors.
"""
# end_points will collect relevant activations for external use, for example
# summaries or losses.
print ("Warning: batch_norm_params is always None in lenet")
end_points = {}
with tf.name_scope(scope, 'lenet', [inputs]):
with scopes.arg_scope([ops.conv2d, ops.fc, ops.batch_norm, ops.dropout],
is_training=is_training):
with scopes.arg_scope([ops.conv2d, ops.fc],
bias=0.0, batch_norm_params=None, seed=seed):
with scopes.arg_scope([ops.conv2d], stride=1, padding='SAME'):
with scopes.arg_scope([ops.max_pool], stride=2, padding='SAME'):
# 32 x 32 x 3
end_points['conv1'] = ops.conv2d(inputs, 20, [5, 5], stride=1, stddev=0.05,
weight_decay=weight_decay, seed=seed+1, scope='conv1')
end_points['pool1'] = ops.max_pool(end_points['conv1'], [2, 2], scope='pool1')
end_points['conv2'] = ops.conv2d(end_points['pool1'], 50, [5, 5], stride=1, stddev=0.05,
weight_decay=weight_decay, seed=seed+2, scope='conv2')
end_points['pool2'] = ops.max_pool(end_points['conv2'], [2, 2], scope='pool2')
end_points['pool2'] = ops.flatten(end_points['pool2'], scope='flatten')
net = ops.fc(end_points['pool2'], 500, stddev=0.048, weight_decay=weight_decay,
seed = seed +3, scope='fc3')
# Final pooling and prediction
with tf.variable_scope('logits'):
logits = ops.fc(net, num_classes, activation=None, stddev=0.0767, weight_decay=weight_decay,
scope='logits', seed = seed +5, restore=restore_logits)
# 10
end_points['logits'] = logits
end_points['predictions'] = tf.nn.softmax(logits, name='predictions')
# There is no aux_logits for LeNet
end_points['aux_logits'] = tf.constant(0)
return logits, end_points
def testCreateFcCreatesWeightsAndBiasesVars(self):
height, width = 3, 3
inputs = tf.random_uniform((5, height * width * 3), seed=1)
with self.test_session():
self.assertFalse(variables.get_variables('fc1/weights'))
self.assertFalse(variables.get_variables('fc1/biases'))
ops.fc(inputs, 32, scope='fc1')
self.assertTrue(variables.get_variables('fc1/weights'))
self.assertTrue(variables.get_variables('fc1/biases'))
def testReuseFCWithBatchNorm(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height * width * 3), seed=1)
with scopes.arg_scope([ops.fc], batch_norm_params={'decay': 0.9}):
net = ops.fc(images, 27, scope='fc1')
net = ops.fc(net, 27, scope='fc1', reuse=True)
self.assertEquals(len(variables.get_variables()), 4)
self.assertEquals(len(variables.get_variables('fc1/BatchNorm')), 3)
def testCreateFcCreatesWeightsAndBiasesVars(self):
height, width = 3, 3
inputs = tf.random_uniform((5, height * width * 3), seed=1)
with self.test_session():
self.assertFalse(variables.get_variables('fc1/weights'))
self.assertFalse(variables.get_variables('fc1/biases'))
ops.fc(inputs, 32, scope='fc1')
self.assertTrue(variables.get_variables('fc1/weights'))
self.assertTrue(variables.get_variables('fc1/biases'))
def testReuseFCWithBatchNorm(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height * width * 3), seed=1)
with scopes.arg_scope([ops.fc], batch_norm_params={'decay': 0.9}):
net = ops.fc(images, 27, scope='fc1')
net = ops.fc(net, 27, scope='fc1', reuse=True)
self.assertEqual(len(variables.get_variables()), 4)
self.assertEqual(len(variables.get_variables('fc1/BatchNorm')), 3)
def testReuseFCWithWD(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
ops.fc(inputs, 32, weight_decay=0.01, scope='fc')
self.assertEquals(len(tf.get_collection(losses.LOSSES_COLLECTION)), 1)
tf.get_variable_scope().reuse_variables()
ops.fc(inputs, 32, weight_decay=0.01, scope='fc')
self.assertEquals(len(tf.get_collection(losses.LOSSES_COLLECTION)), 1)
def testCreateFCWithWD(self):
height, width = 3, 3
with self.test_session() as sess:
inputs = tf.random_uniform((5, height * width * 3), seed=1)
ops.fc(inputs, 32, weight_decay=0.01)
wd = tf.get_collection(losses.LOSSES_COLLECTION)[0]
self.assertEquals(wd.op.name, 'FC/weights/Regularizer/L2Loss/value')
sess.run(tf.initialize_all_variables())
self.assertTrue(sess.run(wd) <= 0.01)
def testFCWithBatchNorm(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height * width * 3), seed=1)
with scopes.arg_scope([ops.fc], batch_norm_params={}):
net = ops.fc(images, 27)
net = ops.fc(net, 27)
self.assertEqual(len(variables.get_variables()), 8)
self.assertEqual(len(variables.get_variables('FC/BatchNorm')), 3)
self.assertEqual(len(variables.get_variables('FC_1/BatchNorm')), 3)
def testCreateFCWithWD(self):
height, width = 3, 3
with self.test_session() as sess:
inputs = tf.random_uniform((5, height * width * 3), seed=1)
ops.fc(inputs, 32, weight_decay=0.01)
wd = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)[0]
self.assertEquals(wd.op.name,
'FC/weights/Regularizer/L2Regularizer/value')
sess.run(tf.global_variables_initializer())
self.assertTrue(sess.run(wd) <= 0.01)
def testFCWithBatchNorm(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height * width * 3), seed=1)
with scopes.arg_scope([ops.fc], batch_norm_params={}):
net = ops.fc(images, 27)
net = ops.fc(net, 27)
self.assertEquals(len(variables.get_variables()), 8)
self.assertEquals(len(variables.get_variables('FC/BatchNorm')), 3)
self.assertEquals(len(variables.get_variables('FC_1/BatchNorm')), 3)
def testFCWithBatchNorm(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height * width * 3), seed=1)
with scopes.arg_scope([ops.fc], batch_norm_params={}):
net = ops.fc(images, 32, scope='fc1')
net = ops.fc(net, 32, scope='fc2')
self.assertEquals(len(tf.get_collection('moving_vars')), 4)
self.assertEquals(len(variables.get_variables('fc1/BatchNorm')), 3)
self.assertEquals(len(variables.get_variables('fc2/BatchNorm')), 3)
def testFCWithBatchNorm(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height * width * 3), seed=1)
with scopes.arg_scope([ops.fc], batch_norm_params={}):
net = ops.fc(images, 32, scope='fc1')
net = ops.fc(net, 32, scope='fc2')
self.assertEquals(len(tf.get_collection('moving_vars')), 4)
self.assertEquals(len(variables.get_variables('fc1/BatchNorm')), 3)
self.assertEquals(len(variables.get_variables('fc2/BatchNorm')), 3)
def testCreateFCWithWD(self):
height, width = 3, 3
with self.test_session() as sess:
inputs = tf.random_uniform((5, height * width * 3), seed=1)
ops.fc(inputs, 32, weight_decay=0.01)
wd = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)[0]
self.assertEqual(wd.op.name,
'FC/weights/Regularizer/L2Regularizer/value')
sess.run(tf.initialize_all_variables())
self.assertTrue(sess.run(wd) <= 0.01)
def testReuseFCWithWD(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
ops.fc(inputs, 32, weight_decay=0.01, scope='fc')
self.assertEquals(len(tf.get_collection(losses.LOSSES_COLLECTION)),
1)
tf.get_variable_scope().reuse_variables()
ops.fc(inputs, 32, weight_decay=0.01, scope='fc')
self.assertEquals(len(tf.get_collection(losses.LOSSES_COLLECTION)),
1)
def testReuseFCWithWD(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
ops.fc(inputs, 32, weight_decay=0.01, scope='fc')
self.assertEqual(len(variables.get_variables()), 2)
self.assertEqual(
len(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)), 1)
ops.fc(inputs, 32, weight_decay=0.01, scope='fc', reuse=True)
self.assertEqual(len(variables.get_variables()), 2)
self.assertEqual(
len(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)), 1)
def testReuseFCWithWD(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
ops.fc(inputs, 32, weight_decay=0.01, scope='fc')
self.assertEquals(len(variables.get_variables()), 2)
self.assertEquals(
len(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)), 1)
ops.fc(inputs, 32, weight_decay=0.01, scope='fc', reuse=True)
self.assertEquals(len(variables.get_variables()), 2)
self.assertEquals(
len(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)), 1)
def testCreateFC(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
output = ops.fc(inputs, 32)
self.assertEquals(output.op.name, 'FC/Relu')
self.assertListEqual(output.get_shape().as_list(), [5, 32])
def inference(images):
"""Build the CIFAR-10 model.
Args:
images: Images returned from distorted_inputs() or inputs().
Returns:
Logits.
"""
# We instantiate all variables using tf.get_variable() instead of
# tf.Variable() in order to share variables across multiple GPU training runs.
# If we only ran this model on a single GPU, we could simplify this function
# by replacing all instances of tf.get_variable() with tf.Variable().
#
with scopes.arg_scope([ops.conv2d, ops.fc], stddev=0.1, bias=0.1, batch_norm_params={}):
# with scopes.arg_scope([ops.conv2d, ops.fc], stddev=0.1, bias=0.1):
with scopes.arg_scope([ops.conv2d], kernel_size=[3,3], padding='SAME'):
with scopes.arg_scope([ops.max_pool], kernel_size=[3,3], padding='SAME'):
net = ops.conv2d(images, num_filters_out=64)
net = ops.conv2d(net, num_filters_out=64)
net = ops.max_pool(net)
net = ops.conv2d(net, num_filters_out=128)
net = ops.conv2d(net, num_filters_out=128)
net = ops.max_pool(net)
net = ops.conv2d(net, num_filters_out=256)
net = ops.conv2d(net, num_filters_out=256)
net = ops.max_pool(net)
net = ops.conv2d(net, num_filters_out=512)
net = ops.conv2d(net, num_filters_out=512)
net = ops.avg_pool(net, kernel_size=[3,3], padding='SAME')
net = ops.flatten(net)
# net = ops.fc(net, num_units_out=1024)
# net = ops.fc(net, num_units_out=256)
net = ops.fc(net, num_units_out=10)
return net
def nin(inputs,
num_classes=10,
is_training=True,
restore_logits=True,
scope=''):
# end_points will collect relevant activations for external use, for example
# summaries or losses.
end_points = {}
with tf.op_scope([inputs], scope, 'nin'):
with scopes.arg_scope([ops.conv2d, ops.fc, ops.batch_norm],
is_training=is_training):
# conv1
end_points['conv1'] = ops.conv2d(inputs,192,[5,5],scope='conv1')
end_points['conv1_1'] = ops.conv2d(end_points['conv1'],160,[1,1],scope='conv1_1')
end_points['conv1_2'] = ops.conv2d(end_points['conv1_1'],96,[1,1],scope='conv1_2')
end_points['pool1'] = ops.max_pool(end_points['conv1_2'],[3,3],stride=2,
padding='SAME',scope='pool1')
net = ops.dropout(end_points['pool1'],0.5)
# conv2
end_points['conv2'] = ops.conv2d(net,192,[5,5],scope='conv2')
end_points['conv2_1'] = ops.conv2d(end_points['conv2'],192,[1,1],scope='conv2_1')
end_points['conv2_2'] = ops.conv2d(end_points['conv2_1'],192,[1,1],scope='conv2_2')
end_points['pool2'] = ops.max_pool(end_points['conv2_2'],[3,3],stride=2,
padding='SAME',scope='pool2')
net = ops.dropout(end_points['pool2'],0.5)
# conv3
end_points['conv3'] = ops.conv2d(net,192,[3,3],scope='conv3')
end_points['conv3_1'] = ops.conv2d(end_points['conv3'],192,[1,1],scope='conv3_1')
end_points['conv3_2'] = ops.conv2d(end_points['conv3_1'],10,[1,1],scope='conv3_2')
net = ops.avg_pool(end_points['conv3_2'],[8,8],scope='avg_pool')
flatten = ops.flatten(net,scope='flatten')
#TODO take care this,using num_classes but 10..
end_points['logits'] = ops.fc(flatten,num_classes,activation=None,scope='fc')
return end_points['logits'],end_points
def testCreateFC(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
output = ops.fc(inputs, 32)
self.assertEqual(output.op.name, 'FC/Relu')
self.assertListEqual(output.get_shape().as_list(), [5, 32])
def nin_dssm(inputs,
num_classes,
num_of_exs,
is_training=True,
restore_logits=True,
scope=''):
# end_points will collect relevant activations for external use, for example
# summaries or losses.
end_points = {}
with tf.op_scope([inputs], scope, 'nin'):
with scopes.arg_scope([ops.conv2d, ops.fc, ops.batch_norm],
is_training=is_training):
# conv1
end_points['conv1'] = ops.conv2d(inputs,192,[5,5],scope='conv1')
end_points['conv1_1'] = ops.conv2d(end_points['conv1'],160,[1,1],scope='conv1_1')
end_points['conv1_2'] = ops.conv2d(end_points['conv1_1'],96,[1,1],scope='conv1_2')
end_points['pool1'] = ops.max_pool(end_points['conv1_2'],[3,3],stride=2,
padding='SAME',scope='pool1')
net = ops.dropout(end_points['pool1'],0.5)
# conv2 96*16*16
end_points['conv2'] = ops.conv2d(net,192,[5,5],scope='conv2')
end_points['conv2_1'] = ops.conv2d(end_points['conv2'],192,[1,1],scope='conv2_1')
end_points['conv2_2'] = ops.conv2d(end_points['conv2_1'],192,[1,1],scope='conv2_2')
end_points['pool2'] = ops.max_pool(end_points['conv2_2'],[3,3],stride=2,
padding='SAME',scope='pool2')
net = ops.dropout(end_points['pool2'],0.5)
# conv3 192*8*8
end_points['conv3'] = ops.conv2d(net,192,[3,3],scope='conv3')
# 192 * 8 * 8
end_points['conv3_1'] = ops.conv2d(end_points['conv3'],192,[1,1],scope='conv3_1')
# 192 * 8 * 8
#TODO using which layer feature?
#firstly,consider conv3_1, and then consider fusion conv3 & conv3_1
end_points['max_pool'] = ops.max_pool(end_points['conv3_1'],[8,8],scope='max_pool')
end_points['avg_pool'] = ops.avg_pool(end_points['conv3_1'],[8,8],scope='avg_pool')
end_points['hybrid_pool'] = 0.9*end_points['max_pool'] + 0.1*end_points['avg_pool']
end_points['feature'] = tf.nn.l2_normalize(tf.squeeze(end_points['hybrid_pool']),dim=1)
#OUTPUT (batch_size * num_negs_and_pos+1) * 192 ,eg. batch_size*3*192
imgs = tf.split(0,num_of_exs ,end_points['feature'])
anchors = imgs[0]
positives = imgs[1]
rst=[tf.reduce_sum(tf.mul(anchors,positives),1)]
for k in xrange(2,num_of_exs):
rst.append(tf.reduce_sum(tf.mul(anchors,imgs[k]),1))
#batch*(negs-1)
end_points['dssm'] = tf.concat(1,[tf.expand_dims(_,-1) for _ in rst])
end_points['conv3_2'] = ops.conv2d(end_points['conv3_1'],10,[1,1],scope='conv3_2')
net = ops.avg_pool(end_points['conv3_2'],[8,8],scope='avg_pool')
flatten = ops.flatten(net,scope='flatten')
#TODO take care this,using num_classes but 10..
end_points['logits'] = ops.fc(flatten,num_classes,activation=None,scope='fc')
return end_points['logits'],end_points['dssm'],end_points
def inception_v3(inputs,
dropout_keep_prob=0.8,
num_classes=1000,
is_training=True,
restore_logits=True,
scope=''):
"""Latest Inception from http://arxiv.org/abs/1512.00567.
"Rethinking the Inception Architecture for Computer Vision"
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens,
Zbigniew Wojna
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: Optional scope for name_scope.
Returns:
a list containing 'logits', 'aux_logits' Tensors.
"""
# end_points will collect relevant activations for external use, for example
# summaries or losses.
end_points = {}
with tf.name_scope(scope, 'inception_v3', [inputs]):
with scopes.arg_scope([ops.conv2d, ops.fc, ops.batch_norm, ops.dropout],
is_training=is_training):
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1, padding='VALID'):
# 299 x 299 x 3
end_points['conv0'] = ops.conv2d(inputs, 32, [3, 3], stride=2,
scope='conv0')
# 149 x 149 x 32
end_points['conv1'] = ops.conv2d(end_points['conv0'], 32, [3, 3],
scope='conv1')
# 147 x 147 x 32
end_points['conv2'] = ops.conv2d(end_points['conv1'], 64, [3, 3],
padding='SAME', scope='conv2')
# 147 x 147 x 64
end_points['pool1'] = ops.max_pool(end_points['conv2'], [3, 3],
stride=2, scope='pool1')
# 73 x 73 x 64
end_points['conv3'] = ops.conv2d(end_points['pool1'], 80, [1, 1],
scope='conv3')
# 73 x 73 x 80.
end_points['conv4'] = ops.conv2d(end_points['conv3'], 192, [3, 3],
scope='conv4')
# 71 x 71 x 192.
end_points['pool2'] = ops.max_pool(end_points['conv4'], [3, 3],
stride=2, scope='pool2')
# 35 x 35 x 192.
net = end_points['pool2']
# Inception blocks
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1, padding='SAME'):
# mixed: 35 x 35 x 256.
with tf.variable_scope('mixed_35x35x256a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 32, [1, 1])
net = tf.concat([branch1x1, branch5x5, branch3x3dbl, branch_pool], 3)
end_points['mixed_35x35x256a'] = net
# mixed_1: 35 x 35 x 288.
with tf.variable_scope('mixed_35x35x288a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 64, [1, 1])
net = tf.concat([branch1x1, branch5x5, branch3x3dbl, branch_pool], 3)
end_points['mixed_35x35x288a'] = net
# mixed_2: 35 x 35 x 288.
with tf.variable_scope('mixed_35x35x288b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 64, [1, 1])
net = tf.concat([branch1x1, branch5x5, branch3x3dbl, branch_pool], 3)
end_points['mixed_35x35x288b'] = net
# mixed_3: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768a'):
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 384, [3, 3], stride=2, padding='VALID')
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3],
stride=2, padding='VALID')
with tf.variable_scope('branch_pool'):
branch_pool = ops.max_pool(net, [3, 3], stride=2, padding='VALID')
net = tf.concat([branch3x3, branch3x3dbl, branch_pool], 3)
end_points['mixed_17x17x768a'] = net
# mixed4: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 128, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 128, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 128, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat([branch1x1, branch7x7, branch7x7dbl, branch_pool], 3)
end_points['mixed_17x17x768b'] = net
# mixed_5: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768c'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 160, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 160, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 160, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat([branch1x1, branch7x7, branch7x7dbl, branch_pool], 3)
end_points['mixed_17x17x768c'] = net
# mixed_6: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768d'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 160, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 160, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 160, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat([branch1x1, branch7x7, branch7x7dbl, branch_pool], 3)
end_points['mixed_17x17x768d'] = net
# mixed_7: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768e'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 192, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 192, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 192, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat([branch1x1, branch7x7, branch7x7dbl, branch_pool], 3)
end_points['mixed_17x17x768e'] = net
# Auxiliary Head logits
aux_logits = tf.identity(end_points['mixed_17x17x768e'])
with tf.variable_scope('aux_logits'):
aux_logits = ops.avg_pool(aux_logits, [5, 5], stride=3,
padding='VALID')
aux_logits = ops.conv2d(aux_logits, 128, [1, 1], scope='proj')
# Shape of feature map before the final layer.
shape = aux_logits.get_shape()
aux_logits = ops.conv2d(aux_logits, 768, shape[1:3], stddev=0.01,
padding='VALID')
aux_logits = ops.flatten(aux_logits)
aux_logits = ops.fc(aux_logits, num_classes, activation=None,
stddev=0.001, restore=restore_logits)
end_points['aux_logits'] = aux_logits
# mixed_8: 8 x 8 x 1280.
# Note that the scope below is not changed to not void previous
# checkpoints.
# (TODO) Fix the scope when appropriate.
with tf.variable_scope('mixed_17x17x1280a'):
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 192, [1, 1])
branch3x3 = ops.conv2d(branch3x3, 320, [3, 3], stride=2,
padding='VALID')
with tf.variable_scope('branch7x7x3'):
branch7x7x3 = ops.conv2d(net, 192, [1, 1])
branch7x7x3 = ops.conv2d(branch7x7x3, 192, [1, 7])
branch7x7x3 = ops.conv2d(branch7x7x3, 192, [7, 1])
branch7x7x3 = ops.conv2d(branch7x7x3, 192, [3, 3],
stride=2, padding='VALID')
with tf.variable_scope('branch_pool'):
branch_pool = ops.max_pool(net, [3, 3], stride=2, padding='VALID')
net = tf.concat([branch3x3, branch7x7x3, branch_pool], 3)
end_points['mixed_17x17x1280a'] = net
# mixed_9: 8 x 8 x 2048.
with tf.variable_scope('mixed_8x8x2048a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 320, [1, 1])
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 384, [1, 1])
branch3x3 = tf.concat([ops.conv2d(branch3x3, 384, [1, 3]),
ops.conv2d(branch3x3, 384, [3, 1])], 3)
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 448, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 384, [3, 3])
branch3x3dbl = tf.concat([ops.conv2d(branch3x3dbl, 384, [1, 3]),
ops.conv2d(branch3x3dbl, 384, [3, 1])], 3)
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat([branch1x1, branch3x3, branch3x3dbl, branch_pool], 3)
end_points['mixed_8x8x2048a'] = net
# mixed_10: 8 x 8 x 2048.
with tf.variable_scope('mixed_8x8x2048b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 320, [1, 1])
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 384, [1, 1])
branch3x3 = tf.concat([ops.conv2d(branch3x3, 384, [1, 3]),
ops.conv2d(branch3x3, 384, [3, 1])], 3)
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 448, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 384, [3, 3])
branch3x3dbl = tf.concat([ops.conv2d(branch3x3dbl, 384, [1, 3]),
ops.conv2d(branch3x3dbl, 384, [3, 1])], 3)
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat([branch1x1, branch3x3, branch3x3dbl, branch_pool], 3)
end_points['mixed_8x8x2048b'] = net
# Final pooling and prediction
with tf.variable_scope('logits'):
shape = net.get_shape()
net = ops.avg_pool(net, shape[1:3], padding='VALID', scope='pool')
# 1 x 1 x 2048
net = ops.dropout(net, dropout_keep_prob, scope='dropout')
net = ops.flatten(net, scope='flatten')
# 2048
logits = ops.fc(net, num_classes, activation=None, scope='logits',
restore=restore_logits)
# 1000
end_points['logits'] = logits
end_points['predictions'] = tf.nn.softmax(logits, name='predictions')
return logits, end_points
def testCreateFCWithoutActivation(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
output = ops.fc(inputs, 32, activation=None)
self.assertEquals(output.op.name, 'FC/xw_plus_b')
def cifar10_alexnet(inputs,
dropout_keep_prob=0.5,
num_classes=10,
is_training=True,
restore_logits=True,
weight_decay=0.004,
seed=1,
scope=''):
"""AlexNet on cifar10 from https://www.tensorflow.org/tutorials/deep_cnn
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: Optional scope for name_scope.
Returns:
a list containing 'logits', 'aux_logits' Tensors.
"""
# end_points will collect relevant activations for external use, for example
# summaries or losses.
print ("Warning: batch_norm_params is always None in cifar10_alexnet")
end_points = {}
with tf.name_scope(scope, 'cifar10_alexnet', [inputs]):
with scopes.arg_scope([ops.conv2d, ops.fc, ops.batch_norm, ops.dropout],
is_training=is_training):
with scopes.arg_scope([ops.conv2d, ops.fc],
bias=0.0, batch_norm_params=None, seed=seed):
with scopes.arg_scope([ops.conv2d], stride=1, padding='SAME'):
with scopes.arg_scope([ops.max_pool], stride=2, padding='SAME'):
# 32 x 32 x 3
end_points['conv1'] = ops.conv2d(inputs, 64, [5, 5], stride=1, stddev=0.05,
weight_decay=0.0, seed=seed+1, scope='conv1')
end_points['pool1'] = ops.max_pool(end_points['conv1'], [3, 3], scope='pool1')
end_points['lrn1'] = ops.lrn(end_points['pool1'], depth_radius=4, bias=1.0, alpha=0.001/9.0, beta=0.75, scope='lrn1')
end_points['conv2'] = ops.conv2d(end_points['lrn1'], 64, [5, 5], stride=1, stddev=0.05, bias=0.1,
weight_decay=0.0, seed=seed+2, scope='conv2')
end_points['lrn2'] = ops.lrn(end_points['conv2'], depth_radius=4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, scope='lrn2')
end_points['pool2'] = ops.max_pool(end_points['lrn2'], [3, 3], scope='pool2')
end_points['pool2'] = ops.flatten(end_points['pool2'], scope='flatten')
end_points['fc3'] = ops.fc(end_points['pool2'], 384, stddev=0.04, weight_decay=weight_decay, bias=0.1,
seed = seed +3, scope='fc3')
net = ops.fc(end_points['fc3'], 192, stddev=0.04, weight_decay=weight_decay, bias=0.1,
seed=seed + 4, scope='fc4')
# Final pooling and prediction
with tf.variable_scope('logits'):
logits = ops.fc(net, num_classes, activation=None, stddev=1/192.0, weight_decay=0.0,
bias=0.0, scope='logits', seed = seed +5, restore=restore_logits)
# 10
end_points['logits'] = logits
end_points['predictions'] = tf.nn.softmax(logits, name='predictions')
# There is no aux_logits for AlexNet
end_points['aux_logits'] = tf.constant(0)
return logits, end_points
def inception_v3(inputs,
dropout_keep_prob=0.8,
num_classes=1000,
is_training=True,
restore_logits=True,
scope=''):
# end_points will collect relevant activations for external use, for example
# summaries or losses.
end_points = {}
with tf.name_scope(scope, 'inception_v3', [inputs]):
with scopes.arg_scope(
[ops.conv2d, ops.fc, ops.batch_norm, ops.dropout],
is_training=is_training):
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1,
padding='VALID'):
# 299 x 299 x 3
end_points['conv0'] = ops.conv2d(inputs,
32, [3, 3],
stride=2,
scope='conv0')
# 149 x 149 x 32
end_points['conv1'] = ops.conv2d(end_points['conv0'],
32, [3, 3],
scope='conv1')
# 147 x 147 x 32
end_points['conv2'] = ops.conv2d(end_points['conv1'],
64, [3, 3],
padding='SAME',
scope='conv2')
# 147 x 147 x 64
end_points['pool1'] = ops.max_pool(end_points['conv2'], [3, 3],
stride=2,
scope='pool1')
# 73 x 73 x 64
end_points['conv3'] = ops.conv2d(end_points['pool1'],
80, [1, 1],
scope='conv3')
# 73 x 73 x 80.
end_points['conv4'] = ops.conv2d(end_points['conv3'],
192, [3, 3],
scope='conv4')
# 71 x 71 x 192.
end_points['pool2'] = ops.max_pool(end_points['conv4'], [3, 3],
stride=2,
scope='pool2')
# 35 x 35 x 192.
net = end_points['pool2']
# Inception blocks
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1,
padding='SAME'):
# mixed: 35 x 35 x 256.
with tf.variable_scope('mixed_35x35x256a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 32, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch5x5, branch3x3dbl,
branch_pool
])
end_points['mixed_35x35x256a'] = net
# mixed_1: 35 x 35 x 288.
with tf.variable_scope('mixed_35x35x288a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 64, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch5x5, branch3x3dbl,
branch_pool
])
end_points['mixed_35x35x288a'] = net
# mixed_2: 35 x 35 x 288.
with tf.variable_scope('mixed_35x35x288b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 64, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch5x5, branch3x3dbl,
branch_pool
])
end_points['mixed_35x35x288b'] = net
# mixed_3: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768a'):
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net,
384, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl,
96, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch_pool'):
branch_pool = ops.max_pool(net, [3, 3],
stride=2,
padding='VALID')
net = tf.concat(
axis=3, values=[branch3x3, branch3x3dbl, branch_pool])
end_points['mixed_17x17x768a'] = net
# mixed4: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 128, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 128, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 128, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch7x7, branch7x7dbl,
branch_pool
])
end_points['mixed_17x17x768b'] = net
# mixed_5: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768c'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 160, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 160, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 160, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch7x7, branch7x7dbl,
branch_pool
])
end_points['mixed_17x17x768c'] = net
# mixed_6: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768d'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 160, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 160, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 160, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch7x7, branch7x7dbl,
branch_pool
])
end_points['mixed_17x17x768d'] = net
# mixed_7: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768e'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 192, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 192, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 192, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch7x7, branch7x7dbl,
branch_pool
])
end_points['mixed_17x17x768e'] = net
# Auxiliary Head logits
aux_logits = tf.identity(end_points['mixed_17x17x768e'])
with tf.variable_scope('aux_logits'):
aux_logits = ops.avg_pool(aux_logits, [5, 5],
stride=3,
padding='VALID')
aux_logits = ops.conv2d(aux_logits,
128, [1, 1],
scope='proj')
# Shape of feature map before the final layer.
shape = aux_logits.get_shape()
aux_logits = ops.conv2d(aux_logits,
768,
shape[1:3],
stddev=0.01,
padding='VALID')
aux_logits = ops.flatten(aux_logits)
aux_logits = ops.fc(aux_logits,
num_classes,
activation=None,
stddev=0.001,
restore=restore_logits)
end_points['aux_logits'] = aux_logits
# mixed_8: 8 x 8 x 1280.
# Note that the scope below is not changed to not void previous
# checkpoints.
# (TODO) Fix the scope when appropriate.
with tf.variable_scope('mixed_17x17x1280a'):
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 192, [1, 1])
branch3x3 = ops.conv2d(branch3x3,
320, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch7x7x3'):
branch7x7x3 = ops.conv2d(net, 192, [1, 1])
branch7x7x3 = ops.conv2d(branch7x7x3, 192, [1, 7])
branch7x7x3 = ops.conv2d(branch7x7x3, 192, [7, 1])
branch7x7x3 = ops.conv2d(branch7x7x3,
192, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch_pool'):
branch_pool = ops.max_pool(net, [3, 3],
stride=2,
padding='VALID')
net = tf.concat(
axis=3, values=[branch3x3, branch7x7x3, branch_pool])
end_points['mixed_17x17x1280a'] = net
# mixed_9: 8 x 8 x 2048.
with tf.variable_scope('mixed_8x8x2048a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 320, [1, 1])
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 384, [1, 1])
branch3x3 = tf.concat(
axis=3,
values=[
ops.conv2d(branch3x3, 384, [1, 3]),
ops.conv2d(branch3x3, 384, [3, 1])
])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 448, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 384, [3, 3])
branch3x3dbl = tf.concat(
axis=3,
values=[
ops.conv2d(branch3x3dbl, 384, [1, 3]),
ops.conv2d(branch3x3dbl, 384, [3, 1])
])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch3x3, branch3x3dbl,
branch_pool
])
end_points['mixed_8x8x2048a'] = net
# mixed_10: 8 x 8 x 2048.
with tf.variable_scope('mixed_8x8x2048b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 320, [1, 1])
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 384, [1, 1])
branch3x3 = tf.concat(
axis=3,
values=[
ops.conv2d(branch3x3, 384, [1, 3]),
ops.conv2d(branch3x3, 384, [3, 1])
])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 448, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 384, [3, 3])
branch3x3dbl = tf.concat(
axis=3,
values=[
ops.conv2d(branch3x3dbl, 384, [1, 3]),
ops.conv2d(branch3x3dbl, 384, [3, 1])
])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch3x3, branch3x3dbl,
branch_pool
])
end_points['mixed_8x8x2048b'] = net
# Final pooling and prediction
with tf.variable_scope('logits'):
shape = net.get_shape()
net = ops.avg_pool(net,
shape[1:3],
padding='VALID',
scope='pool')
# 1 x 1 x 2048
net = ops.dropout(net, dropout_keep_prob, scope='dropout')
net = ops.flatten(net, scope='flatten')
# 2048
logits = ops.fc(net,
num_classes,
activation=None,
scope='logits',
restore=restore_logits)
# 1000
end_points['logits'] = logits
end_points['predictions'] = tf.nn.softmax(
logits, name='predictions')
return logits, end_points
def testCreateFCWithScope(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
output = ops.fc(inputs, 32, scope='fc1')
self.assertEquals(output.op.name, 'fc1/Relu')
def testCreateFCWithoutActivation(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
output = ops.fc(inputs, 32, activation=None)
self.assertEqual(output.op.name, 'FC/xw_plus_b')
def inception_v3(inputs,
dropout_keep_prob=0.8,
num_classes=1001,
is_training=True,
restore_logits=True,
scope=''):
"""Latest Inception from http://arxiv.org/abs/1512.00567.
"Rethinking the Inception Architecture for Computer Vision"
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens,
Zbigniew Wojna
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: Optional scope for op_scope.
Returns:
a list containing 'logits', 'aux_logits' Tensors.
"""
# end_points will collect relevant activations for external use, for example
# summaries or losses.
end_points = {}
with tf.op_scope([inputs], scope, 'baxNet'):
with scopes.arg_scope([ops.conv2d, ops.fc, ops.batch_norm, ops.dropout],
is_training=is_training):
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1, padding='VALID'):
# 256 x 256 x 3
end_points['conv0'] = ops.conv2d(inputs, 8, [5, 5], stride=1,
scope='conv0', padding='SAME')
end_points['batch_norm1'] = ops.batch_norm(end_points['conv0'], scope='batch_norm1')
# 256 x 256 x 32
end_points['conv1'] = ops.conv2d(end_points['batch_norm1'], 16, [3, 3],
scope='conv1', padding='SAME')
end_points['batch_norm2'] = ops.batch_norm(end_points['conv1'], scope='batch_norm2')
# 128 x 128 x 64
end_points['conv2'] = ops.conv2d(end_points['batch_norm2'], 16, [3, 3],
scope='conv2', padding='SAME')
end_points['batch_norm3'] = ops.batch_norm(end_points['conv2'], scope='batch_norm3')
in_net = end_points['batch_norm3']
print('IN_NET SHAPE')
print(in_net.get_shape())
curr_filters = 16
base_layer_num = [32,16,8,4]
for i in xrange(1,5):
for j in xrange(1,base_layer_num[i-1] + i):
with tf.variable_scope('res%d_%d' % (i,j)):
if (j < (base_layer_num[i-1] + i - 1)):
curr_padding = 'SAME'
curr_stride = 1
else:
curr_filters = 2*curr_filters
curr_padding = 'SAME'
curr_stride = 2
conv1_1 = ops.conv2d(in_net, curr_filters, [3, 3], padding=curr_padding, stride=curr_stride, scope='conv1_1')
batch_norm1_1 = ops.batch_norm(conv1_1, scope='batch_norm1_1')
conv1_2 = ops.conv2d(batch_norm1_1, curr_filters, [3, 3], padding='SAME', scope='conv1_2')
if (j < (base_layer_num[i-1] + i - 1)):
combined = in_net + conv1_2
else:
combined = ops.conv2d(in_net, curr_filters, [1, 1], padding='SAME', stride=2, scope='combined')
combined = combined + conv1_2
print('DOWN SAMPLE')
print(in_net.get_shape())
print(combined.get_shape())
batch_norm1_2 = ops.batch_norm(combined, scope='batch_norm1_2')
in_net = batch_norm1_2
end_points['res%d_%d' %(i,j)] = in_net
# for i in xrange(1,int(np.log2(in_net.get_shape()[1])) + 1):
# print('SHAPPEEEE')
print(in_net.get_shape())
for i in xrange(1,3):
with tf.variable_scope('res_final%d' % i):
conv1_1 = ops.conv2d(in_net, curr_filters, [3, 3], padding='SAME', stride=2, scope='conv1_1')
batch_norm1_1 = ops.batch_norm(conv1_1, scope='batch_norm1_1')
conv1_2 = ops.conv2d(batch_norm1_1, curr_filters, [3, 3], padding='SAME', scope='conv1_2')
combined = ops.conv2d(in_net, curr_filters, [1, 1], padding='SAME', stride=2, scope='combined')
combined = combined + conv1_2
batch_norm1_2 = ops.batch_norm(combined, scope='batch_norm1_2')
in_net = batch_norm1_2
end_points['res_final%d' % i] = in_net
with tf.variable_scope('logits'):
shape = in_net.get_shape()
print('FINAL SHAPE')
print(shape)
if (shape[1] > 1):
in_net = ops.avg_pool(in_net, shape[1:3], padding='VALID', scope='avg_pool')
in_net = ops.flatten(in_net, scope='flatten')
logits = ops.fc(in_net, num_classes, activation=None, scope='logits',
restore=restore_logits)
end_points['logits'] = logits
end_points['predictions'] = tf.nn.softmax(logits, name='predictions')
return logits, end_points
def vgg(inputs,
num_classes=1000,
is_training=True,
restore_logits=True,
scope=''):
# end_points will collect relevant activations for external use, for example
# summaries or losses.
end_points = {}
with tf.op_scope([inputs], scope, 'vgg'):
with scopes.arg_scope([ops.conv2d, ops.fc, ops.batch_norm],
is_training=is_training):
# conv1
end_points['conv1'] = ops.repeat_op(2,
inputs,
ops.conv2d,
64, [3, 3],
scope='conv1')
end_points['pool1'] = ops.max_pool(end_points['conv1'], [2, 2],
scope='pool1')
# conv2
end_points['conv2'] = ops.repeat_op(2,
end_points['pool1'],
ops.conv2d,
128, [3, 3],
scope='conv2')
end_points['pool2'] = ops.max_pool(end_points['conv2'], [2, 2],
scope='pool2')
# conv3
end_points['conv3'] = ops.repeat_op(2,
end_points['pool2'],
ops.conv2d,
256, [3, 3],
scope='conv3')
end_points['pool3'] = ops.max_pool(end_points['conv3'], [2, 2],
scope='pool3')
# conv4
end_points['conv4'] = ops.repeat_op(2,
end_points['pool3'],
ops.conv2d,
512, [3, 3],
scope='conv4')
end_points['pool4'] = ops.max_pool(end_points['conv4'], [2, 2],
scope='pool4')
# conv5
end_points['conv5'] = ops.repeat_op(2,
end_points['pool4'],
ops.conv2d,
512, [3, 3],
scope='conv5')
end_points['pool5'] = ops.max_pool(end_points['conv5'], [2, 2],
scope='pool5')
end_points['flatten5'] = ops.flatten(end_points['pool5'],
scope='flatten5')
end_points['fc6'] = ops.fc(end_points['flatten5'],
4096,
scope='fc6')
end_points['dropout6'] = ops.dropout(end_points['fc6'],
0.5,
scope='dropout6')
end_points['fc7'] = ops.fc(end_points['dropout6'],
4096,
scope='fc7')
end_points['dropout7'] = ops.dropout(end_points['fc7'],
0.5,
scope='dropout7')
logits = ops.fc(end_points['fc7'],
num_classes,
activation=None,
scope='fc8')
return logits, end_points
def inception_v3(inputs,
dropout_keep_prob=0.8,
num_classes=1000,
is_training=True,
restore_logits=True,
scope=''):
"""Latest Inception from http://arxiv.org/abs/1512.00567.
"Rethinking the Inception Architecture for Computer Vision"
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens,
Zbigniew Wojna
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: Optional scope for name_scope.
Returns:
a list containing 'logits', 'aux_logits' Tensors.
"""
# end_points will collect relevant activations for external use, for example
# summaries or losses.
end_points = {}
with tf.name_scope(scope, 'inception_v3', [inputs]):
with scopes.arg_scope(
[ops.conv2d, ops.fc, ops.batch_norm, ops.dropout],
is_training=is_training):
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1,
padding='VALID'):
# 299 x 299 x 3
end_points['conv0'] = ops.conv2d(inputs,
32, [3, 3],
stride=2,
scope='conv0')
# 149 x 149 x 32
end_points['conv1'] = ops.conv2d(end_points['conv0'],
32, [3, 3],
scope='conv1')
# 147 x 147 x 32
end_points['conv2'] = ops.conv2d(end_points['conv1'],
64, [3, 3],
padding='SAME',
scope='conv2')
# 147 x 147 x 64
end_points['pool1'] = ops.max_pool(end_points['conv2'], [3, 3],
stride=2,
scope='pool1')
# 73 x 73 x 64
end_points['conv3'] = ops.conv2d(end_points['pool1'],
80, [1, 1],
scope='conv3')
# 73 x 73 x 80.
end_points['conv4'] = ops.conv2d(end_points['conv3'],
192, [3, 3],
scope='conv4')
# 71 x 71 x 192.
end_points['pool2'] = ops.max_pool(end_points['conv4'], [3, 3],
stride=2,
scope='pool2')
# 35 x 35 x 192.
net = end_points['pool2']
# Inception blocks
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1,
padding='SAME'):
# mixed: 35 x 35 x 256.
with tf.variable_scope('mixed_35x35x256a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 32, [1, 1])
net = tf.concat(
[branch1x1, branch5x5, branch3x3dbl, branch_pool], 3)
end_points['mixed_35x35x256a'] = net
# mixed_1: 35 x 35 x 288.
with tf.variable_scope('mixed_35x35x288a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 64, [1, 1])
net = tf.concat(
[branch1x1, branch5x5, branch3x3dbl, branch_pool], 3)
end_points['mixed_35x35x288a'] = net
# mixed_2: 35 x 35 x 288.
with tf.variable_scope('mixed_35x35x288b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 64, [1, 1])
net = tf.concat(
[branch1x1, branch5x5, branch3x3dbl, branch_pool], 3)
end_points['mixed_35x35x288b'] = net
# mixed_3: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768a'):
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net,
384, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl,
96, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch_pool'):
branch_pool = ops.max_pool(net, [3, 3],
stride=2,
padding='VALID')
net = tf.concat([branch3x3, branch3x3dbl, branch_pool], 3)
end_points['mixed_17x17x768a'] = net
# mixed4: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 128, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 128, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 128, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(
[branch1x1, branch7x7, branch7x7dbl, branch_pool], 3)
end_points['mixed_17x17x768b'] = net
# mixed_5: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768c'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 160, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 160, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 160, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(
[branch1x1, branch7x7, branch7x7dbl, branch_pool], 3)
end_points['mixed_17x17x768c'] = net
# mixed_6: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768d'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 160, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 160, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 160, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(
[branch1x1, branch7x7, branch7x7dbl, branch_pool], 3)
end_points['mixed_17x17x768d'] = net
# mixed_7: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768e'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 192, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 192, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 192, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(
[branch1x1, branch7x7, branch7x7dbl, branch_pool], 3)
end_points['mixed_17x17x768e'] = net
# Auxiliary Head logits
aux_logits = tf.identity(end_points['mixed_17x17x768e'])
with tf.variable_scope('aux_logits'):
aux_logits = ops.avg_pool(aux_logits, [5, 5],
stride=3,
padding='VALID')
aux_logits = ops.conv2d(aux_logits,
128, [1, 1],
scope='proj')
# Shape of feature map before the final layer.
shape = aux_logits.get_shape()
aux_logits = ops.conv2d(aux_logits,
768,
shape[1:3],
stddev=0.01,
padding='VALID')
aux_logits = ops.flatten(aux_logits)
aux_logits = ops.fc(aux_logits,
num_classes,
activation=None,
stddev=0.001,
restore=restore_logits)
end_points['aux_logits'] = aux_logits
# mixed_8: 8 x 8 x 1280.
# Note that the scope below is not changed to not void previous
# checkpoints.
# (TODO) Fix the scope when appropriate.
with tf.variable_scope('mixed_17x17x1280a'):
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 192, [1, 1])
branch3x3 = ops.conv2d(branch3x3,
320, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch7x7x3'):
branch7x7x3 = ops.conv2d(net, 192, [1, 1])
branch7x7x3 = ops.conv2d(branch7x7x3, 192, [1, 7])
branch7x7x3 = ops.conv2d(branch7x7x3, 192, [7, 1])
branch7x7x3 = ops.conv2d(branch7x7x3,
192, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch_pool'):
branch_pool = ops.max_pool(net, [3, 3],
stride=2,
padding='VALID')
net = tf.concat([branch3x3, branch7x7x3, branch_pool], 3)
end_points['mixed_17x17x1280a'] = net
# mixed_9: 8 x 8 x 2048.
with tf.variable_scope('mixed_8x8x2048a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 320, [1, 1])
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 384, [1, 1])
branch3x3 = tf.concat([
ops.conv2d(branch3x3, 384, [1, 3]),
ops.conv2d(branch3x3, 384, [3, 1])
], 3)
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 448, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 384, [3, 3])
branch3x3dbl = tf.concat([
ops.conv2d(branch3x3dbl, 384, [1, 3]),
ops.conv2d(branch3x3dbl, 384, [3, 1])
], 3)
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(
[branch1x1, branch3x3, branch3x3dbl, branch_pool], 3)
end_points['mixed_8x8x2048a'] = net
# mixed_10: 8 x 8 x 2048.
with tf.variable_scope('mixed_8x8x2048b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 320, [1, 1])
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 384, [1, 1])
branch3x3 = tf.concat([
ops.conv2d(branch3x3, 384, [1, 3]),
ops.conv2d(branch3x3, 384, [3, 1])
], 3)
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 448, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 384, [3, 3])
branch3x3dbl = tf.concat([
ops.conv2d(branch3x3dbl, 384, [1, 3]),
ops.conv2d(branch3x3dbl, 384, [3, 1])
], 3)
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(
[branch1x1, branch3x3, branch3x3dbl, branch_pool], 3)
end_points['mixed_8x8x2048b'] = net
# Final pooling and prediction
with tf.variable_scope('logits'):
shape = net.get_shape()
net = ops.avg_pool(net,
shape[1:3],
padding='VALID',
scope='pool')
# 1 x 1 x 2048
net = ops.dropout(net, dropout_keep_prob, scope='dropout')
net = ops.flatten(net, scope='flatten')
# 2048
logits_2048 = net
logits = ops.fc(net,
num_classes,
activation=None,
scope='logits',
restore=restore_logits)
# 1000
end_points['logits'] = logits
end_points['predictions'] = tf.nn.softmax(
logits, name='predictions')
return logits, end_points, logits_2048
def vgg_16(inputs,
dropout_keep_prob=0.8,
num_classes=10,
is_training=True,
restore_logits=True,
scope=''):
"""Latest Inception from http://arxiv.org/abs/1512.00567.
"Rethinking the Inception Architecture for Computer Vision"
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens,
Zbigniew Wojna
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: Optional scope for op_scope.
Returns:
a list containing 'logits', 'aux_logits' Tensors.
"""
# end_points will collect relevant activations for external use, for example
# summaries or losses.
dropout_keep_prob = 0.4 if training else 1.0
end_points = {}
with tf.op_scope([inputs], scope, 'vgg_16'):
with scopes.arg_scope([ops.conv2d, ops.fc, ops.batch_norm, ops.dropout],
is_training=is_training):
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1, padding='SAME'):
# assume input_op shape is 224x224x3
# block 1 -- outputs 112x112x64
end_points['conv1_1'] = ops.conv2d(inputs, 64, [3, 3], stride=1,
scope='conv1_1')
end_points['conv1_2'] = ops.conv2d(end_points['conv1_1'], 64, [3, 3],
scope='conv1_2')
end_points['pool1'] = ops.max_pool(end_points['conv1_2'], [2, 2],
stride=2, scope='pool1')
# block 2 -- outputs 56x56x128
end_points['conv2_1'] = ops.conv2d(end_points['pool1'], 128, [3, 3],
scope='conv2_1')
end_points['conv2_2'] = ops.conv2d(end_points['conv2_1'], 128, [3, 3],
scope='conv2_2')
end_points['pool2'] = ops.max_pool(end_points['conv2_2'], [2, 2],
stride=2, scope='pool2')
# block 3 -- outputs 28x28x256
end_points['conv3_1'] = ops.conv2d(end_points['pool2'], 256, [3, 3],
scope='conv3_1')
end_points['conv3_2'] = ops.conv2d(end_points['conv3_1'], 256, [3, 3],
scope='conv3_2')
end_points['pool3'] = ops.max_pool(end_points['conv3_2'], [2, 2],
stride=2, scope='pool3')
# block 4 -- outputs 14x14x512
end_points['conv4_1'] = ops.conv2d(end_points['pool3'], 512, [3, 3],
scope='conv4_1')
end_points['conv4_2'] = ops.conv2d(end_points['conv4_1'], 512, [3, 3],
scope='conv4_2')
end_points['pool4'] = ops.max_pool(end_points['conv4_2'], [2, 2],
stride=2, scope='pool4')
# block 5 -- outputs 7x7x512
end_points['conv5_1'] = ops.conv2d(end_points['pool4'], 512, [3, 3],
scope='conv5_1')
end_points['conv5_2'] = ops.conv2d(end_points['conv5_1'], 512, [3, 3],
scope='conv5_2')
end_points['pool5'] = ops.max_pool(end_points['conv5_2'], [2, 2],
stride=2, scope='pool5')
net = end_points['pool5']
# Final pooling and prediction
with tf.variable_scope('logits'):
# flatten
net = ops.flatten(net, scope='flatten')
# fully connected
end_points['fc6'] = ops.fc(net, 1000, activation=None, scope='fc6',
restore=restore_logits)
end_points['fc6_drop'] = ops.dropout(end_points['fc6'], dropout_keep_prob, scope='fc6_drop')
end_points['fc7'] = ops.fc(end_points['fc6_drop'], 50, activation=None, scope='fc7',
restore=restore_logits)
end_points['fc7_drop'] = ops.dropout(end_points['fc7'], dropout_keep_prob, scope='fc7_drop')
end_points['fc8'] = ops.fc(end_points['fc7_drop'], num_classes, activation=None, scope='fc8',
restore=restore_logits)
end_points['fc8_drop'] = ops.dropout(end_points['fc8'], dropout_keep_prob, scope='fc8_drop')
logits = end_points['fc8_drop']
end_points['logits'] = logits
end_points['predictions'] = tf.nn.softmax(logits, name='predictions')
return logits, end_points
def alexnet(inputs,
dropout_keep_prob=0.5,
num_classes=1000,
is_training=True,
restore_logits=True,
seed=1,
weight_decay=0.0005,
scope=''):
"""AlexNet from https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: Optional scope for name_scope.
Returns:
a list containing 'logits', 'aux_logits' Tensors.
"""
# end_points will collect relevant activations for external use, for example
# summaries or losses.
#print ("INFO: batch norm in alexnet is disabled")
end_points = {}
with tf.name_scope(scope, 'alexnet', [inputs]):
with scopes.arg_scope([ops.conv2d, ops.fc, ops.batch_norm, ops.dropout],
is_training=is_training):
with scopes.arg_scope([ops.conv2d, ops.fc],
weight_decay=weight_decay, stddev=0.01, bias=0.1,
#batch_norm_params=None,
weights_initializer=tf.truncated_normal_initializer):
with scopes.arg_scope([ops.conv2d],
stride=1, padding='SAME'):
with scopes.arg_scope([ops.max_pool],
stride=2, padding='VALID'):
# 224 x 224 x 3
end_points['conv1_1'] = ops.conv2d(inputs, 48, [11, 11], stride=4, bias=0.0, seed = seed +1, scope='conv1_1')
end_points['conv1_2'] = ops.conv2d(inputs, 48, [11, 11], stride=4, bias=0.0, seed = seed +2, scope='conv1_2')
end_points['lrn1_1'] = ops.lrn(end_points['conv1_1'], scope='lrn1_1')
end_points['lrn1_2'] = ops.lrn(end_points['conv1_2'], scope='lrn1_2')
end_points['pool1_1'] = ops.max_pool(end_points['lrn1_1'], [3, 3], scope='pool1_1')
end_points['pool1_2'] = ops.max_pool(end_points['lrn1_2'], [3, 3], scope='pool1_2')
# 27 x 27 x 48 x 2
end_points['conv2_1'] = ops.conv2d(end_points['pool1_1'], 128, [5, 5], seed = seed +3, scope='conv2_1')
end_points['conv2_2'] = ops.conv2d(end_points['pool1_2'], 128, [5, 5], seed = seed +4, scope='conv2_2')
end_points['lrn2_1'] = ops.lrn(end_points['conv2_1'], scope='lrn2_1')
end_points['lrn2_2'] = ops.lrn(end_points['conv2_2'], scope='lrn2_2')
end_points['pool2_1'] = ops.max_pool(end_points['lrn2_1'], [3, 3], scope='pool2_1')
end_points['pool2_2'] = ops.max_pool(end_points['lrn2_2'], [3, 3], scope='pool2_2')
end_points['pool2'] = tf.concat([end_points['pool2_1'],end_points['pool2_2']],3)
# 13 x 13 x 256
end_points['conv3_1'] = ops.conv2d(end_points['pool2'], 192, [3, 3], bias=0.0, seed = seed +5, scope='conv3_1')
end_points['conv3_2'] = ops.conv2d(end_points['pool2'], 192, [3, 3], bias=0.0, seed = seed +6, scope='conv3_2')
# 13 x 13 x 192 x 2
end_points['conv4_1'] = ops.conv2d(end_points['conv3_1'], 192, [3, 3], seed = seed +7, scope='conv4_1')
end_points['conv4_2'] = ops.conv2d(end_points['conv3_2'], 192, [3, 3], seed = seed +8, scope='conv4_2')
# 13 x 13 x 192 x 2
end_points['conv5_1'] = ops.conv2d(end_points['conv4_1'], 128, [3, 3], seed = seed +9, scope='conv5_1')
end_points['conv5_2'] = ops.conv2d(end_points['conv4_2'], 128, [3, 3], seed = seed +10, scope='conv5_2')
end_points['pool5_1'] = ops.max_pool(end_points['conv5_1'], [3, 3], scope='pool5_1')
end_points['pool5_2'] = ops.max_pool(end_points['conv5_2'], [3, 3], scope='pool5_2')
end_points['pool5'] = tf.concat([end_points['pool5_1'], end_points['pool5_2']], 3)
end_points['pool5'] = ops.flatten(end_points['pool5'], scope='flatten')
end_points['fc6'] = ops.fc(end_points['pool5'], 4096, stddev=0.005, seed = seed +11, scope='fc6')
end_points['dropout6'] = ops.dropout(end_points['fc6'], dropout_keep_prob, scope='dropout6')
end_points['fc7'] = ops.fc(end_points['dropout6'], 4096, stddev=0.005, seed = seed +12, scope='fc7')
net = ops.dropout(end_points['fc7'], dropout_keep_prob, scope='dropout7')
# Final pooling and prediction
with tf.variable_scope('logits'):
# 4096
logits = ops.fc(net, num_classes, activation=None, bias=0.0, seed = seed +13, scope='logits',
restore=restore_logits)
# 1000
end_points['logits'] = logits
end_points['predictions'] = tf.nn.softmax(logits, name='predictions')
# There is no aux_logits for AlexNet
end_points['aux_logits'] = tf.constant(0)
return logits, end_points
def inception_v3(inputs,
num_classes=2,
is_training=True,
dropout_keep_prob=0.8,
min_depth=16,
depth_multiplier=1.0,
spatial_squeeze=True,
reuse=None,
create_aux_logits=True,
scope='InceptionV3',
global_pool=False):
"""Inception model from http://arxiv.org/abs/1512.00567.
"Rethinking the Inception Architecture for Computer Vision"
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens,
Zbigniew Wojna.
With the default arguments this method constructs the exact model defined in
the paper. However, one can experiment with variations of the inception_v3
network by changing arguments dropout_keep_prob, min_depth and
depth_multiplier.
The default image size used to train this network is 299x299.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes. If 0 or None, the logits layer
is omitted and the input features to the logits layer (before dropout)
are returned instead.
is_training: whether is training or not.
dropout_keep_prob: the percentage of activation values that are retained.
min_depth: Minimum depth value (number of channels) for all convolution ops.
Enforced when depth_multiplier < 1, and not an active constraint when
depth_multiplier >= 1.
depth_multiplier: Float multiplier for the depth (number of channels)
for all convolution ops. The value must be greater than zero. Typical
usage will be to set this value in (0, 1) to reduce the number of
parameters or computation cost of the model.
prediction_fn: a function to get predictions out of logits.
spatial_squeeze: if True, logits is of shape [B, C], if false logits is of
shape [B, 1, 1, C], where B is batch_size and C is number of classes.
reuse: whether or not the network and its variables should be reused. To be
able to reuse 'scope' must be given.
create_aux_logits: Whether to create the auxiliary logits.
scope: Optional variable_scope.
global_pool: Optional boolean flag to control the avgpooling before the
logits layer. If false or unset, pooling is done with a fixed window
that reduces default-sized inputs to 1x1, while larger inputs lead to
larger outputs. If true, any input size is pooled down to 1x1.
Returns:
net: a Tensor with the logits (pre-softmax activations) if num_classes
is a non-zero integer, or the non-dropped-out input to the logits layer
if num_classes is 0 or None.
end_points: a dictionary from components of the network to the corresponding
activation.
Raises:
ValueError: if 'depth_multiplier' is less than or equal to zero.
"""
if depth_multiplier <= 0:
raise ValueError('depth_multiplier is not greater than zero.')
depth = lambda d: max(int(d * depth_multiplier), min_depth)
#with tf.variable_scope(scope, 'InceptionV3', [inputs], reuse=reuse) as scope:
with tf.name_scope(scope, 'inception_v3', [inputs]) as scope:
with scopes.arg_scope([ops.batch_norm, ops.dropout],
is_training=is_training):
net, end_points = inception_v3_base(
inputs,
scope=scope,
min_depth=min_depth,
depth_multiplier=depth_multiplier,
is_training=is_training)
# Auxiliary Head logits
if create_aux_logits and num_classes:
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1,
padding='SAME'):
aux_logits = tf.identity(end_points['Mixed_6e'])
with tf.variable_scope('aux_logits'):
aux_logits = ops.avg_pool(aux_logits, [5, 5],
stride=3,
padding='VALID')
aux_logits = ops.conv2d(aux_logits,
128, [1, 1],
scope='proj')
# Shape of feature map before the final layer.
shape = aux_logits.get_shape()
aux_logits = ops.conv2d(aux_logits,
768,
shape[1:3],
stddev=0.01,
padding='VALID')
aux_logits = ops.flatten(aux_logits)
aux_logits = ops.fc(aux_logits,
num_classes,
activation=None,
stddev=0.001,
restore=reuse)
end_points['aux_logits'] = aux_logits
with tf.variable_scope('logits'):
shape = net.get_shape()
net = ops.avg_pool(net,
shape[1:3],
padding='VALID',
scope='pool')
# 1 x 1 x 2048
net = ops.dropout(net, dropout_keep_prob, scope='dropout')
net = ops.flatten(net, scope='flatten')
# 2048
logits = ops.fc(net,
num_classes,
activation=None,
scope='logits',
restore=reuse)
# 1000
end_points['logits'] = logits
end_points['predictions'] = tf.nn.softmax(logits,
name='predictions')
return net, logits, end_points
def alexnet(inputs,
dropout_keep_prob=0.8,
num_classes=1000,
is_training=True,
restore_logits=True,
scope=''):
"""Latest Inception from http://arxiv.org/abs/1512.00567.
"Rethinking the Inception Architecture for Computer Vision"
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens,
Zbigniew Wojna
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: Optional scope for op_scope.
Returns:
a list containing 'logits', 'aux_logits' Tensors.
"""
# end_points will collect relevant activations for external use, for example
# summaries or losses.
end_points = {}
with tf.op_scope([inputs], scope, 'alexnet'):
with scopes.arg_scope([ops.conv2d, ops.fc, ops.batch_norm, ops.dropout],
is_training=is_training):
# conv and pool will do padding
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
padding='SAME'):
# define the initial distribution of filter weight
with scopes.arg_scope([ops.conv2d], stddev=0.01):
end_points['conv1'] = ops.conv2d(inputs, 96, [11, 11], stride=4,
scope='conv1')
end_points['pool1'] = ops.max_pool(end_points['conv1'], [3, 3],
stride=2, scope='pool1')
end_points['conv2'] = ops.conv2d(end_points['pool1'], 256, [5, 5],
bias=1.0, scope='conv2')
end_points['pool2'] = ops.max_pool(end_points['conv2'], [3, 3],
stride=2, scope='pool2')
end_points['conv3'] = ops.conv2d(end_points['pool2'], 384, [3, 3],
scope='conv3')
end_points['conv4'] = ops.conv2d(end_points['conv3'], 384, [3, 3],
bias=1.0, scope='conv4')
end_points['conv5'] = ops.conv2d(end_points['conv4'], 256, [3, 3],
bias=1.0, scope='conv5')
end_points['pool5'] = ops.max_pool(end_points['conv5'], [3, 3],
stride=2, scope='pool5')
# reshape the 4d tensor into 2d
end_points['flatten'] = ops.flatten(end_points['pool5'], scope='flatten')
# define the initial distribution of fc weight
with scopes.arg_scope([ops.fc], stddev=0.005, bias=1.0):
# define the dropout ratio
with scopes.arg_scope([ops.dropout], keep_prob=dropout_keep_prob):
end_points['fc6'] = ops.fc(end_points['flatten'], 4096, scope='fc6')
end_points['drop6'] = ops.dropout(end_points['fc6'], scope='drop6')
end_points['fc7'] = ops.fc(end_points['drop6'], 4096, scope='fc7')
end_points['drop7'] = ops.dropout(end_points['fc7'], scope='drop7')
end_points['fc8'] = ops.fc(end_points['drop7'], num_classes,
activation=None,
scope='fc8', restore=restore_logits)
return end_points['fc8'], end_points
def resnet34(inputs,
num_classes=1000,
is_training=True,
restore_logits=True,
scope=''):
end_points = {}
with tf.op_scope([inputs], scope, 'resnet'):
with scopes.arg_scope([ops.conv2d, ops.batch_norm],
is_training=is_training):
# 224 x 224 x 3
end_points['conv1'] = ops.conv2d(inputs,
64, [7, 7],
stride=2,
scope='conv1')
end_points['pool1'] = ops.max_pool(end_points['conv1'], [3, 3],
stride=2,
padding='SAME',
scope='pool1')
# 56 * 56
#TODO (using loop)
end_points['conv2_1'] = block34(end_points['pool1'], 64, 3,
'res2_1')
end_points['conv2_2'] = block34(end_points['conv2_1'], 64, 3,
'res2_2')
end_points['conv2_3'] = block34(end_points['conv2_2'], 64, 3,
'res2_3')
# 56 * 56
end_points['conv3_1'] = block34(end_points['conv2_3'],
128,
3,
'res3_1',
stride=2,
ex=True)
end_points['conv3_2'] = block34(end_points['conv3_1'], 128, 3,
'res3_2')
end_points['conv3_3'] = block34(end_points['conv3_2'], 128, 3,
'res3_3')
end_points['conv3_4'] = block34(end_points['conv3_3'], 128, 3,
'res3_4')
# 28 * 28
end_points['conv4_1'] = block34(end_points['conv3_4'],
256,
3,
'res4_1',
stride=2,
ex=True)
end_points['conv4_2'] = block34(end_points['conv4_1'], 256, 3,
'res4_2')
end_points['conv4_3'] = block34(end_points['conv4_2'], 256, 3,
'res4_3')
end_points['conv4_4'] = block34(end_points['conv4_3'], 256, 3,
'res4_4')
end_points['conv4_5'] = block34(end_points['conv4_4'], 256, 3,
'res4_5')
end_points['conv4_6'] = block34(end_points['conv4_5'], 256, 3,
'res4_6')
# 14 * 14
end_points['conv5_1'] = block34(end_points['conv4_6'],
512,
3,
'res5_1',
stride=2,
ex=True)
end_points['conv5_2'] = block34(end_points['conv5_1'], 512, 3,
'res5_2')
end_points['conv5_3'] = block34(end_points['conv5_2'], 512, 3,
'res5_3')
#7 * 7 * 512
end_points['avg'] = ops.avg_pool(end_points['conv5_3'], [7, 7],
stride=1,
padding='SAME',
scope='avg_pooling')
end_points['flatten'] = ops.flatten(end_points['avg'],
scope='flatten')
end_points['logits'] = ops.fc(end_points['flatten'],
num_classes,
scope='logits')
return end_points['logits'], end_points
def testCreateFCWithScope(self):
height, width = 3, 3
with self.test_session():
inputs = tf.random_uniform((5, height * width * 3), seed=1)
output = ops.fc(inputs, 32, scope='fc1')
self.assertEqual(output.op.name, 'fc1/Relu')
