def __init__(self, width, height, channels, classes):
super(LeNet, self).__init__()
# activation_function = "tanh"
activation_function = "relu"
self._number_of_classes = classes
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x_image = tf.reshape(x, [-1, width, height, channels],
name="input_reshape")
with tf.variable_scope("conv1"):
out = conv(x_image, 5, 5, 20, activation=activation_function)
out = max_pool_2x2(out)
with tf.variable_scope("conv2"):
out = conv(out, 5, 5, 50, activation=activation_function)
out = max_pool_2x2(out)
dims = out.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2"):
flatten = tf.reshape(out, [-1, int(flatten_size)])
with tf.variable_scope("fc1"):
out = fc(flatten, 500)
if activation_function == "relu":
out = tf.nn.relu(out)
elif activation_function == "tanh":
out = tf.nn.tanh(out)
with tf.variable_scope("fc2"):
self._logits = fc(out, classes)
if classes > 1:
self._predictions = tf.nn.softmax(self._logits)
else:
self._predictions = self._logits
def __init__(self, width=28, height=28, channels=1, classes=10,
hidden_layers=[], dropout=[], activation_fn=nn.relu):
super(MultiLayerPerceptron, self).__init__()
self._number_of_classes = classes
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
# Hidden dropout
dropout_default = tf.zeros([2], dtype=tf.float32)
self._hidden_dropout = tf.placeholder_with_default(dropout_default,
[None],
name="hidden_dropout")
# Input dropout
self._input_dropout = tf.placeholder_with_default(tf.constant(0.0, dtype=tf.float32),
[],
name="input_dropout")
# Activations
self._activations = tf.placeholder_with_default([0, 0],
[None],
name="activations")
self._inputs = x
#out = tf.nn.dropout(x, keep_prob=tf.constant(1.0, dtype=tf.float32) - self._input_dropout)
out = x
for idx, h in enumerate(hidden_layers):
with tf.variable_scope("fc%d" % idx):
if 1 == self._activations[idx]:
activation_fn = nn.tanh
else:
activation_fn = nn.relu
out = fc(out, h, activation_fn=activation_fn)
if self._hidden_dropout.get_shape()[0] > idx:
with tf.variable_scope("dropout%d" % idx):
out = tf.nn.dropout(out, keep_prob=1.0 - self._hidden_dropout[idx])
with tf.variable_scope("fc%d" % len(hidden_layers)):
out = fc(out, classes)
self._logits = out
if classes > 1:
self._predictions = tf.nn.softmax(out)
else:
self._predictions = self._logits
def __init__(self, width, height, channels, classes, number_of_gpus=4):
super(LeNet, self).__init__()
# activation_function = "tanh"
activation_function = "relu"
self._number_of_classes = classes
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x_image = tf.reshape(x, [-1, width, height, channels],
name="input_reshape")
if number_of_gpus <= 1:
gpus = 1
else:
gpus = number_of_gpus
if gpus > 1:
x_image_arr = tf.split(x_image, gpus, axis=0)
else:
x_image_arr = [x_image]
classes_arr = [classes] * gpus
logits_arr = [0.0] * gpus
for gpu in range(gpus):
with tf.device('/gpu:' + str(gpu)):
with tf.variable_scope("conv1_" + str(gpu)):
out = conv(x_image_arr[gpu], 5, 5, 20, activation=activation_function)
out = max_pool_2x2(out)
with tf.variable_scope("conv2_" + str(gpu)):
out = conv(out, 5, 5, 50, activation=activation_function)
out = max_pool_2x2(out)
dims = out.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2_" + str(gpu)):
flatten = tf.reshape(out, [-1, int(flatten_size)])
with tf.variable_scope("fc1_" + str(gpu)):
out = fc(flatten, 500)
if activation_function == "relu":
out = tf.nn.relu(out)
elif activation_function == "tanh":
out = tf.nn.tanh(out)
with tf.variable_scope("fc2_" + str(gpu)):
logits_arr[gpu] = fc(out, classes_arr[gpu])
self._logits = tf.concat(logits_arr, 0)
# print(self._logits.graph.as_graph_def())
if classes > 1:
self._predictions = tf.nn.softmax(self._logits)
else:
self._predictions = self._logits
def __init__(self, width=299, height=299, channels=3, classes=10):
super(InceptionV4, self).__init__()
assert width == height, "width and height must be the same"
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x = tf.reshape(x, [-1, width, height, channels])
self._number_of_classes = classes
max_w = 299
min_w = 299 // 3
with tf.variable_scope("resize1"):
if width < min_w:
x = tf.image.resize_images(x, [min_w, min_w])
elif width == 299:
pass # do nothing
else:
x = tf.image.resize_images(x, [max_w, max_w])
with tf.variable_scope("stem"):
out = stem(x)
for i in range(4):
with tf.variable_scope("incA" + str(i)):
out = inceptionA(out)
with tf.variable_scope("redA"):
out = reductionA(out, k=192, l=224, m=256, n=384)
for i in range(7):
with tf.variable_scope("incB" + str(i)):
out = inceptionB(out)
with tf.variable_scope("redB"):
out = reductionB(out)
for i in range(3):
with tf.variable_scope("incC" + str(i)):
out = inceptionC(out)
a, b = out.get_shape().as_list()[1:3]
out = tf.nn.avg_pool(out,
ksize=[1, a, b, 1],
strides=[1, 1, 1, 1],
padding="VALID")
dims = out.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2"):
out = tf.reshape(out, [-1, int(flatten_size)])
with tf.variable_scope("fc1"):
out = fc(out, classes)
self._logits = out
if classes > 1:
self._predictions = tf.nn.softmax(self._logits)
else:
self._predictions = self._logits
def __init__(self, width=299, height=299, channels=3, classes=10):
super(InceptionV3, self).__init__()
self._number_of_classes = classes
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x = tf.reshape(x, [-1, width, height, channels])
max_w = 299
min_w = 299 // 3
with tf.variable_scope("resize1"):
if width < min_w:
x = tf.image.resize_images(x, [min_w, min_w])
elif width > 299:
x = tf.image.resize_images(x, [max_w, max_w])
# Stage 1
out = conv3x3(x,
32,
stride=2,
padding="VALID",
batch_norm=use_batch_norm)
out = conv3x3(out, 32, padding="VALID", batch_norm=use_batch_norm)
out = conv3x3(out, 64, batch_norm=use_batch_norm)
out = max_pool_3x3(out, padding="VALID")
# Stage 2
out = conv3x3(out, 80, padding="VALID", batch_norm=use_batch_norm)
out = conv3x3(out, 192, padding="VALID", batch_norm=use_batch_norm)
out = max_pool_3x3(out, padding="VALID")
# Stage 3
out = inception7A(out, 64, 64, 96, 96, 48, 64, 32)
out = inception7A(out, 64, 64, 96, 96, 48, 64, 64)
out = inception7A(out, 64, 64, 96, 96, 48, 64, 64)
out = inception7B(out, 384, 64, 96, 96)
# Stage 4
out = inception7C(out, 192, 128, 128, 192, 128, 128, 128, 128, 192,
192)
out = inception7C(out, 192, 160, 160, 192, 160, 160, 160, 160, 192,
192)
out = inception7C(out, 192, 160, 160, 192, 160, 160, 160, 160, 192,
192)
out = inception7C(out, 192, 192, 192, 192, 192, 192, 192, 192, 192,
192)
out = inception7D(out, 192, 320, 192, 192, 192, 192)
# Stage 5
out = inception7E(out, 320, 384, 384, 384, 448, 384, 384, 384, 192,
avg_pool_3x3)
out = inception7E(out, 320, 384, 384, 384, 448, 384, 384, 384, 192,
max_pool_3x3)
out = tf.nn.avg_pool(out,
ksize=[1, 8, 8, 1],
strides=[1, 1, 1, 1],
padding='SAME')
dims = out.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2"):
out = tf.reshape(out, [-1, int(flatten_size)])
with tf.variable_scope("fc1"):
out = fc(out, classes)
self._logits = out
if classes > 1:
self._predictions = tf.nn.softmax(self._logits)
else:
self._predictions = self._logits
def __init__(self, width, height, channels, classes):
super(AlexNet, self).__init__()
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x = tf.reshape(self._inputs, [-1, width, height, channels])
self._number_of_classes = classes
with tf.variable_scope("resize1"):
if width < 224:
x = tf.image.resize_images(x, [48, 48])
else:
x = tf.image.resize_images(x, [224, 224])
with tf.variable_scope("conv1"):
conv1 = conv11x11(x, 96, stride=4, padding='VALID')
pool1 = max_pool_3x3(conv1, padding='VALID')
lrn1 = tf.nn.lrn(pool1, 5, bias=1.0, alpha=0.0001, beta=0.75)
with tf.variable_scope("conv2"):
conv2 = conv5x5(lrn1, 256)
pool2 = max_pool_3x3(conv2, padding='VALID')
lrn2 = tf.nn.lrn(pool2, 5, bias=1.0, alpha=0.0001, beta=0.75)
with tf.variable_scope("conv3"):
conv3 = conv3x3(lrn2, 384)
with tf.variable_scope("conv4"):
conv4 = conv3x3(conv3, 384)
with tf.variable_scope("conv5"):
conv5 = conv3x3(conv4, 256)
pool3 = max_pool_3x3(conv5)
lrn3 = tf.nn.lrn(pool3, 5, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
dims = lrn3.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2"):
flatten = tf.reshape(lrn3, [-1, int(flatten_size)])
with tf.variable_scope("fc1"):
fc1 = fc(flatten, 4096)
relu1 = tf.nn.relu(fc1)
with tf.variable_scope("fc2"):
fc2 = fc(relu1, 4096)
relu2 = tf.nn.relu(fc2)
with tf.variable_scope("fc3"):
y = fc(relu2, classes)
self._logits = y
if classes > 1:
self._predictions = tf.nn.softmax(y)
else:
self._predictions = self._logits
def __init__(self, width, height, channels, classes):
super(AlexNet, self).__init__()
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x = tf.reshape(self._inputs, [-1, width, height, channels])
self._number_of_classes = classes
with tf.variable_scope("resize1"):
if width < 224:
x = tf.image.resize_images(x, [48, 48])
else:
x = tf.image.resize_images(x, [224, 224])
with tf.variable_scope("conv1"):
conv1 = conv11x11(x, 96, stride=4, padding='VALID')
pool1 = max_pool_3x3(conv1, padding='VALID')
lrn1 = tf.nn.lrn(pool1, 5, bias=1.0, alpha=0.0001, beta=0.75)
with tf.variable_scope("conv2"):
conv2 = conv5x5(lrn1, 256)
pool2 = max_pool_3x3(conv2, padding='VALID')
lrn2 = tf.nn.lrn(pool2, 5, bias=1.0, alpha=0.0001, beta=0.75)
with tf.variable_scope("conv3"):
conv3 = conv3x3(lrn2, 384)
with tf.variable_scope("conv4"):
conv4 = conv3x3(conv3, 384)
with tf.variable_scope("conv5"):
conv5 = conv3x3(conv4, 256)
pool3 = max_pool_3x3(conv5)
lrn3 = tf.nn.lrn(pool3, 5, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
dims = lrn3.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2"):
flatten = tf.reshape(lrn3, [-1, int(flatten_size)])
with tf.variable_scope("fc1"):
fc1 = fc(flatten, 4096)
relu1 = tf.nn.relu(fc1)
with tf.variable_scope("fc2"):
fc2 = fc(relu1, 4096)
relu2 = tf.nn.relu(fc2)
with tf.variable_scope("fc3"):
y = fc(relu2, classes)
self._logits = y
if classes > 1:
self._predictions = tf.nn.softmax(y)
else:
self._predictions = self._logits
def __init__(self, width=299, height=299, channels=3, classes=10):
super(ResNet, self).__init__()
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
x = tf.reshape(x, [-1, width, height, channels])
self._number_of_classes = classes
# normalizer_params = { 'is_training': is_training() }
# adapted from https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/learn/resnet.py
# Configurations for each bottleneck group.
# activation = tf.nn.relu
BottleneckGroup = namedtuple('BottleneckGroup',
['num_blocks', 'num_filters', 'bottleneck_size'])
groups = [
BottleneckGroup(3, 128, 32), BottleneckGroup(3, 256, 64),
# BottleneckGroup(3, 512, 128), BottleneckGroup(3, 1024, 256)
]
# First convolution expands to 64 channels
with tf.variable_scope('conv_layer1'):
# net = convolution2d(
# x, 64, 7, normalizer_fn=batch_norm, normalizer_params=normalizer_params, activation_fn=activation)
net = conv(x, 7, 7, 64)
# Max pool
net = tf.nn.max_pool(net, [1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')
# First chain of resnets
with tf.variable_scope('conv_layer2'):
# net = convolution2d(net, groups[0].num_filters, 1, padding='VALID')
net = conv(net, 1, 1, groups[0].num_filters, padding='VALID')
# Create the bottleneck groups, each of which contains `num_blocks`
# bottleneck groups.
for group_i, group in enumerate(groups):
for block_i in range(group.num_blocks):
name = 'group_%d/block_%d' % (group_i, block_i)
# 1x1 convolution responsible for reducing dimension
with tf.variable_scope(name + '/conv_in'):
# conv = convolution2d(
# net,
# group.bottleneck_size,
# 1,
# padding='VALID',
# activation_fn=activation,
# normalizer_fn=batch_norm,
# normalizer_params=normalizer_params)
convLayer = conv(net, 1, 1, group.bottleneck_size, padding='VALID')
with tf.variable_scope(name + '/conv_bottleneck'):
# conv = convolution2d(
# conv,
# group.bottleneck_size,
# 3,
# padding='SAME',
# activation_fn=activation,
# normalizer_fn=batch_norm,
# normalizer_params=normalizer_params)
convLayer = conv(convLayer, 3, 3, group.bottleneck_size)
# 1x1 convolution responsible for restoring dimension
with tf.variable_scope(name + '/conv_out'):
input_dim = net.get_shape()[-1].value
# conv = convolution2d(
# conv,
# input_dim,
# 1,
# padding='VALID',
# activation_fn=activation,
# normalizer_fn=batch_norm,
# normalizer_params=normalizer_params)
convLayer = conv(convLayer, 1, 1, input_dim, padding='VALID')
# shortcut connections that turn the network into its counterpart
# residual function (identity shortcut)
net = convLayer + net
try:
# upscale to the next group size
next_group = groups[group_i + 1]
with tf.variable_scope('block_%d/conv_upscale' % group_i):
# net = convolution2d(
# net,
# next_group.num_filters,
# 1,
# activation_fn=None,
# biases_initializer=None,
# padding='SAME')
net = conv(net, 1, 1, next_group.num_filters)
except IndexError:
pass
net_shape = net.get_shape().as_list()
net = tf.nn.avg_pool(
net,
ksize=[1, net_shape[1], net_shape[2], 1],
strides=[1, 1, 1, 1],
padding='VALID')
out = net
dims = out.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
out = tf.reshape(out, [-1, int(flatten_size)])
out = fc(out, classes)
self._logits = out
if classes > 1:
self._predictions = tf.nn.softmax(self._logits)
else:
self._predictions = self._logits
def __init__(self, width=299, height=299, channels=3, classes=10):
super(InceptionV3, self).__init__()
self._number_of_classes = classes
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x = tf.reshape(x, [-1, width, height, channels])
max_w = 299
min_w = 299 // 3
with tf.variable_scope("resize1"):
if width < min_w:
x = tf.image.resize_images(x, [min_w, min_w])
elif width > 299:
x = tf.image.resize_images(x, [max_w, max_w])
# Stage 1
out = conv3x3(x, 32, stride=2, padding="VALID", batch_norm=use_batch_norm)
out = conv3x3(out, 32, padding="VALID", batch_norm=use_batch_norm)
out = conv3x3(out, 64, batch_norm=use_batch_norm)
out = max_pool_3x3(out, padding="VALID")
# Stage 2
out = conv3x3(out, 80, padding="VALID", batch_norm=use_batch_norm)
out = conv3x3(out, 192, padding="VALID", batch_norm=use_batch_norm)
out = max_pool_3x3(out, padding="VALID")
# Stage 3
out = inception7A(out, 64, 64, 96, 96, 48, 64, 32)
out = inception7A(out, 64, 64, 96, 96, 48, 64, 64)
out = inception7A(out, 64, 64, 96, 96, 48, 64, 64)
out = inception7B(out, 384, 64, 96, 96)
# Stage 4
out = inception7C(out, 192, 128, 128, 192, 128, 128, 128, 128, 192, 192)
out = inception7C(out, 192, 160, 160, 192, 160, 160, 160, 160, 192, 192)
out = inception7C(out, 192, 160, 160, 192, 160, 160, 160, 160, 192, 192)
out = inception7C(out, 192, 192, 192, 192, 192, 192, 192, 192, 192, 192)
out = inception7D(out, 192, 320, 192, 192, 192, 192)
# Stage 5
out = inception7E(out, 320, 384, 384, 384, 448, 384, 384, 384, 192, avg_pool_3x3)
out = inception7E(out, 320, 384, 384, 384, 448, 384, 384, 384, 192, max_pool_3x3)
out = tf.nn.avg_pool(out, ksize=[1, 8, 8, 1], strides=[1, 1, 1, 1], padding='SAME')
dims = out.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2"):
out = tf.reshape(out, [-1, int(flatten_size)])
with tf.variable_scope("fc1"):
out = fc(out, classes)
self._logits = out
if classes > 1:
self._predictions = tf.nn.softmax(self._logits)
else:
self._predictions = self._logits
def __init__(self, width=299, height=299, channels=3, classes=10):
super(InceptionV4, self).__init__()
assert width == height, "width and height must be the same"
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x = tf.reshape(x, [-1, width, height, channels])
self._number_of_classes = classes
max_w = 299
min_w = 299 // 3
with tf.variable_scope("resize1"):
if width < min_w:
x = tf.image.resize_images(x, [min_w, min_w])
elif width == 299:
pass # do nothing
else:
x = tf.image.resize_images(x, [max_w, max_w])
with tf.variable_scope("stem"):
out = stem(x)
for i in range(4):
with tf.variable_scope("incA" + str(i)):
out = inceptionA(out)
with tf.variable_scope("redA"):
out = reductionA(out, k=192, l=224, m=256, n=384)
for i in range(7):
with tf.variable_scope("incB" + str(i)):
out = inceptionB(out)
with tf.variable_scope("redB"):
out = reductionB(out)
for i in range(3):
with tf.variable_scope("incC" + str(i)):
out = inceptionC(out)
a, b = out.get_shape().as_list()[1:3]
out = tf.nn.avg_pool(out, ksize=[1, a, b, 1],
strides=[1, 1, 1, 1], padding="VALID")
dims = out.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2"):
out = tf.reshape(out, [-1, int(flatten_size)])
with tf.variable_scope("fc1"):
out = fc(out, classes)
self._logits = out
if classes > 1:
self._predictions = tf.nn.softmax(self._logits)
else:
self._predictions = self._logits
def __init__(self, width, height, channels, classes, number_of_gpus=4):
super(LeNet, self).__init__()
# activation_function = "tanh"
activation_function = "relu"
self._number_of_classes = classes
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x_image = tf.reshape(x, [-1, width, height, channels],
name="input_reshape")
if number_of_gpus <= 1:
gpus = 1
else:
gpus = number_of_gpus
if gpus > 1:
x_image_arr = tf.split(x_image, gpus, axis=0)
else:
x_image_arr = [x_image]
classes_arr = [classes] * gpus
logits_arr = [0.0] * gpus
for gpu in range(gpus):
with tf.device('/gpu:' + str(gpu)):
with tf.variable_scope("conv1"):
out = conv(x_image_arr[gpu],
5,
5,
20,
activation=activation_function)
out = max_pool_2x2(out)
with tf.variable_scope("conv2"):
out = conv(out, 5, 5, 50, activation=activation_function)
out = max_pool_2x2(out)
dims = out.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2"):
flatten = tf.reshape(out, [-1, int(flatten_size)])
with tf.variable_scope("fc1"):
out = fc(flatten, 500)
if activation_function == "relu":
out = tf.nn.relu(out)
elif activation_function == "tanh":
out = tf.nn.tanh(out)
with tf.variable_scope("fc2"):
logits_arr[gpu] = fc(out, classes_arr[gpu])
self._logits = tf.concat(logits_arr, 0)
# print(self._logits.graph.as_graph_def())
if classes > 1:
self._predictions = tf.nn.softmax(self._logits)
else:
self._predictions = self._logits
def __init__(self, width=28, height=28, channels=1, classes=10):
super(VGG16, self).__init__()
assert width == height, "width and height must be the same"
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x = tf.reshape(x, [-1, width, height, channels])
self._number_of_classes = classes
with tf.variable_scope("resize1"):
if width < 224:
x = tf.image.resize_images(x, [48, 48])
elif width > 224:
x = tf.image.resize_images(x, [224, 224])
with tf.variable_scope("conv1"):
out = conv3x3(x, 64, stride=1)
out = max_pool_2x2(out, stride=2)
with tf.variable_scope("conv2"):
out = conv3x3(out, 128, stride=1)
out = max_pool_2x2(out, stride=2)
with tf.variable_scope("conv3"):
out = conv3x3(out, 256, stride=1)
with tf.variable_scope("conv4"):
out = conv3x3(out, 256, stride=1)
out = max_pool_2x2(out, stride=2)
with tf.variable_scope("conv5"):
out = conv3x3(out, 512, stride=1)
with tf.variable_scope("conv6"):
out = conv3x3(out, 512, stride=1)
out = max_pool_2x2(out, stride=2)
with tf.variable_scope("conv7"):
out = conv3x3(out, 512, stride=1)
with tf.variable_scope("conv8"):
out = conv3x3(out, 512, stride=1)
out = max_pool_2x2(out, stride=2)
dims = out.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2"):
out = tf.reshape(out, [-1, int(flatten_size)])
# fully connected
with tf.variable_scope("fc1"):
out = fc(out, 4096)
out = tf.nn.relu(out)
with tf.variable_scope("fc2"):
out = fc(out, 4096)
out = tf.nn.relu(out)
with tf.variable_scope("fc3"):
y = fc(out, classes)
self._logits = y
if classes > 1:
self._predictions = tf.nn.softmax(y)
else:
self._predictions = self._logits
def __init__(self, width=28, height=28, channels=1, classes=10):
super(VGG16, self).__init__()
assert width == height, "width and height must be the same"
size = width * height * channels
x = tf.placeholder(tf.float32, [None, size], name="x")
self._inputs = x
with tf.variable_scope("reshape1"):
x = tf.reshape(x, [-1, width, height, channels])
self._number_of_classes = classes
with tf.variable_scope("resize1"):
if width < 224:
x = tf.image.resize_images(x, [48, 48])
elif width > 224:
x = tf.image.resize_images(x, [224, 224])
with tf.variable_scope("conv1"):
out = conv3x3(x, 64, stride=1)
out = max_pool_2x2(out, stride=2)
with tf.variable_scope("conv2"):
out = conv3x3(out, 128, stride=1)
out = max_pool_2x2(out, stride=2)
with tf.variable_scope("conv3"):
out = conv3x3(out, 256, stride=1)
with tf.variable_scope("conv4"):
out = conv3x3(out, 256, stride=1)
out = max_pool_2x2(out, stride=2)
with tf.variable_scope("conv5"):
out = conv3x3(out, 512, stride=1)
with tf.variable_scope("conv6"):
out = conv3x3(out, 512, stride=1)
out = max_pool_2x2(out, stride=2)
with tf.variable_scope("conv7"):
out = conv3x3(out, 512, stride=1)
with tf.variable_scope("conv8"):
out = conv3x3(out, 512, stride=1)
out = max_pool_2x2(out, stride=2)
dims = out.get_shape().as_list()
flatten_size = 1
for d in dims[1:]:
flatten_size *= d
with tf.variable_scope("reshape2"):
out = tf.reshape(out, [-1, int(flatten_size)])
# fully connected
with tf.variable_scope("fc1"):
out = fc(out, 4096)
out = tf.nn.relu(out)
with tf.variable_scope("fc2"):
out = fc(out, 4096)
out = tf.nn.relu(out)
with tf.variable_scope("fc3"):
y = fc(out, classes)
self._logits = y
if classes > 1:
self._predictions = tf.nn.softmax(y)
else:
self._predictions = self._logits
