def vgg_16(inputs,
num_classes=1000,
is_training=True,
dropout_keep_prob=0.5,
spatial_squeeze=True,
scope='vgg_16'):
"""Oxford Net VGG 16-Layers version D Example.
Note: All the fully_connected layers have been transformed to conv2d layers.
To use in classification mode, resize input to 224x224.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
spatial_squeeze: whether or not should squeeze the spatial dimensions of the
outputs. Useful to remove unnecessary dimensions for classification.
scope: Optional scope for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
with tf.variable_scope(scope, 'vgg_16', [inputs]) as sc:
end_points_collection = sc.name + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
with slim.arg_scope([slim.conv2d, slim.fully_connected, slim.max_pool2d],
outputs_collections=end_points_collection):
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
net = slim.max_pool2d(net, [2, 2], scope='pool1')
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
net = slim.max_pool2d(net, [2, 2], scope='pool2')
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
net = slim.max_pool2d(net, [2, 2], scope='pool3')
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
net = slim.max_pool2d(net, [2, 2], scope='pool4')
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
net = slim.max_pool2d(net, [2, 2], scope='pool5')
# Use conv2d instead of fully_connected layers.
net = slim.conv2d(net, 4096, [7, 7], padding='VALID', scope='fc6')
net = slim.dropout(net, dropout_keep_prob, is_training=is_training,
scope='dropout6')
net = slim.conv2d(net, 4096, [1, 1], scope='fc7')
net = slim.dropout(net, dropout_keep_prob, is_training=is_training,
scope='dropout7')
net = slim.conv2d(net, 1000, [1, 1],
activation_fn=None,
normalizer_fn=None,
scope='fc8')
# Convert end_points_collection into a end_point dict.
end_points = slim.utils.convert_collection_to_dict(end_points_collection)
if spatial_squeeze:
# net = tf.squeeze(net, [1, 2], name='fc8/squeezed')
net = custom_layers.spatial_squeeze(net)
net = custom_layers.pad_logits(net, pad=(num_classes - 1000, 0))
end_points[sc.name + '/fc8'] = net
return net, end_points
def vgg_16(inputs,
num_classes=1000,
is_training=True,
dropout_keep_prob=0.5,
spatial_squeeze=True,
scope='vgg_16'):
"""Oxford Net VGG 16-Layers version D Example.
Note: All the fully_connected layers have been transformed to conv2d layers.
To use in classification mode, resize input to 224x224.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
spatial_squeeze: whether or not should squeeze the spatial dimensions of the
outputs. Useful to remove unnecessary dimensions for classification.
scope: Optional scope for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
with tf.variable_scope(scope, 'vgg_16', [inputs]) as sc:
end_points_collection = sc.name + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
with slim.arg_scope(
[slim.conv2d, slim.fully_connected, slim.max_pool2d],
outputs_collections=end_points_collection):
net = slim.repeat(inputs,
2,
slim.conv2d,
64, [3, 3],
scope='conv1')
net = slim.max_pool2d(net, [2, 2], scope='pool1')
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
net = slim.max_pool2d(net, [2, 2], scope='pool2')
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
net = slim.max_pool2d(net, [2, 2], scope='pool3')
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
net = slim.max_pool2d(net, [2, 2], scope='pool4')
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
net = slim.max_pool2d(net, [2, 2], scope='pool5')
# Use conv2d instead of fully_connected layers.
net = slim.conv2d(net, 4096, [7, 7], padding='VALID', scope='fc6')
net = slim.dropout(net,
dropout_keep_prob,
is_training=is_training,
scope='dropout6')
net = slim.conv2d(net, 4096, [1, 1], scope='fc7')
net = slim.dropout(net,
dropout_keep_prob,
is_training=is_training,
scope='dropout7')
net = slim.conv2d(net,
1000, [1, 1],
activation_fn=None,
normalizer_fn=None,
scope='fc8')
# Convert end_points_collection into a end_point dict.
end_points = slim.utils.convert_collection_to_dict(
end_points_collection)
if spatial_squeeze:
# net = tf.squeeze(net, [1, 2], name='fc8/squeezed')
net = custom_layers.spatial_squeeze(net)
net = custom_layers.pad_logits(net,
pad=(num_classes - 1000, 0))
end_points[sc.name + '/fc8'] = net
return net, end_points
def mobilenets(inputs,
num_classes=1000,
width_multiplier=1.0,
is_training=True,
dropout_keep_prob=0.5,
scope='MobileNets'):
"""MobileNets implementation.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
scope: Optional scope for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
kernel_size = [3, 3]
padding = [(kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2]
def mobilenet_block(net, num_out_channels, stride=[1, 1], scope=None):
"""Basic MobileNet block combining:
- depthwise conv + BN + relu
- 1x1 conv + BN + relu
"""
with tf.variable_scope(scope, 'block', [net]) as sc:
num_out_channels = int(num_out_channels * width_multiplier)
if stride[0] == 1 and stride[1] == 1:
# Depthwise convolution with stride=1
net = custom_layers.depthwise_convolution2d(net,
kernel_size,
depth_multiplier=1,
stride=stride,
scope='conv_dw')
else:
# Mimic CAFFE padding if stride > 1.
net = custom_layers.pad2d(net, pad=padding)
net = custom_layers.depthwise_convolution2d(net,
kernel_size,
padding='VALID',
depth_multiplier=1,
stride=stride,
scope='conv_dw')
# Pointwise convolution.
net = slim.conv2d(net, num_out_channels, [1, 1], scope='conv_pw')
return net
with tf.variable_scope(scope, 'MobileNets', [inputs]) as sc:
end_points = {}
# First full convolution...
net = custom_layers.pad2d(inputs, pad=padding)
net = slim.conv2d(net,
32,
kernel_size,
stride=[2, 2],
padding='VALID',
scope='conv1')
# net = slim.conv2d(inputs, 32, [ksize, ksize], stride=[2, 2], scope='conv1')
# Then, MobileNet blocks!
net = mobilenet_block(net, 64, scope='block2')
net = mobilenet_block(net, 128, stride=[2, 2], scope='block3')
net = mobilenet_block(net, 128, scope='block4')
net = mobilenet_block(net, 256, stride=[2, 2], scope='block5')
net = mobilenet_block(net, 256, scope='block6')
net = mobilenet_block(net, 512, stride=[2, 2], scope='block7')
# Intermediate blocks...
for i in range(5):
net = mobilenet_block(net, 512, scope='block%i' % (i + 8))
# Final blocks.
net = mobilenet_block(net, 1024, stride=[2, 2], scope='block13')
net = mobilenet_block(net, 1024, scope='block14')
# Spatial pooling + fully connected layer.
net = custom_layers.spatial_mean(net,
keep_dims=True,
scope='spatial_mean14')
net = slim.conv2d(net,
1000, [1, 1],
activation_fn=None,
normalizer_fn=None,
normalizer_params=None,
biases_initializer=tf.zeros_initializer(),
scope='conv_fc15')
net = custom_layers.spatial_squeeze(net)
# net = slim.fully_connected(net, 1000, scope='fc15')
# Logits padding...
net = custom_layers.pad_logits(net, pad=(num_classes - 1000, 0))
return net, end_points
def mobilenets_btree(inputs,
num_classes=1000,
kernel_size=[3, 3],
width_multiplier=1.0,
dropouts=[0.5],
pad_logits=True,
is_training=True,
reuse=None,
scope='MobileNets'):
"""MobileNets implementation.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
scope: Optional scope for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
# MobileNets kernel size and padding (for layers with stride > 1).
# kernel_size = [3, 3]
padding = [(kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2]
def mobilenet_block(net, num_out_channels, stride=[1, 1], scope=None):
"""Basic MobileNet block combining:
- depthwise conv + BN + relu
- 1x1 conv + BN + relu
"""
with tf.variable_scope(scope, 'block', [net]) as sc:
num_out_channels = int(num_out_channels * width_multiplier)
if stride[0] == 1 and stride[1] == 1:
# Depthwise convolution with stride=1
net = custom_layers.depthwise_convolution2d(net,
kernel_size,
depth_multiplier=1,
stride=stride,
scope='conv_dw')
else:
# Mimic CAFFE padding if stride > 1 => usually better accuracy.
net = custom_layers.pad2d(net, pad=padding)
net = custom_layers.depthwise_convolution2d(net,
kernel_size,
padding='VALID',
depth_multiplier=1,
stride=stride,
scope='conv_dw')
# Pointwise convolution.
net = slim.conv2d(net, num_out_channels, [1, 1], scope='conv_pw')
return net
def mobilenet_block_btree_v1(net,
num_out_channels,
stride=[1, 1],
split=2,
scope=None):
"""Basic MobileNet block combining:
- depthwise conv + BN + relu
- 1x1 conv + BN + relu
"""
with tf.variable_scope(scope, 'block', [net]) as sc:
num_out_channels = int(num_out_channels * width_multiplier)
# Depthwise convolution with stride=1
net = custom_layers.depthwise_convolution2d(net,
kernel_size,
depth_multiplier=1,
stride=stride,
scope='conv_dw')
# Split-pointwise convolution.
net = btree_layers.conv2d_1x1_split(net,
num_out_channels,
split=split,
scope='conv_pw_split')
return net
def mobilenet_block_btree_v2(net,
num_out_channels,
stride=[1, 1],
split=2,
scope=None):
"""Combination of ResNets block and B-tree.
"""
with tf.variable_scope(scope, 'block', [net]) as sc:
# Start with Batch Norm.
net = custom_layers.batch_norm(net)
# Depthwise convolution with stride=1
net = custom_layers.depthwise_convolution2d(net,
kernel_size,
depth_multiplier=1,
stride=stride,
activation_fn=None,
scope='conv_dw')
# Split-pointwise convolution.
num_out_channels = int(num_out_channels * width_multiplier)
net = btree_layers.conv2d_1x1_split(net,
num_out_channels,
split=split,
activation_fn=None,
normalizer_fn=None,
scope='conv_pw_split')
return net
with tf.variable_scope(scope, 'MobileNets', [inputs], reuse=reuse) as sc:
end_points = {}
# First full convolution...
net = custom_layers.pad2d(inputs, pad=padding)
net = slim.conv2d(net,
32,
kernel_size,
stride=[2, 2],
padding='VALID',
scope='conv1')
# net = slim.conv2d(inputs, 32, kernel_size, stride=[2, 2],
# padding='SAME', scope='conv1')
# Then, MobileNet blocks!
net = mobilenet_block(net, 64, scope='block2')
net = mobilenet_block(net, 128, stride=[2, 2], scope='block3')
net = mobilenet_block(net, 128, scope='block4')
net = mobilenet_block(net, 256, stride=[2, 2], scope='block5')
net = mobilenet_block(net, 256, scope='block6')
net = mobilenet_block(net, 512, stride=[2, 2], scope='block7')
# Intermediate blocks...
for i in range(8, 16):
with tf.variable_scope(scope, 'resblock_%i' % i, [net]) as sc:
# Residual block...
res = net
net = mobilenet_block_btree_v2(net,
512,
split=4,
scope='block%i_a' % i)
net = btree_layers.translate_channels(
net, delta=64, scope='ch_translate_%i_a' % i)
net = mobilenet_block_btree_v2(net,
512,
split=4,
scope='block%i_b' % i)
net = btree_layers.translate_channels(
net, delta=-64, scope='ch_translate_%i_b' % i)
net = mobilenet_block_btree_v2(net,
512,
split=4,
scope='block%i_c' % i)
net = tf.add(res, net, 'residual_sum_%i' % i)
net = custom_layers.batch_norm(net)
# Final blocks.
net = mobilenet_block(net, 1024, stride=[2, 2], scope='block13')
net = mobilenet_block(net, 1024, scope='block14')
# Spatial pooling + fully connected layer.
net = custom_layers.spatial_mean(net,
keep_dims=True,
scope='spatial_mean14')
net = slim.conv2d(net,
num_classes, [1, 1],
activation_fn=None,
normalizer_fn=None,
normalizer_params=None,
biases_initializer=tf.zeros_initializer(),
scope='conv_fc15')
net = custom_layers.spatial_squeeze(net)
# Logits padding: get everyone to the same number of classes.
if pad_logits:
net = custom_layers.pad_logits(net, pad=(num_classes - 1000, 0))
return net, end_points