def discriminator(inp, reuse=False):
with tf.variable_scope('Encoder', reuse=reuse):
# 64
inp = gaussnoise(inp, std=0.05)
conv1 = conv2d(inp, 128, kernel=3, strides=2, name=dname + 'conv1')
conv1 = lrelu(conv1, 0.2)
# 32
conv2 = tf.nn.dropout(conv1, keep_prob)
conv2 = conv2d(conv2, 256, kernel=3, strides=2, name=dname + 'conv2')
conv2 = batchnorm(conv2, is_training=is_train, name=dname + 'bn2')
conv2 = lrelu(conv2, 0.2)
# 16
conv3 = tf.nn.dropout(conv2, keep_prob)
conv3 = conv2d(conv3, 512, kernel=3, strides=2, name=dname + 'conv3')
conv3 = batchnorm(conv3, is_training=is_train, name=dname + 'bn3')
conv3 = lrelu(conv3, 0.2)
# 8
conv3b = conv2d(conv3, 512, kernel=3, strides=1, name=dname + 'conv3b')
conv3b = batchnorm(conv3b, is_training=is_train, name=dname + 'bn3b')
conv3b = lrelu(conv3b, 0.2)
conv4 = tf.nn.dropout(conv3b, keep_prob)
conv4 = conv2d(conv4, 1024, kernel=3, strides=2, name=dname + 'conv4')
conv4 = batchnorm(conv4, is_training=is_train, name=dname + 'bn4')
conv4 = lrelu(conv4, 0.2)
# 4
flat = flatten(conv4)
# Classifier
clspred = linear(flat, n_classes, name=dname + 'cpred')
# Decoder
g1 = conv2d(conv4, nout=512, kernel=3, name=dname + 'deconv1')
g1 = batchnorm(g1, is_training=tf.constant(True), name=dname + 'bn1g')
g1 = lrelu(g1, 0.2)
g2 = nnupsampling(g1, [8, 8])
g2 = conv2d(g2, nout=256, kernel=3, name=dname + 'deconv2')
g2 = batchnorm(g2, is_training=tf.constant(True), name=dname + 'bn2g')
g2 = lrelu(g2, 0.2)
g3 = nnupsampling(g2, [16, 16])
g3 = conv2d(g3, nout=128, kernel=3, name=dname + 'deconv3')
g3 = batchnorm(g3, is_training=tf.constant(True), name=dname + 'bn3g')
g3 = lrelu(g3, 0.2)
g4 = nnupsampling(g3, [32, 32])
g4 = conv2d(g4, nout=64, kernel=3, name=dname + 'deconv4')
g4 = batchnorm(g4, is_training=tf.constant(True), name=dname + 'bn4g')
g4 = lrelu(g4, 0.2)
g5 = nnupsampling(g4, [64, 64])
g5 = conv2d(g5, nout=32, kernel=3, name=dname + 'deconv5')
g5 = batchnorm(g5, is_training=tf.constant(True), name=dname + 'bn5g')
g5 = lrelu(g5, 0.2)
g5b = conv2d(g5, nout=3, kernel=3, name=dname + 'deconv5b')
g5b = tf.nn.tanh(g5b)
return clspred, g5b
def ModelSimple(X, is_training):
h_1 = lrelu(batch_norm(conv2d(X, 32, name='conv1'),
is_training, scope='bn1'), name='lrelu1')
h_2 = lrelu(batch_norm(conv2d(h_1, 64, name='conv2'),
is_training, scope='bn2'), name='lrelu2')
h_3 = lrelu(batch_norm(conv2d(h_2, 64, name='conv3'),
is_training, scope='bn3'), name='lrelu3')
h_3_flat = tf.reshape(h_3, [-1, 64 * 4 * 4])
return linear(h_3_flat, 10)
def ModelSimple(X, is_training):
h_1 = lrelu(batch_norm(conv2d(X, 32, name='conv1'),
is_training, scope='bn1'), name='lrelu1')
h_2 = lrelu(batch_norm(conv2d(h_1, 64, name='conv2'),
is_training, scope='bn2'), name='lrelu2')
h_3 = lrelu(batch_norm(conv2d(h_2, 64, name='conv3'),
is_training, scope='bn3'), name='lrelu3')
h_3_flat = tf.reshape(h_3, [-1, 64 * 4 * 4])
return linear(h_3_flat, 10)
def D_base(name, x):
with tf.variable_scope(name):
conv1 = conv('conv1', x, 4 * 4, 64, 2, 0, True)
l1 = lrelu('lrelu1', conv1)
conv2 = conv('conv2', l1, 4 * 4, 128, 2, 0, True)
ins2 = ins_norm('ins2', conv2)
l2 = lrelu('lrelu2', ins2)
conv3 = conv('conv3', l2, 4 * 4, 256, 2, 0, True)
ins3 = ins_norm('ins3', conv3)
l3 = lrelu('lrelu3', ins3)
conv4 = conv('conv4', l3, 4 * 4, 512, 1, 0, True)
ins4 = ins_norm('ins4', conv4)
l4 = lrelu('lrelu4', ins4)
conv5 = conv('conv5', l4, 4 * 4, 1, 1, 0, True)
return [l1, l2, l3, l4, conv5]
def generator(inp_z, inp_y, reuse=False):
with tf.variable_scope('Generator', reuse=reuse):
inp = tf.concat([inp_z, inp_y], 1)
sz = 4
g1 = linear(inp, 512 * sz * sz, name=gname + 'deconv1')
g1 = batchnorm(g1, is_training=tf.constant(True), name=gname + 'bn1g')
g1 = lrelu(g1, 0.2)
g1_reshaped = tf.reshape(g1, [-1, 512, sz, sz])
print 'genreshape: ' + str(g1_reshaped.get_shape().as_list())
g2 = nnupsampling(g1_reshaped, [8, 8])
g2 = conv2d(g2, nout=512, kernel=3, name=gname + 'deconv2')
g2 = batchnorm(g2, is_training=tf.constant(True), name=gname + 'bn2g')
g2 = lrelu(g2, 0.2)
g3 = nnupsampling(g2, [16, 16])
g3 = conv2d(g3, nout=256, kernel=3, name=gname + 'deconv3')
g3 = batchnorm(g3, is_training=tf.constant(True), name=gname + 'bn3g')
g3 = lrelu(g3, 0.2)
g4 = nnupsampling(g3, [32, 32])
g4 = conv2d(g4, nout=128, kernel=3, name=gname + 'deconv4')
g4 = batchnorm(g4, is_training=tf.constant(True), name=gname + 'bn4g')
g4 = lrelu(g4, 0.2)
g4b = conv2d(g4, nout=128, kernel=3, name=gname + 'deconv4b')
g4b = batchnorm(g4b,
is_training=tf.constant(True),
name=gname + 'bn4bg')
g4b = lrelu(g4b, 0.2)
g5 = nnupsampling(g4b, [64, 64])
g5 = conv2d(g5, nout=64, kernel=3, name=gname + 'deconv5')
g5 = batchnorm(g5, is_training=tf.constant(True), name=gname + 'bn5g')
g5 = lrelu(g5, 0.2)
g5b = conv2d(g5, nout=64, kernel=3, name=gname + 'deconv5b')
g5b = batchnorm(g5b,
is_training=tf.constant(True),
name=gname + 'bn5bg')
g5b = lrelu(g5b, 0.2)
g6 = nnupsampling(g5b, [128, 128])
g6 = conv2d(g6, nout=32, kernel=3, name=gname + 'deconv6')
g6 = batchnorm(g6, is_training=tf.constant(True), name=gname + 'bn6g')
g6 = lrelu(g6, 0.2)
g6b = conv2d(g6, nout=3, kernel=3, name=gname + 'deconv6b')
g6b = tf.nn.tanh(g6b)
g6b_64 = pool(g6b, fsize=3, strides=2, op='avg')
return g6b_64, g6b
def net_preloaded(vgg_layers,
input_image,
relu_kind=None,
pooling_type='avg',
padding='SAME'):
"""
This function read the vgg layers and create the net architecture
We need the input image to know the dimension of the input layer of the net
"""
net = {}
shapes = {}
_, height, width, numberChannels = input_image.shape # In order to have the right shape of the input
current = tf.Variable(
np.zeros((1, height, width, numberChannels), dtype=np.float32))
shapes['input'] = current.get_shape().as_list()
net['input'] = current
print('input', current.get_shape())
for i, name in enumerate(VGG19_LAYERS):
kind = name[:4]
shapes[name] = current.get_shape().as_list()
if (kind == 'conv'):
#if(VGG19_mat=='texturesyn_normalizedvgg.mat'):
# Only way to get the weight of the kernel of convolution
# Inspired by http://programtalk.com/vs2/python/2964/facenet/tmp/vggverydeep19.py/
kernels = vgg_layers[i][0][0][2][0][0]
bias = vgg_layers[i][0][0][2][0][1]
# matconvnet: weights are [width, height, in_channels, out_channels]
# tensorflow: weights are [height, width, in_channels, out_channels]
kernels = tf.constant(np.transpose(kernels, (1, 0, 2, 3)))
bias = tf.constant(bias.reshape(-1))
current = st.conv_layer(current, kernels, bias, name, padding)
# Update the variable named current to have the right size
elif (kind == 'relu'):
if (relu_kind == 'l'): # leakly relu
current = lrelu(current, leak=leak)
else:
current = tf.nn.relu(current, name=name)
elif (kind == 'pool'):
current = st.pool_layer(current, name, pooling_type, padding)
net[name] = current
print(name, current.get_shape())
assert len(net) == len(VGG19_LAYERS) + 1 # Test if the length is right
return (net, shapes)
def generator(inp_z, inp_y):
with tf.variable_scope('Generator'):
inp = tf.concat([inp_z, inp_y], 1)
g1 = linear(inp, 512 * 4 * 4, name=gname + 'deconv1')
g1 = batchnorm(g1, is_training=tf.constant(True), name=gname + 'bn1g')
g1 = lrelu(g1, 0.2)
g1_reshaped = tf.reshape(g1, [-1, 512, 4, 4])
print 'genreshape: ' + str(g1_reshaped.get_shape().as_list())
g2 = nnupsampling(g1_reshaped, [8, 8])
g2 = conv2d(g2, nout=256, kernel=3, name=gname + 'deconv2')
g2 = batchnorm(g2, is_training=tf.constant(True), name=gname + 'bn2g')
g2 = lrelu(g2, 0.2)
g3 = nnupsampling(g2, [16, 16])
g3 = conv2d(g3, nout=128, kernel=3, name=gname + 'deconv3')
g3 = batchnorm(g3, is_training=tf.constant(True), name=gname + 'bn3g')
g3 = lrelu(g3, 0.2)
g3b = conv2d(g3, nout=128, kernel=3, name=gname + 'deconv3b')
g3b = batchnorm(g3b,
is_training=tf.constant(True),
name=gname + 'bn3bg')
g3b = lrelu(g3b, 0.2)
g4 = nnupsampling(g3b, [32, 32])
g4 = conv2d(g4, nout=64, kernel=3, name=gname + 'deconv4')
g4 = batchnorm(g4, is_training=tf.constant(True), name=gname + 'bn4g')
g4 = lrelu(g4, 0.2)
g4b = conv2d(g4, nout=64, kernel=3, name=gname + 'deconv4b')
g4b = batchnorm(g4b,
is_training=tf.constant(True),
name=gname + 'bn4bg')
g4b = lrelu(g4b, 0.2)
g5 = nnupsampling(g4b, [64, 64])
g5 = conv2d(g5, nout=32, kernel=3, name=gname + 'deconv5')
g5 = batchnorm(g5, is_training=tf.constant(True), name=gname + 'bn5g')
g5 = lrelu(g5, 0.2)
g5b = conv2d(g5, nout=3, kernel=3, name=gname + 'deconv5b')
g5b = tf.nn.tanh(g5b)
g5b_32 = pool(g5b, fsize=3, strides=2, op='avg', pad='SAME')
return g5b_32, g5b
def discriminator(inp, reuse=False):
with tf.variable_scope('Encoder', reuse=reuse):
# 32
inp = gaussnoise(inp, std=0.05)
conv1 = conv2d(inp, 96, kernel=3, strides=1, name=dname + 'conv1')
conv1 = lrelu(conv1, 0.2)
conv1b = conv2d(conv1, 96, kernel=3, strides=2, name=dname + 'conv1b')
conv1b = batchnorm(conv1b, is_training=is_train, name=dname + 'bn1b')
conv1b = lrelu(conv1b, 0.2)
conv1b = tf.nn.dropout(conv1b, keep_prob)
# 16
conv2 = conv2d(conv1b, 192, kernel=3, strides=1, name=dname + 'conv2')
conv2 = batchnorm(conv2, is_training=is_train, name=dname + 'bn2')
conv2 = lrelu(conv2, 0.2)
conv2b = conv2d(conv2, 192, kernel=3, strides=2, name=dname + 'conv2b')
conv2b = batchnorm(conv2b, is_training=is_train, name=dname + 'bn2b')
conv2b = lrelu(conv2b, 0.2)
conv2b = tf.nn.dropout(conv2b, keep_prob)
# 8
conv3 = conv2d(conv2b, 256, kernel=3, strides=1, name=dname + 'conv3')
conv3 = batchnorm(conv3, is_training=is_train, name=dname + 'bn3')
conv3 = lrelu(conv3, 0.2)
conv3b = conv2d(conv3, 256, kernel=1, strides=1, name=dname + 'conv3b')
conv3b = batchnorm(conv3b, is_training=is_train, name=dname + 'bn3b')
conv3b = lrelu(conv3b, 0.2)
conv4 = conv2d(conv3b, 512, kernel=1, strides=1, name=dname + 'conv4')
conv4 = batchnorm(conv4, is_training=is_train, name=dname + 'bn4')
conv4 = lrelu(conv4, 0.2)
flat = flatten(conv4)
# Classifier
clspred = linear(flat, n_classes, name=dname + 'cpred')
# Decoder
g2 = conv2d(conv4, nout=256, kernel=3, name=dname + 'deconv2')
g2 = batchnorm(g2, is_training=tf.constant(True), name=dname + 'bn2g')
g2 = lrelu(g2, 0.2)
g3 = nnupsampling(g2, [16, 16])
g3 = conv2d(g3, nout=128, kernel=3, name=dname + 'deconv3')
g3 = batchnorm(g3, is_training=tf.constant(True), name=dname + 'bn3g')
g3 = lrelu(g3, 0.2)
g3b = conv2d(g3, nout=128, kernel=3, name=dname + 'deconv3b')
g3b = batchnorm(g3b,
is_training=tf.constant(True),
name=dname + 'bn3bg')
g3b = lrelu(g3b, 0.2)
g4 = nnupsampling(g3b, [32, 32])
g4 = conv2d(g4, nout=64, kernel=3, name=dname + 'deconv4')
g4 = batchnorm(g4, is_training=tf.constant(True), name=dname + 'bn4g')
g4 = lrelu(g4, 0.2)
g4b = conv2d(g4, nout=3, kernel=3, name=dname + 'deconv4b')
g4b = tf.nn.tanh(g4b)
return clspred, g4b
def autoencoder(input_shape=[None, 784],
n_filters=[1, 10, 10, 10],
filter_sizes=[3, 3, 3, 3],
corruption=False):
"""Build a deep denoising autoencoder w/ tied weights.
Parameters
----------
input_shape : list, optional
Description
n_filters : list, optional
Description
filter_sizes : list, optional
Description
Returns
-------
x : Tensor
Input placeholder to the network
z : Tensor
Inner-most latent representation
y : Tensor
Output reconstruction of the input
cost : Tensor
Overall cost to use for training
Raises
------
ValueError
Description
"""
# %%
# input to the network
x = tf.placeholder(
tf.float32, input_shape, name='x')
# %%
# Optionally apply denoising autoencoder
if corruption:
x_noise = corrupt(x)
else:
x_noise = x
# %%
# ensure 2-d is converted to square tensor.
if len(x.get_shape()) == 2:
x_dim = np.sqrt(x_noise.get_shape().as_list()[1])
if x_dim != int(x_dim):
raise ValueError('Unsupported input dimensions')
x_dim = int(x_dim)
x_tensor = tf.reshape(
x_noise, [-1, x_dim, x_dim, n_filters[0]])
elif len(x_noise.get_shape()) == 4:
x_tensor = x_noise
else:
raise ValueError('Unsupported input dimensions')
current_input = x_tensor
# %%
# Build the encoder
encoder = []
shapes = []
for layer_i, n_output in enumerate(n_filters[1:]):
n_input = current_input.get_shape().as_list()[3]
shapes.append(current_input.get_shape().as_list())
W = tf.Variable(
tf.random_uniform([
filter_sizes[layer_i],
filter_sizes[layer_i],
n_input, n_output],
-1.0 / math.sqrt(n_input),
1.0 / math.sqrt(n_input)))
b = tf.Variable(tf.zeros([n_output]))
encoder.append(W)
output = lrelu(
tf.add(tf.nn.conv2d(
current_input, W, strides=[1, 2, 2, 1], padding='SAME'), b))
current_input = output
# %%
# store the latent representation
z = current_input
encoder.reverse()
shapes.reverse()
# %%
# Build the decoder using the same weights
for layer_i, shape in enumerate(shapes):
W = encoder[layer_i]
b = tf.Variable(tf.zeros([W.get_shape().as_list()[2]]))
output = lrelu(tf.add(
tf.nn.deconv2d(
current_input, W,
tf.pack([tf.shape(x)[0], shape[1], shape[2], shape[3]]),
strides=[1, 2, 2, 1], padding='SAME'), b))
current_input = output
# %%
# now have the reconstruction through the network
y = current_input
# cost function measures pixel-wise difference
cost = tf.reduce_sum(tf.square(y - x_tensor))
# %%
return {'x': x, 'z': z, 'y': y, 'cost': cost}
from connections import conv2d, linear from datasets import MNIST # %% Setup input to the network and true output label. These are # simply placeholders which we'll fill in later. mnist = MNIST() x = tf.placeholder(tf.float32, [None, 784]) y = tf.placeholder(tf.float32, [None, 10]) x_tensor = tf.reshape(x, [-1, 28, 28, 1]) # %% Define the network: bn1 = batch_norm(-1, name='bn1') bn2 = batch_norm(-1, name='bn2') bn3 = batch_norm(-1, name='bn3') h_1 = lrelu(bn1(conv2d(x_tensor, 32, name='conv1')), name='lrelu1') h_2 = lrelu(bn2(conv2d(h_1, 64, name='conv2')), name='lrelu2') h_3 = lrelu(bn3(conv2d(h_2, 64, name='conv3')), name='lrelu3') h_3_flat = tf.reshape(h_3, [-1, 64 * 4 * 4]) h_4 = linear(h_3_flat, 10) y_pred = tf.nn.softmax(h_4) # %% Define loss/eval/training functions cross_entropy = -tf.reduce_sum(y * tf.log(y_pred)) train_step = tf.train.AdamOptimizer().minimize(cross_entropy) correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.argmax(y, 1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float')) # %% We now create a new session to actually perform the initialization the # variables:
def autoencoder(input_shape=[None, 784],
n_filters=[1, 10, 10, 10],
filter_sizes=[3, 3, 3, 3],
corruption=False):
"""Build a deep denoising autoencoder w/ tied weights.
Parameters
----------
input_shape : list, optional
Description
n_filters : list, optional
Description
filter_sizes : list, optional
Description
Returns
-------
x : Tensor
Input placeholder to the network
z : Tensor
Inner-most latent representation
y : Tensor
Output reconstruction of the input
cost : Tensor
Overall cost to use for training
Raises
------
ValueError
Description
"""
# %%
# input to the network
x = tf.placeholder(tf.float32, input_shape, name='x')
# %%
# Optionally apply denoising autoencoder
if corruption:
x_noise = corrupt(x)
else:
x_noise = x
# %%
# ensure 2-d is converted to square tensor.
if len(x.get_shape()) == 2:
x_dim = np.sqrt(x_noise.get_shape().as_list()[1])
if x_dim != int(x_dim):
raise ValueError('Unsupported input dimensions')
x_dim = int(x_dim)
x_tensor = tf.reshape(x_noise, [-1, x_dim, x_dim, n_filters[0]])
elif len(x_noise.get_shape()) == 4:
x_tensor = x_noise
else:
raise ValueError('Unsupported input dimensions')
current_input = x_tensor
# %%
# Build the encoder
encoder = []
shapes = []
for layer_i, n_output in enumerate(n_filters[1:]):
n_input = current_input.get_shape().as_list()[3]
shapes.append(current_input.get_shape().as_list())
W = tf.Variable(
tf.random_uniform([
filter_sizes[layer_i], filter_sizes[layer_i], n_input, n_output
], -1.0 / math.sqrt(n_input), 1.0 / math.sqrt(n_input)))
b = tf.Variable(tf.zeros([n_output]))
encoder.append(W)
output = lrelu(
tf.add(
tf.nn.conv2d(current_input,
W,
strides=[1, 2, 2, 1],
padding='SAME'), b))
current_input = output
# %%
# store the latent representation
z = current_input
encoder.reverse()
shapes.reverse()
# %%
# Build the decoder using the same weights
for layer_i, shape in enumerate(shapes):
W = encoder[layer_i]
b = tf.Variable(tf.zeros([W.get_shape().as_list()[2]]))
output = lrelu(
tf.add(
tf.nn.deconv2d(
current_input,
W,
tf.pack([tf.shape(x)[0], shape[1], shape[2], shape[3]]),
strides=[1, 2, 2, 1],
padding='SAME'), b))
current_input = output
# %%
# now have the reconstruction through the network
y = current_input
# cost function measures pixel-wise difference
cost = tf.reduce_sum(tf.square(y - x_tensor))
# %%
return {'x': x, 'z': z, 'y': y, 'cost': cost}
