def discrim(X, w, w2, g2, b2, w3, g3, b3, wy, by):
h = lrelu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2)))
h2 = lrelu(batchnorm(dnn_conv(h, w2, subsample=(2, 2), border_mode=(2, 2)), g=g2, b=b2))
h3 = lrelu(batchnorm(dnn_conv(h2, w3, subsample=(2, 2), border_mode=(2, 2)), g=g3, b=b3))
h3 = T.flatten(h3, 2)
y = -softplus(T.dot(h3, wy)+by)
return y
def discrim(X, w, w2, g2, b2, w3, g3, b3, w4, g4, b4, w5, g5, b5, w6, g6, b6,
wy):
h = lrelu(dnn_conv(X, w, subsample=(1, 1), border_mode=(1, 1)))
h2 = lrelu(
batchnorm(dnn_conv(h, w2, subsample=(2, 2), border_mode=(1, 1)),
g=g2,
b=b2))
h3 = lrelu(
batchnorm(dnn_conv(h2, w3, subsample=(1, 1), border_mode=(1, 1)),
g=g3,
b=b3))
h4 = lrelu(
batchnorm(dnn_conv(h3, w4, subsample=(2, 2), border_mode=(1, 1)),
g=g4,
b=b4))
h5 = lrelu(
batchnorm(dnn_conv(h4, w5, subsample=(1, 1), border_mode=(1, 1)),
g=g5,
b=b5))
h6 = lrelu(
batchnorm(dnn_conv(h5, w6, subsample=(2, 2), border_mode=(1, 1)),
g=g6,
b=b6))
h6 = T.flatten(h6, 2)
y = sigmoid(T.dot(h6, wy))
return y
def discrim( t, w1, w2, g2, b2, w3, g3, b3, w4, g4, b4, w5 ):
h1 = lrelu( dnn_conv( t, w1, subsample=( 2, 2 ), border_mode = ( 2, 2 ) ) )
h2 = lrelu( batchnorm( dnn_conv( h1, w2, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g2, b = b2 ) )
h3 = lrelu( batchnorm( dnn_conv( h2, w3, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g3, b = b3 ) )
h4 = lrelu( batchnorm( dnn_conv( h3, w4, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g4, b = b4 ) )
yd = sigmoid( T.dot( T.flatten( h4, 2 ), w5 ) )
return yd
def discrim(X, w, w2, w3, wy):
h = lrelu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2)))
h2 = lrelu(batchnorm(dnn_conv(h, w2, subsample=(2, 2), border_mode=(2, 2))))
h2 = T.flatten(h2, 2)
h3 = lrelu(batchnorm(T.dot(h2, w3)))
y = sigmoid(T.dot(h3, wy))
return y
def discrim(X, w, b, w2, g2, b2, w3, g3, b3, w4, g4, b4, wy, wy1):
h0 = dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2))
if args.db1:
h0 += b.dimshuffle('x', 0, 'x', 'x')
h1 = lrelu(h0)
h1 = dropout(h1, args.dropout)
h1 = dnn_conv(h1, w2, subsample=(2, 2), border_mode=(2, 2))
if args.dbn:
h1 = batchnorm(h1, g=g2, b=b2)
else:
h1 += b2.dimshuffle('x', 0, 'x', 'x')
h2 = lrelu(h1)
h2 = dropout(h2, args.dropout)
h2 = dnn_conv(h2, w3, subsample=(2, 2), border_mode=(2, 2))
if args.dbn:
h2 = batchnorm(h2, g=g3, b=b3)
else:
h2 += b3.dimshuffle('x', 0, 'x', 'x')
h3 = lrelu(h2)
h3 = dropout(h3, args.dropout)
h3 = dnn_conv(h3, w4, subsample=(2, 2), border_mode=(2, 2))
if args.dbn:
h3 = batchnorm(h3, g=g4, b=b4)
else:
h3 += b4.dimshuffle('x', 0, 'x', 'x')
h4 = lrelu(h3)
h4 = dropout(h4, args.dropout)
h4 = T.flatten(h4, 2)
y = sigmoid(T.dot(h4, wy))
y1 = sigmoid(T.dot(h4, wy1))
return y, y1
def encoder( s, w1, w2, g2, b2, w3, g3, b3, w4, g4, b4, w5, g5, b5 ):
h1 = lrelu( dnn_conv( s, w1, subsample=( 2, 2 ), border_mode = ( 2, 2 ) ) )
h2 = lrelu( batchnorm( dnn_conv( h1, w2, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g2, b = b2 ) )
h3 = lrelu( batchnorm( dnn_conv( h2, w3, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g3, b = b3 ) )
h4 = lrelu( batchnorm( dnn_conv( h3, w4, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g4, b = b4 ) )
z = lrelu( batchnorm( dnn_conv( h4, w5, subsample = ( 1, 1 ), border_mode = ( 0, 0 ) ), g = g5, b = b5 ) )
return T.flatten( z, 2 )
def denseConvlayer(layer_inputs, bottleneck_scale, growth_rate, is_training):
# Build the bottleneck operation
net = layer_inputs
net_temp = tf.identity(net)
net = batchnorm(net, is_training)
net = prelu_tf(net, name='Prelu_1')
net = conv2(net,
kernel=1,
output_channel=bottleneck_scale * growth_rate,
stride=1,
use_bias=False,
scope='conv1x1')
net = batchnorm(net, is_training)
net = prelu_tf(net, name='Prelu_2')
net = conv2(net,
kernel=3,
output_channel=growth_rate,
stride=1,
use_bias=False,
scope='conv3x3')
# Concatenate the processed feature to the feature
net = tf.concat([net_temp, net], axis=3)
return net
def discrim(X, w, b, w2, g2, b2, w3, g3, b3, w4, g4, b4, wy, wy1):
h0 = dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2))
if args.db1:
h0 += b.dimshuffle('x', 0, 'x', 'x')
h1 = lrelu(h0)
h1 = dropout(h1, args.dropout)
h1 = dnn_conv(h1, w2, subsample=(2, 2), border_mode=(2, 2))
if args.dbn:
h1 = batchnorm(h1, g=g2, b=b2)
else:
h1 += b2.dimshuffle('x', 0, 'x', 'x')
h2 = lrelu(h1)
h2 = dropout(h2, args.dropout)
h2 = dnn_conv(h2, w3, subsample=(2, 2), border_mode=(2, 2))
if args.dbn:
h2 = batchnorm(h2, g=g3, b=b3)
else:
h2 += b3.dimshuffle('x', 0, 'x', 'x')
h3 = lrelu(h2)
h3 = dropout(h3, args.dropout)
h3 = dnn_conv(h3, w4, subsample=(2, 2), border_mode=(2, 2))
if args.dbn:
h3 = batchnorm(h3, g=g4, b=b4)
else:
h3 += b4.dimshuffle('x', 0, 'x', 'x')
h4 = lrelu(h3)
h4 = dropout(h4, args.dropout)
h4 = T.flatten(h4, 2)
y = sigmoid(T.dot(h4, wy))
y1 = sigmoid(T.dot(h4, wy1))
return y, y1
def gen(Z, w, g, b, w2, g2, b2, w3, g3, b3,wx):
h0 = relu(batchnorm(T.dot(Z, w), g=g, b=b))
h0 = h0.reshape((h0.shape[0], ngf*4, 2, 2))
h1 = relu(batchnorm(deconv(h0, w2, subsample=(2, 2), border_mode=(2, 2)), g=g2, b=b2))
h2 = relu(batchnorm(deconv(h1, w3, subsample=(2, 2), border_mode=(2, 2)), g=g3, b=b3))
x = tanh(deconv(h2, wx, subsample=(2, 2), border_mode=(2, 2)))
return x
def apply(self, batch_size=None, rand_vals=None):
"""
Apply this generator module. Pass _either_ batch_size or rand_vals.
"""
assert not ((batch_size is None) and
(rand_vals is None)), "need either batch_size or rand_vals"
if rand_vals is None:
rand_shape = (batch_size, self.rand_dim)
if self.rand_type == 'normal':
rand_vals = self.rng.normal(size=rand_shape, avg=0.0, std=1.0, \
dtype=theano.config.floatX)
else:
rand_vals = self.rng.uniform(size=rand_shape, low=-1.0, high=1.0, \
dtype=theano.config.floatX)
else:
rand_shape = (rand_vals.shape[0], self.rand_dim)
rand_vals = rand_vals.reshape(rand_shape)
# transform random values into fc layer
h1 = T.dot(rand_vals, self.w1)
if self.apply_bn_1:
h1 = batchnorm(h1, g=self.g1, b=self.b1)
h1 = relu(h1)
# transform from fc layer to output
h2 = T.dot(h1, self.w2)
if self.apply_bn_2:
h2 = batchnorm(h2, g=self.g2, b=self.b2)
if self.final_relu:
h2 = relu(h2)
return h2
def gen_test_tanh(_z, _params, _pls, n_layers=3, n_f=128, init_sz=4):
tan_z = tanh(_z)
[gw0, gg0, gb0] = _params[0:3]
hs = []
u = _pls[0]
s = _pls[n_layers + 1]
h0 = relu(
batchnorm(T.dot(T.clip(tan_z, -1.0, 1.0), gw0), u=u, s=s, g=gg0,
b=gb0))
h1 = h0.reshape((h0.shape[0], n_f * 2**n_layers, init_sz, init_sz))
hs.extend([h0, h1])
for n in range(n_layers):
[w, g, b] = _params[3 * (n + 1):3 * (n + 2)]
hin = hs[-1]
u = _pls[n + 1]
s = _pls[n + n_layers + 2]
hout = relu(
batchnorm(deconv(hin, w, subsample=(2, 2), border_mode=(2, 2)),
u=u,
s=s,
g=g,
b=b))
hs.append(hout)
x = tanh(deconv(hs[-1], _params[-1], subsample=(2, 2), border_mode=(2, 2)))
return x
def discrim(X, w, w2, g2, b2, w3, g3, b3, wy, by):
h0 = dropout(relu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2))), p=0.5)
h1 = dropout(relu(batchnorm(dnn_conv(h0, w2, subsample=(2, 2), border_mode=(2, 2)), g=g2, b=b2)), p=0.5)
h2 = dropout(relu(batchnorm(dnn_conv(h1, w3, subsample=(2, 2), border_mode=(2, 2)), g=g3, b=b3)), p=0.5)
h2 = T.flatten(h2, 2)
y = -relu(T.dot(h2, wy)+by)
return y
def apply(self, batch_size=None, rand_vals=None):
"""
Apply this generator module. Pass _either_ batch_size or rand_vals.
"""
assert not ((batch_size is None) and (rand_vals is None)), "need either batch_size or rand_vals"
if rand_vals is None:
rand_shape = (batch_size, self.rand_dim)
if self.rand_type == 'normal':
rand_vals = self.rng.normal(size=rand_shape, avg=0.0, std=1.0, \
dtype=theano.config.floatX)
else:
rand_vals = self.rng.uniform(size=rand_shape, low=-1.0, high=1.0, \
dtype=theano.config.floatX)
else:
rand_shape = (rand_vals.shape[0], self.rand_dim)
rand_vals = rand_vals.reshape(rand_shape)
# transform random values into fc layer
h1 = T.dot(rand_vals, self.w1)
if self.apply_bn_1:
h1 = batchnorm(h1, g=self.g1, b=self.b1)
h1 = relu(h1)
# transform from fc layer to output
h2 = T.dot(h1, self.w2)
if self.apply_bn_2:
h2 = batchnorm(h2, g=self.g2, b=self.b2)
if self.final_relu:
h2 = relu(h2)
return h2
def apply(self, input):
"""
Apply this discriminator module to the given input. This produces a
collection of filter responses for feedforward and a spatial grid of
discriminator outputs.
"""
bm = int((self.filt_dim - 1) / 2) # use "same" mode convolutions
ss = self.ds_stride # stride for "learned downsampling"
# apply first conv layer
h1 = dnn_conv(input, self.w1, subsample=(1, 1), border_mode=(bm, bm))
if self.apply_bn_1:
h1 = batchnorm(h1, g=self.g1, b=self.b1)
h1 = lrelu(h1)
# apply second conv layer (may include downsampling)
if self.use_pooling:
h2 = dnn_conv(h1, self.w2, subsample=(1, 1), border_mode=(bm, bm))
if self.apply_bn_2:
h2 = batchnorm(h2, g=self.g2, b=self.b2)
h2 = lrelu(h2)
h2 = dnn_pool(h2, (ss,ss), stride=(ss, ss), mode='max', pad=(0, 0))
else:
h2 = dnn_conv(h1, self.w2, subsample=(ss, ss), border_mode=(bm, bm))
if self.apply_bn_2:
h2 = batchnorm(h2, g=self.g2, b=self.b2)
h2 = lrelu(h2)
# apply discriminator layer
y = dnn_conv(h2, self.wd, subsample=(1, 1), border_mode=(bm, bm))
y = sigmoid(T.flatten(y, 2)) # flatten to (batch_size, num_preds)
return h2, y
def gen(Z, w1, g1, b1, w2, g2, b2, w3, g3, b3, w4, g4, b4, wx):
Gl1 = relu(batchnorm(T.dot(Z, w1), g=g1, b=b1))
Gl1 = Gl1.reshape((Gl1.shape[0], Channel[-1], Convlayersize[-1],
Convlayersize[-1], Convlayersize[-1]))
input_shape = (None, None, Convlayersize[-1], Convlayersize[-1],
Convlayersize[-1])
filter_shape = (Channel[-1], Channel[-2], kernal[-1], kernal[-1],
kernal[-1])
Gl2 = relu(
batchnorm(conv(Gl1,
w2,
filter_shape=filter_shape,
input_shape=input_shape,
conv_mode='deconv'),
g=g2,
b=b2))
input_shape = (None, None, Convlayersize[-2], Convlayersize[-2],
Convlayersize[-2])
filter_shape = (Channel[-2], Channel[-3], kernal[-2], kernal[-2],
kernal[-2])
Gl3 = relu(
batchnorm(conv(Gl2,
w3,
filter_shape=filter_shape,
input_shape=input_shape,
conv_mode='deconv'),
g=g3,
b=b3))
input_shape = (None, None, Convlayersize[-3], Convlayersize[-3],
Convlayersize[-3])
filter_shape = (Channel[-3], Channel[-4], kernal[-3], kernal[-3],
kernal[-3])
Gl4 = relu(
batchnorm(conv(Gl3,
w4,
filter_shape=filter_shape,
input_shape=input_shape,
conv_mode='deconv'),
g=g4,
b=b4))
input_shape = (None, None, Convlayersize[-4], Convlayersize[-4],
Convlayersize[-4])
filter_shape = (Channel[-4], Channel[-5], kernal[-4], kernal[-4],
kernal[-4])
GlX = sigmoid(
conv(Gl4,
wx,
filter_shape=filter_shape,
input_shape=input_shape,
conv_mode='deconv'))
return GlX
def gen(Z, w, g, b, w2, g2, b2, w3, g3, b3, w4, g4, b4, wx):
h = relu(batchnorm(T.dot(Z, w), g=g, b=b))
h = h.reshape((h.shape[0], ngf*8, 4, 4))
h2 = relu(batchnorm(deconv(h, w2, subsample=(2, 2), border_mode=(2, 2)), g=g2, b=b2))
h3 = relu(batchnorm(deconv(h2, w3, subsample=(2, 2), border_mode=(2, 2)), g=g3, b=b3))
h4 = relu(batchnorm(deconv(h3, w4, subsample=(2, 2), border_mode=(2, 2)), g=g4, b=b4))
x = tanh(deconv(h4, wx, subsample=(2, 2), border_mode=(2, 2)))
return x
def domain_discrim( st, w1, w2, g2, b2, w3, g3, b3, w4, g4, b4, w5, g5, b5, w6 ):
h1 = lrelu( dnn_conv( st, w1, subsample=( 2, 2 ), border_mode = ( 2, 2 ) ) )
h2 = lrelu( batchnorm( dnn_conv( h1, w2, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g2, b = b2 ) )
h3 = lrelu( batchnorm( dnn_conv( h2, w3, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g3, b = b3 ) )
h4 = lrelu( batchnorm( dnn_conv( h3, w4, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g4, b = b4 ) )
h5 = lrelu( batchnorm( dnn_conv( h4, w5, subsample = ( 1, 1 ), border_mode = ( 0, 0 ) ), g = g5, b = b5 ) )
ydd = sigmoid( T.dot( T.flatten( h5, 2 ), w6 ) )
return ydd
def feature_function(input_data, is_train=True):
h0 = relu(batchnorm(X=dnn_conv(input_data, conv_w0, subsample=(2, 2), border_mode=(2, 2)), g=bn_w0, b=bn_b0))
h1 = relu(batchnorm(X=dnn_conv(h0, conv_w1, subsample=(2, 2), border_mode=(2, 2)), g=bn_w1, b=bn_b1))
h2 = relu(batchnorm(X=dnn_conv(h1, conv_w2, subsample=(2, 2), border_mode=(2, 2)), g=bn_w2, b=bn_b2))
h3 = relu(batchnorm(X=dnn_conv(h2, conv_w3, subsample=(2, 2), border_mode=(2, 2)), g=bn_w3, b=bn_b3))
h3 = T.flatten(h3, 2)
f = tanh(T.dot(h3, linear_w4)+linear_b4)
return f
def gen(Z, w, w2, w3, gwx):
h = relu(batchnorm(T.dot(Z, w)))
h2 = relu(batchnorm(T.dot(h, w2)))
h2 = h2.reshape((h2.shape[0], ngf * 2, 7, 7))
h3 = relu(batchnorm(deconv(h2, w3, subsample=(2, 2), border_mode=(2, 2))))
x = sigmoid(deconv(h3, gwx, subsample=(2, 2), border_mode=(2, 2)))
return x
def discrim(X, w, w2, g2, b2, w3, g3, b3, w4, g4, b4, wy):
h = lrelu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2)))
h2 = lrelu(batchnorm(dnn_conv(h, w2, subsample=(2, 2), border_mode=(2, 2)), g=g2, b=b2))
h3 = lrelu(batchnorm(dnn_conv(h2, w3, subsample=(2, 2), border_mode=(2, 2)), g=g3, b=b3))
h4 = lrelu(batchnorm(dnn_conv(h3, w4, subsample=(2, 2), border_mode=(2, 2)), g=g4, b=b4))
h4 = T.flatten(h4, 2)
y = sigmoid(T.dot(h4, wy))
return y
def gen(Z, w, g, b, w2, g2, b2, w3, g3, b3, w4, g4, b4, wx):
h = relu(batchnorm(T.dot(Z, w), g=g, b=b))
h = h.reshape((h.shape[0], ngf*8, 4, 4))
h2 = relu(batchnorm(deconv(h, w2, subsample=(2, 2), border_mode=(2, 2)), g=g2, b=b2))
h3 = relu(batchnorm(deconv(h2, w3, subsample=(2, 2), border_mode=(2, 2)), g=g3, b=b3))
h4 = relu(batchnorm(deconv(h3, w4, subsample=(2, 2), border_mode=(2, 2)), g=g4, b=b4))
x = tanh(deconv(h4, wx, subsample=(2, 2), border_mode=(2, 2)))
return x
def discrim(X, w, w2, g2, b2, w3, g3, b3, wy, wa):
h0 = dropout(relu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2))), p=0.5)
h1 = dropout(relu(batchnorm(dnn_conv(h0, w2, subsample=(2, 2), border_mode=(2, 2)), g=g2, b=b2)), p=0.5)
h2 = dropout(relu(batchnorm(dnn_conv(h1, w3, subsample=(2, 2), border_mode=(2, 2)), g=g3, b=b3)), p=0.5)
h2 = T.flatten(h2, 2)
y = square(T.dot(h2, wy))
y = T.dot(T.log(1+y), T.exp(wa))
return y
def decoder( z, w1, g1, b1, w2, g2, b2, w3, g3, b3, w4, g4, b4, w5 ):
h1 = relu( batchnorm( T.dot( z, w1 ), g = g1, b = b1 ) )
h1 = h1.reshape( (h1.shape[ 0 ], nf * 8, 4, 4 ) )
h2 = relu( batchnorm( deconv( h1, w2, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g2, b = b2 ) )
h3 = relu( batchnorm( deconv( h2, w3, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g3, b = b3 ) )
h4 = relu( batchnorm( deconv( h3, w4, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g4, b = b4 ) )
t = tanh( deconv( h4, w5, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ) )
return t
def discrim(X, w, w2, g2, b2, w3, g3, b3, w4, g4, b4, wy):
h = lrelu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2)))
h2 = lrelu(batchnorm(dnn_conv(h, w2, subsample=(2, 2), border_mode=(2, 2)), g=g2, b=b2))
h3 = lrelu(batchnorm(dnn_conv(h2, w3, subsample=(2, 2), border_mode=(2, 2)), g=g3, b=b3))
h4 = lrelu(batchnorm(dnn_conv(h3, w4, subsample=(2, 2), border_mode=(2, 2)), g=g4, b=b4))
h4 = T.flatten(h4, 2)
y = sigmoid(T.dot(h4, wy))
return y
def generator_function(hidden_data, is_train=True):
h0_0 = relu(batchnorm(X=T.dot(hidden_data, linear_w0_0), g=bn_w0_0, b=bn_b0_0))
h0_1 = relu(batchnorm(X=T.dot(h0_0, linear_w0_1), g=bn_w0_1, b=bn_b0_1))
h0 = h0_1.reshape((h0_1.shape[0], num_gen_filters0, init_image_size, init_image_size))
h1 = relu(batchnorm(deconv(h0, conv_w1, subsample=(2, 2), border_mode=(2, 2)), g=bn_w1, b=bn_b1))
h2 = relu(batchnorm(deconv(h1, conv_w2, subsample=(2, 2), border_mode=(2, 2)), g=bn_w2, b=bn_b2))
h3 = relu(batchnorm(deconv(h2, conv_w3, subsample=(2, 2), border_mode=(2, 2)), g=bn_w3, b=bn_b3))
output = tanh(deconv(h3, conv_w4, subsample=(2, 2), border_mode=(2, 2))+conv_b4.dimshuffle('x', 0, 'x', 'x'))
return output
def converter( Z, w5d, g5d, b5d, w4d, g4d, b4d, w3d, g3d, b3d, w2d, g2d, b2d, w1d ):
h5d = relu( batchnorm( dnn_conv( Z, w5d, subsample = ( 1, 1 ), border_mode = ( 0, 0 ) ), g = g5d, b = b5d ) )
h5d = h5d.reshape( ( h5d.shape[ 0 ], nf * 8, 4, 4 ) )
h4d = relu( batchnorm( deconv( h5d, w4d, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g4d, b = b4d ) )
h3d = relu( batchnorm( deconv( h4d, w3d, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g3d, b = b3d ) )
h2d = relu( batchnorm( deconv( h3d, w2d, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g2d, b = b2d ) )
h1d = tanh( deconv( h2d, w1d, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ) )
y = h1d
return y
def discrim(X, w, w2, g2, b2, w3, g3, b3, w4, g4, b4, w5, g5, b5, w6, g6, b6, wy):
h = lrelu(dnn_conv(X, w, subsample=(1, 1), border_mode=(1, 1)))
h2 = lrelu(batchnorm(dnn_conv(h, w2, subsample=(2, 2), border_mode=(1, 1)), g=g2, b=b2))
h3 = lrelu(batchnorm(dnn_conv(h2, w3, subsample=(1, 1), border_mode=(1, 1)), g=g3, b=b3))
h4 = lrelu(batchnorm(dnn_conv(h3, w4, subsample=(2, 2), border_mode=(1, 1)), g=g4, b=b4))
h5 = lrelu(batchnorm(dnn_conv(h4, w5, subsample=(1, 1), border_mode=(1, 1)), g=g5, b=b5))
h6 = lrelu(batchnorm(dnn_conv(h5, w6, subsample=(2, 2), border_mode=(1, 1)), g=g6, b=b6))
h6 = T.flatten(h6, 2)
y = sigmoid(T.dot(h6, wy))
return y
def discrim(X):
current_input = dropout(X, 0.3)
### encoder ###
cv1 = relu(
dnn_conv(current_input, aew1, subsample=(1, 1), border_mode=(1, 1)))
cv2 = relu(
batchnorm(dnn_conv(cv1, aew2, subsample=(4, 4), border_mode=(2, 2)),
g=aeg2,
b=aeb2))
cv3 = relu(
batchnorm(dnn_conv(cv2, aew3, subsample=(1, 1), border_mode=(1, 1)),
g=aeg3,
b=aeb3))
cv4 = relu(
batchnorm(dnn_conv(cv3, aew4, subsample=(4, 4), border_mode=(2, 2)),
g=aeg4,
b=aeb4))
cv5 = relu(
batchnorm(dnn_conv(cv4, aew5, subsample=(1, 1), border_mode=(1, 1)),
g=aeg5,
b=aeb5))
cv6 = relu(
batchnorm(dnn_conv(cv5, aew6, subsample=(4, 4), border_mode=(0, 0)),
g=aeg6,
b=aeb6))
### decoder ###
dv6 = relu(
batchnorm(deconv(cv6, aew6, subsample=(4, 4), border_mode=(0, 0)),
g=aeg6t,
b=aeb6t))
dv5 = relu(
batchnorm(deconv(dv6, aew5, subsample=(1, 1), border_mode=(1, 1)),
g=aeg5t,
b=aeb5t))
dv4 = relu(
batchnorm(deconv(dv5, aew4, subsample=(4, 4), border_mode=(2, 2)),
g=aeg4t,
b=aeb4t))
dv3 = relu(
batchnorm(deconv(dv4, aew3, subsample=(1, 1), border_mode=(1, 1)),
g=aeg3t,
b=aeb3t))
dv2 = relu(
batchnorm(deconv(dv3, aew2, subsample=(4, 4), border_mode=(2, 2)),
g=aeg2t,
b=aeb2t))
dv1 = tanh(deconv(dv2, aew1, subsample=(1, 1), border_mode=(1, 1)))
rX = dv1
mse = T.sqrt(T.sum(T.abs_(T.flatten(X - rX, 2)), axis=1)) + T.sqrt(
T.sum(T.flatten((X - rX)**2, 2), axis=1)) # L1 and L2 loss
return T.flatten(cv6, 2), rX, mse
def generator_function(hidden_data, is_train=True):
# layer 0 (linear)
h0 = relu(batchnorm(X=T.dot(hidden_data, linear_w0), g=linear_bn_w0, b=linear_bn_b0))
h0 = h0.reshape((h0.shape[0], num_gen_filters0, init_image_size, init_image_size))
# layer 1 (deconv)
h1 = relu(batchnorm(deconv(h0, conv_w1, subsample=(2, 2), border_mode=(2, 2)), g=conv_bn_w1, b=conv_bn_b1))
# layer 2 (deconv)
h2 = relu(batchnorm(deconv(h1, conv_w2, subsample=(2, 2), border_mode=(2, 2)), g=conv_bn_w2, b=conv_bn_b2))
# layer 3 (deconv)
output = tanh(deconv(h2, conv_w3, subsample=(2, 2), border_mode=(2, 2))+conv_b3.dimshuffle('x', 0, 'x', 'x'))
return output
def gen(Z, Y, w, w2, w3, wx):
yb = Y.dimshuffle(0, 1, 'x', 'x')
Z = T.concatenate([Z, Y], axis=1)
h = relu(batchnorm(T.dot(Z, w)))
h = T.concatenate([h, Y], axis=1)
h2 = relu(batchnorm(T.dot(h, w2)))
h2 = h2.reshape((h2.shape[0], ngf*2, npx_, npx_))
h2 = conv_cond_concat(h2, yb)
h3 = relu(batchnorm(deconv(h2, w3, subsample=(2, 2), border_mode=(2, 2))))
h3 = conv_cond_concat(h3, yb)
x = sigmoid(deconv(h3, wx, subsample=(2, 2), border_mode=(2, 2)))
return x
def discrim(X, Y, w, w2, w3, wy):
yb = Y.dimshuffle(0, 1, 'x', 'x')
X = conv_cond_concat(X, yb)
h = lrelu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2)))
h = conv_cond_concat(h, yb)
h2 = lrelu(batchnorm(dnn_conv(h, w2, subsample=(2, 2), border_mode=(2, 2))))
h2 = T.flatten(h2, 2)
h2 = T.concatenate([h2, Y], axis=1)
h3 = lrelu(batchnorm(T.dot(h2, w3)))
h3 = T.concatenate([h3, Y], axis=1)
y = sigmoid(T.dot(h3, wy))
return y
def gen(Z, Y, w, w2, w3, wx):
yb = Y.dimshuffle(0, 1, 'x', 'x')
Z = T.concatenate([Z, Y], axis=1)
h = relu(batchnorm(T.dot(Z, w)))
h = T.concatenate([h, Y], axis=1)
h2 = relu(batchnorm(T.dot(h, w2)))
h2 = h2.reshape((h2.shape[0], ngf * 2, temp, temp))
h2 = conv_cond_concat(h2, yb)
h3 = relu(batchnorm(deconv(h2, w3, subsample=(2, 2), border_mode=(2, 2))))
h3 = conv_cond_concat(h3, yb)
x = sigmoid(deconv(h3, wx, subsample=(2, 2), border_mode=(2, 2)))
return x
def model(X,
h2_u, h3_u,
h2_s, h3_s,
w, w2, g2, b2, w3, g3, b3, wy
):
h = lrelu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2)))
h2 = lrelu(batchnorm(dnn_conv(h, w2, subsample=(2, 2), border_mode=(2, 2)), g=g2, b=b2, u=h2_u, s=h2_s))
h3 = lrelu(batchnorm(dnn_conv(h2, w3, subsample=(2, 2), border_mode=(2, 2)), g=g3, b=b3, u=h3_u, s=h3_s))
h = T.flatten(dnn_pool(h, (4, 4), (4, 4), mode='max'), 2)
h2 = T.flatten(dnn_pool(h2, (2, 2), (2, 2), mode='max'), 2)
h3 = T.flatten(dnn_pool(h3, (1, 1), (1, 1), mode='max'), 2)
f = T.concatenate([h, h2, h3], axis=1)
return [f]
def model(X,
h2_u, h3_u,
h2_s, h3_s,
w, w2, g2, b2, w3, g3, b3, wy
):
h = lrelu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2)))
h2 = lrelu(batchnorm(dnn_conv(h, w2, subsample=(2, 2), border_mode=(2, 2)), g=g2, b=b2, u=h2_u, s=h2_s))
h3 = lrelu(batchnorm(dnn_conv(h2, w3, subsample=(2, 2), border_mode=(2, 2)), g=g3, b=b3, u=h3_u, s=h3_s))
h = T.flatten(dnn_pool(h, ws=(4, 4), stride=(4, 4), mode='max'), 2)
h2 = T.flatten(dnn_pool(h2, ws=(2, 2), stride=(2, 2), mode='max'), 2)
h3 = T.flatten(dnn_pool(h3, ws=(1, 1), stride=(1, 1), mode='max'), 2)
f = T.concatenate([h, h2, h3], axis=1)
return [f]
def discrim(X, Y, w, w2, w3, wy):
yb = Y.dimshuffle(0, 1, 'x', 'x')
X = conv_cond_concat(X, yb)
h = lrelu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2)))
h = conv_cond_concat(h, yb)
h2 = lrelu(batchnorm(dnn_conv(h, w2, subsample=(2, 2),
border_mode=(2, 2))))
h2 = T.flatten(h2, 2)
h2 = T.concatenate([h2, Y], axis=1)
h3 = lrelu(batchnorm(T.dot(h2, w3)))
h3 = T.concatenate([h3, Y], axis=1)
y = sigmoid(T.dot(h3, wy))
return y
def gen(Z, Y, w, w2, w3, wx):
#Z: (nbatch, nz) = (128, 100)
#Y: (nbatch, ny) = (128, 10)
#w: (nz+ny, ngfc) = (110, 1024)
#w2: (ngfc+ny, ngf*2*7*7) = (1024+10, 64*2*7*7) = (1034, 6272)
#w3: (ngf*2+ny, ngf, 5, 5) = (128+10, 64, 5, 5 ) = (138, 64, 5, 5)
#wx: (ngf+ny, nc, 5, 5) = (64+10, 1, 5, 5) = (74, 1, 5, 5)
print '\n@@@@ gen()'
printVal('Y', Y)
printVal('w', w) #matrix
printVal('w2', w2) #matrix
printVal('w3', w3) #tensor
printVal('wx', wx) #tensor
# Yの要素の並びの入れ替え。数字の引数は、次元番号。'x' は ブロードキャスト
#(G1)
yb = Y.dimshuffle(0, 1, 'x', 'x')
# yb は４次元テンソル
printVal('yb', yb)
# 行列 Z と Y を結合（横方向）:いわゆる Conditional GAN の形にする。
#(G2)
Z = T.concatenate([Z, Y], axis=1) # Z: (128, 110)
#(G3)
# Z*w をバッチ正規化して、ReLU 適用
t1 = T.dot(Z, w) #full connect : t1: (128, 1024)
printVal('t1', t1)
h = relu(batchnorm(t1))
# h: (128, 1024)
#(G4)
h = T.concatenate([h, Y], axis=1) # h: (128, 1034)
#(G5)
h2 = relu(batchnorm(T.dot(h, w2) #NOT full connect
))
#(G6)
h2 = h2.reshape((h2.shape[0], ngf * 2, 7, 7))
#(G7)
h3, yb2 = conv_cond_concat2(h2, yb) #XXX
#(G8)デコンボリューション:論文によれば、空間プーリングの代わりに適用する
d = deconv(h3, w3, subsample=(2, 2), border_mode=(2, 2))
printVal('d', d) # (128, 64, 14, 14)
#h3 = relu(batchnorm(deconv(h2, w3, subsample=(2, 2), border_mode=(2, 2))))
#(G9)
h4 = relu(batchnorm(d))
#(G10)
h5, yb3 = conv_cond_concat2(h4, yb)
#(G11)
x = sigmoid(deconv(h5, wx, subsample=(2, 2), border_mode=(2, 2)))
return x, yb, yb2, d, h3, h5
def bnorm_statistics(X, w, w2, g2, b2, w3, g3, b3, wy):
h = lrelu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2)))
h2 = dnn_conv(h, w2, subsample=(2, 2), border_mode=(2, 2))
h2_u, h2_s = mean_and_var(h2)
h2 = lrelu(batchnorm(h2, g=g2, b=b2))
h3 = dnn_conv(h2, w3, subsample=(2, 2), border_mode=(2, 2))
h3_u, h3_s = mean_and_var(h3)
h3 = lrelu(batchnorm(h3, g=g3, b=b3))
h_us = [h2_u, h3_u]
h_ss = [h2_s, h3_s]
return h_us, h_ss
def bnorm_statistics(X, w, w2, g2, b2, w3, g3, b3, wy):
h = lrelu(dnn_conv(X, w, subsample=(2, 2), border_mode=(2, 2)))
h2 = dnn_conv(h, w2, subsample=(2, 2), border_mode=(2, 2))
h2_u, h2_s = mean_and_var(h2)
h2 = lrelu(batchnorm(h2, g=g2, b=b2))
h3 = dnn_conv(h2, w3, subsample=(2, 2), border_mode=(2, 2))
h3_u, h3_s = mean_and_var(h3)
h3 = lrelu(batchnorm(h3, g=g3, b=b3))
h_us = [h2_u, h3_u]
h_ss = [h2_s, h3_s]
return h_us, h_ss
def apply(self, input, rand_vals=None):
"""
Apply this generator module to some input.
"""
batch_size = input.shape[0]
bm = int((self.filt_dim - 1) / 2) # use "same" mode convolutions
ss = self.us_stride # stride for "learned upsampling"
if self.use_pooling:
# "unpool" the input if desired
input = input.repeat(ss, axis=2).repeat(ss, axis=3)
# get shape for random values that will augment input
rand_shape = (batch_size, self.rand_chans, input.shape[2],
input.shape[3])
if self.use_rand:
# augment input with random channels
if rand_vals is None:
if self.rand_type == 'normal':
rand_vals = self.rng.normal(size=rand_shape, avg=0.0, std=1.0, \
dtype=theano.config.floatX)
else:
rand_vals = self.rng.uniform(size=rand_shape, low=-1.0, high=1.0, \
dtype=theano.config.floatX)
rand_vals = rand_vals.reshape(rand_shape)
# stack random values on top of input
full_input = T.concatenate([rand_vals, input], axis=1)
else:
# don't augment input with random channels
full_input = input
# apply first convolution, perhaps with fractional striding
if self.use_pooling:
h1 = dnn_conv(full_input,
self.w1,
subsample=(1, 1),
border_mode=(bm, bm))
else:
# apply first conv layer (with fractional stride for upsampling)
h1 = deconv(full_input,
self.w1,
subsample=(ss, ss),
border_mode=(bm, bm))
if self.apply_bn_1:
h1 = batchnorm(h1, g=self.g1, b=self.b1)
h1 = relu(h1)
# apply second conv layer
h2 = dnn_conv(h1, self.w2, subsample=(1, 1), border_mode=(bm, bm))
if self.apply_bn_2:
h2 = batchnorm(h2, g=self.g2, b=self.b2)
h2 = relu(h2)
return h2
def make_conv_layer(X,
input_size,
output_size,
input_filters,
output_filters,
name,
index,
weights=None,
filter_sz=5):
is_deconv = output_size >= input_size
w_size = (input_filters, output_filters, filter_sz,
filter_sz) if is_deconv else (output_filters, input_filters,
filter_sz, filter_sz)
if weights is None:
w = gifn(w_size, '%sw%i' % (name, index))
g = gain_ifn((output_filters), '%sg%i' % (name, index))
b = bias_ifn((output_filters), '%sb%i' % (name, index))
else:
w, g, b = weights
conv_method = deconv if is_deconv else dnn_conv
activation = relu if is_deconv else lrelu
sub = output_size / input_size if is_deconv else input_size / output_size
if filter_sz == 3:
bm = 1
else:
bm = 2
layer = activation(
batchnorm(conv_method(X, w, subsample=(sub, sub),
border_mode=(bm, bm)),
g=g,
b=b))
return layer, [w, g, b]
def apply(self, batch_size=None, rand_vals=None):
"""
Apply this generator module. Pass _either_ batch_size or rand_vals.
"""
assert not ((batch_size is None) and
(rand_vals is None)), "need either batch_size or rand_vals"
if rand_vals is None:
rand_shape = (batch_size, self.rand_dim)
if self.rand_type == 'normal':
rand_vals = self.rng.normal(size=rand_shape, avg=0.0, std=1.0, \
dtype=theano.config.floatX)
else:
rand_vals = self.rng.uniform(size=rand_shape, low=-1.0, high=1.0, \
dtype=theano.config.floatX)
else:
rand_shape = (rand_vals.shape[0], self.rand_dim)
rand_vals = rand_vals.reshape(rand_shape)
# transform random values linearly
h1 = T.dot(rand_vals, self.w1)
if self.apply_bn:
h1 = batchnorm(h1, g=self.g1, b=self.b1)
if self.final_relu:
h1 = relu(h1)
return h1
##############
# EYE BUFFER #
##############
def encoder_feature_function(input_data):
# layer 0 (conv)
h0 = dnn_conv(input_data, conv_w0, subsample=(2, 2), border_mode=(2, 2))
h0 = relu(batchnorm(h0, g=bn_w0, b=bn_b0))
# layer 1 (conv)
h1 = dnn_conv( h0, conv_w1, subsample=(2, 2), border_mode=(2, 2))
h1 = relu(batchnorm(h1, g=bn_w1, b=bn_b1))
# layer 2 (conv)
h2 = dnn_conv( h1, conv_w2, subsample=(2, 2), border_mode=(2, 2))
h2 = relu(batchnorm(h2, g=bn_w2, b=bn_b2))
# layer 3 (conv)
h3 = dnn_conv( h2, conv_w3, subsample=(2, 2), border_mode=(2, 2))
h3 = relu(batchnorm(h3, g=bn_w3, b=bn_b3))
# feature
feature = T.flatten(h3, 2)
return feature
def make_conv_layer(X, input_size, output_size, input_filters,
output_filters, name, index,
weights = None, filter_sz = 5):
is_deconv = output_size >= input_size
w_size = (input_filters, output_filters, filter_sz, filter_sz) \
if is_deconv else (output_filters, input_filters, filter_sz, filter_sz)
if weights is None:
w = gifn(w_size, '%sw%i' %(name, index))
g = gain_ifn((output_filters), '%sg%i' %(name, index))
b = bias_ifn((output_filters), '%sb%i' %(name, index))
else:
w,g,b = weights
conv_method = deconv if is_deconv else dnn_conv
activation = relu if is_deconv else lrelu
sub = output_size / input_size if is_deconv else input_size / output_size
if filter_sz == 3:
bm = 1
else:
bm = 2
layer = activation(batchnorm(conv_method(X, w, subsample=(sub, sub), border_mode=(bm, bm)), g=g, b=b))
return layer, [w,g,b]
def encoder_function(input_data):
# layer 0 (conv)
h0 = dnn_conv(input_data, conv_w0, subsample=(2, 2), border_mode=(2, 2))
h0 = relu(batchnorm(h0, g=bn_w0, b=bn_b0))
# layer 1 (conv)
h1 = dnn_conv(h0, conv_w1, subsample=(2, 2), border_mode=(2, 2))
h1 = relu(batchnorm(h1, g=bn_w1, b=bn_b1))
# layer 2 (conv)
h2 = dnn_conv(h1, conv_w2, subsample=(2, 2), border_mode=(2, 2))
h2 = relu(batchnorm(h2, g=bn_w2, b=bn_b2))
# layer 3 (conv)
h3 = dnn_conv(h2, conv_w3, subsample=(2, 2), border_mode=(2, 2))
h3 = T.flatten(relu(batchnorm(h3, g=bn_w3, b=bn_b3)), 2)
# layer output
hidden_data = T.dot(h3, hidden_w) + hidden_b
return hidden_data
def gen_postlearn(_z, _params, n_layers=3, n_f=128, init_sz=4):
output = []
[gw0, gg0, gb0] = _params[0:3]
hs = []
h0_o = T.dot(_z, gw0)
output = [h0_o]
h0 = relu(batchnorm(h0_o, g=gg0, b=gb0))
h1 = h0.reshape((h0.shape[0], n_f * 2**n_layers, init_sz, init_sz))
hs.extend([h0, h1])
for n in range(n_layers):
[w, g, b] = _params[3 * (n + 1):3 * (n + 2)]
hin = hs[-1]
h_o = deconv(hin, w, subsample=(2, 2), border_mode=(2, 2))
hout = relu(batchnorm(h_o, g=g, b=b))
hs.append(hout)
output.append(h_o)
x = tanh(deconv(hs[-1], _params[-1], subsample=(2, 2), border_mode=(2, 2)))
return x, output
def gen(Z, Y):
yb = Y.dimshuffle(0, 1, 'x', 'x')
Z = T.concatenate([Z, Y], axis=1)
h = relu(batchnorm(T.dot(Z, gw), g=gg, b=gb))
h = h.reshape((h.shape[0], ngf * 4, 4, 4))
h = conv_cond_concat(h, yb)
h2 = relu(
batchnorm(deconv(h, gw2, subsample=(2, 2), border_mode=(2, 2)),
g=gg2,
b=gb2))
h2 = conv_cond_concat(h2, yb)
h3 = relu(
batchnorm(deconv(h2, gw3, subsample=(2, 2), border_mode=(2, 2)),
g=gg3,
b=gb3))
h3 = conv_cond_concat(h3, yb)
x = tanh(deconv(h3, gw4, subsample=(2, 2), border_mode=(2, 2)))
return x
def encoder(X, w1, g1, b1, w2, g2, b2, w3, g3, b3, w4, g4, b4, wz):
filter_shape = (Channel[1], Channel[0], kernal[0], kernal[0], kernal[0])
Dl1 = lrelu(batchnorm(conv(X, w1, filter_shape=filter_shape), g=g1, b=b1))
filter_shape = (Channel[2], Channel[1], kernal[1], kernal[1], kernal[1])
Dl2 = lrelu(batchnorm(conv(Dl1, w2, filter_shape=filter_shape), g=g2,
b=b2))
filter_shape = (Channel[3], Channel[2], kernal[2], kernal[2], kernal[2])
Dl3 = lrelu(batchnorm(conv(Dl2, w3, filter_shape=filter_shape), g=g3,
b=b3))
filter_shape = (Channel[4], Channel[3], kernal[3], kernal[3], kernal[3])
Dl4 = lrelu(batchnorm(conv(Dl3, w4, filter_shape=filter_shape), g=g4,
b=b4))
Dl4 = T.flatten(Dl4, 2)
DlZ = sigmoid(T.dot(Dl4, wz))
return DlZ
def apply(self, input, rand_vals=None):
"""
Apply this generator module to some input.
"""
batch_size = input.shape[0]
bm = int((self.filt_dim - 1) / 2) # use "same" mode convolutions
ss = self.us_stride # stride for "learned upsampling"
if self.use_pooling:
# "unpool" the input if desired
input = input.repeat(ss, axis=2).repeat(ss, axis=3)
# get shape for random values that will augment input
rand_shape = (batch_size, self.rand_chans, input.shape[2], input.shape[3])
if self.use_rand:
# augment input with random channels
if rand_vals is None:
if self.rand_type == 'normal':
rand_vals = self.rng.normal(size=rand_shape, avg=0.0, std=1.0, \
dtype=theano.config.floatX)
else:
rand_vals = self.rng.uniform(size=rand_shape, low=-1.0, high=1.0, \
dtype=theano.config.floatX)
rand_vals = rand_vals.reshape(rand_shape)
# stack random values on top of input
full_input = T.concatenate([rand_vals, input], axis=1)
else:
# don't augment input with random channels
full_input = input
# apply first convolution, perhaps with fractional striding
if self.use_pooling:
h1 = dnn_conv(full_input, self.w1, subsample=(1, 1), border_mode=(bm, bm))
else:
# apply first conv layer (with fractional stride for upsampling)
h1 = deconv(full_input, self.w1, subsample=(ss, ss), border_mode=(bm, bm))
if self.apply_bn_1:
h1 = batchnorm(h1, g=self.g1, b=self.b1)
h1 = relu(h1)
# apply second conv layer
h2 = dnn_conv(h1, self.w2, subsample=(1, 1), border_mode=(bm, bm))
if self.apply_bn_2:
h2 = batchnorm(h2, g=self.g2, b=self.b2)
h2 = relu(h2)
return h2
def gen(_z, _params, n_layers=3, n_f=128, init_sz=4, nc=3):
[gw0, gg0, gb0] = _params[0:3]
hs = []
h0 = relu(batchnorm(T.dot(_z, gw0), g=gg0, b=gb0))
h1 = h0.reshape((h0.shape[0], n_f * 2**n_layers, init_sz, init_sz))
hs.extend([h0, h1])
for n in range(n_layers):
[w, g, b] = _params[3 * (n + 1):3 * (n + 2)]
hin = hs[-1]
hout = relu(
batchnorm(deconv(hin, w, subsample=(2, 2), border_mode=(2, 2)),
g=g,
b=b))
hs.append(hout)
x = deconv(hs[-1], _params[-1], subsample=(2, 2), border_mode=(2, 2))
if nc == 3:
x_f = tanh(x)
if nc == 1:
x_f = sigmoid(x)
return x_f
def residual_block(inputs, output_channels, stride, scope):
with tf.variable_scope(scope):
net = ops.conv3d(inputs,
3,
output_channels,
stride,
use_bias=False,
scope='conv_1')
if (FLAGS.GAN_type == 'GAN'):
net = ops.batchnorm(net, FLAGS.is_training)
net = ops.prelu_tf(net)
net = ops.conv3d(net,
3,
output_channels,
stride,
use_bias=False,
scope='conv_2')
if (FLAGS.GAN_type == 'GAN'):
net = ops.batchnorm(net, FLAGS.is_training)
net = net + inputs
return net
def generator_function(hidden_data, is_train=True):
# layer 0 (linear)
h0 = T.dot(hidden_data, linear_w0)
h0 = h0 + t_rng.normal(size=h0.shape, std=0.01, dtype=t_floatX)
h0 = relu(batchnorm(X=h0, g=linear_bn_w0, b=linear_bn_b0))
h0 = h0.reshape((h0.shape[0], num_gen_filters0, init_image_size, init_image_size))
# layer 1 (deconv)
h1 = deconv(h0, conv_w1, subsample=(2, 2), border_mode=(2, 2))
h1 = h1 + t_rng.normal(size=h1.shape, std=0.01, dtype=t_floatX)
h1 = relu(batchnorm(h1, g=conv_bn_w1, b=conv_bn_b1))
# layer 2 (deconv)
h2 = deconv(h1, conv_w2, subsample=(2, 2), border_mode=(2, 2))
h2 = h2 + t_rng.normal(size=h2.shape, std=0.01, dtype=t_floatX)
h2 = relu(batchnorm(h2, g=conv_bn_w2, b=conv_bn_b2))
# layer 3 (deconv)
h3 = deconv(h2, conv_w3, subsample=(2, 2), border_mode=(2, 2))
h3 = h3 + t_rng.normal(size=h3.shape, std=0.01, dtype=t_floatX)
h3 = relu(batchnorm(h3, g=conv_bn_w3, b=conv_bn_b3))
# layer 4 (deconv)
output = tanh(deconv(h3, conv_w4, subsample=(2, 2), border_mode=(2, 2))+conv_b4.dimshuffle('x', 0, 'x', 'x'))
return output
def transitionLayer(layer_inputs, output_channel, is_training):
net = layer_inputs
net = batchnorm(net, is_training)
net = prelu_tf(net)
net = conv2(net,
1,
output_channel,
stride=1,
use_bias=False,
scope='conv1x1')
return net
def gen(Z, w, g, b, w2, g2, b2, w3, g3, b3, w4, g4, b4, w5, g5, b5, w6, g6, b6,
wx):
h = relu(batchnorm(T.dot(Z, w), g=g, b=b))
h = h.reshape((h.shape[0], ngf * 4, 4, 4))
h2 = relu(
batchnorm(deconv(h, w2, subsample=(2, 2), border_mode=(1, 1)),
g=g2,
b=b2))
h3 = relu(
batchnorm(deconv(h2, w3, subsample=(1, 1), border_mode=(1, 1)),
g=g3,
b=b3))
h4 = relu(
batchnorm(deconv(h3, w4, subsample=(2, 2), border_mode=(1, 1)),
g=g4,
b=b4))
h5 = relu(
batchnorm(deconv(h4, w5, subsample=(1, 1), border_mode=(1, 1)),
g=g5,
b=b5))
h6 = relu(
batchnorm(deconv(h5, w6, subsample=(2, 2), border_mode=(1, 1)),
g=g6,
b=b6))
x = tanh(deconv(h6, wx, subsample=(1, 1), border_mode=(1, 1)))
return x
def gen_test(_z,
_params,
_batchnorm,
n_layers=3,
n_f=128,
init_sz=4,
nc=3,
use_tanh=False):
if use_tanh:
_z = tanh(_z)
# gw0 : weight of dense layer(0)
# gg0 , gb0 : params of batchnorm layer
[gw0, gg0, gb0] = _params[0:3]
hs = []
u = _batchnorm[0]
s = _batchnorm[n_layers + 1]
# Clip z => Dense => BatchNorm => ReLU
h0 = relu(
batchnorm(T.dot(T.clip(_z, -1.0, 1.0), gw0), u=u, s=s, g=gg0, b=gb0))
# reshape to 4D
h1 = h0.reshape((h0.shape[0], n_f * 2**n_layers, init_sz, init_sz))
hs.extend([h0, h1])
for n in range(n_layers):
[w, g, b] = _params[3 * (n + 1):3 * (n + 2)]
hin = hs[-1]
u = _batchnorm[n + 1]
s = _batchnorm[n + n_layers + 2]
hout = relu(
batchnorm(deconv(hin, w, subsample=(2, 2), border_mode=(2, 2)),
u=u,
s=s,
g=g,
b=b))
hs.append(hout)
x = deconv(hs[-1], _params[-1], subsample=(2, 2), border_mode=(2, 2))
if nc == 3:
x_f = tanh(x)
if nc == 1:
x_f = sigmoid(x)
return x_f
def converter( IS, w1, w2, g2, b2, w3, g3, b3, w4, g4, b4, w5, g5, b5,
w5d, g5d, b5d, w4d, g4d, b4d, w3d, g3d, b3d, w2d, g2d, b2d, w1d ):
h1 = lrelu( dnn_conv( IS, w1, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ) )
h2 = lrelu( batchnorm( dnn_conv( h1, w2, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g2, b = b2 ) )
h3 = lrelu( batchnorm( dnn_conv( h2, w3, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g3, b = b3 ) )
h4 = lrelu( batchnorm( dnn_conv( h3, w4, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g4, b = b4 ) )
h5 = lrelu( batchnorm( dnn_conv( h4, w5, subsample = ( 1, 1 ), border_mode = ( 0, 0 ) ), g = g5, b = b5 ) )
h5d = relu( batchnorm( dnn_conv( h5, w5d, subsample = ( 1, 1 ), border_mode = ( 0, 0 ) ), g = g5d, b = b5d ) )
h5d = h5d.reshape( ( h5d.shape[0], nf * 8, 4, 4 ) )
h4d = relu( batchnorm( deconv( h5d, w4d, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g4d, b = b4d ) )
h3d = relu( batchnorm( deconv( h4d, w3d, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g3d, b = b3d ) )
h2d = relu( batchnorm( deconv( h3d, w2d, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ), g = g2d, b = b2d ) )
h1d = tanh( deconv( h2d, w1d, subsample = ( 2, 2 ), border_mode = ( 2, 2 ) ) )
y = h1d
return y
def disc_test(_x, _params, _batchnorm, n_layers=3):
w = _params[0]
h0 = lrelu(dnn_conv(_x, w, subsample=(2, 2), border_mode=(2, 2)))
hs = [h0]
for n in range(n_layers):
hin = hs[-1]
w, g, b = _params[1 + 3 * n:1 + 3 * (n + 1)]
u = _batchnorm[n]
s = _batchnorm[n + n_layers]
hout = lrelu(batchnorm(dnn_conv(hin, w, subsample=(2, 2), border_mode=(2, 2)), u=u, s=s, g=g, b=b))
hs.append(hout)
h = T.flatten(hs[-1], 2)
y = sigmoid(T.dot(h, _params[-1]))
return y
def discriminator_block(inputs, output_channel, kernel_size, stride,
scope):
with tf.variable_scope(scope):
net = ops.conv3d(inputs,
kernel_size,
output_channel,
stride,
use_bias=False,
scope='conv1')
if (FLAGS.GAN_type == 'GAN'):
net = ops.batchnorm(net, FLAGS.is_training)
net = ops.lrelu(net, 0.2)
return net
def predict_test(_x, _params, _batchnorm, n_layers=3):
w = _params[0]
h0 = lrelu(dnn_conv(_x, w, subsample=(2, 2), border_mode=(2, 2)))
hs = [h0]
for n in range(n_layers):
hin = hs[-1]
w, g, b = _params[1 + 3 * n:1 + 3 * (n + 1)]
u = _batchnorm[n]
s = _batchnorm[n + n_layers]
hout = lrelu(batchnorm(dnn_conv(hin, w, subsample=(2, 2), border_mode=(2, 2)), u=u, s=s, g=g, b=b))
hs.append(hout)
h = T.flatten(hs[-1], 2)
y = tanh(T.dot(h, _params[-1]))
return y
def gen(Z, Y, w, w2, w3, wx):
print '\n@@@@ gen()'
printVal('Z', Z) # matrix
#printVal( 'Y', Y ) # matrix
printVal('w', w) # matrix
printVal('w2', w2) # matrix
printVal('w3', w3) # tensor
printVal('wx', wx) # tensor
# Yの要素の並びの入れ替え。数字の引数は、次元番号。'x' は ブロードキャスト
# 並び替えの前後で、全体の要素数は変わらない。
yb = Y.dimshuffle(0, 1, 'x', 'x')
# yb は４次元テンソル
#printVal('yb', yb)
# 行列 Z と Y を結合（横方向）
Z = T.concatenate([Z, Y], axis=1) # matrix
# Z*w(Full Connect) をバッチ正規化して、ReLU 適用
tmp_a = T.dot(Z, w) # dot(matrix, matrix)->matrix
printVal('dot(Z,w) -> tmp_a', tmp_a)
h = relu(batchnorm(T.dot(Z, w))) #CCC
h = T.concatenate([h, Y], axis=1) #CCC
printVal('h', h) # matrix
h2 = relu(batchnorm(T.dot(h, w2))) #CCC
printVal('h2', h2) #h2:matrix
h2r = h2.reshape((h2.shape[0], GEN_NUM_FILTER * 2, 7, 7)) #CCC
printVal('h2r', h2r) #h2r:tensor
h2ry = conv_cond_concat(h2r, yb) #
printVal('h2ry', h2ry) #h2:tensor
# デコンボリューション:論文によれば、空間プーリングの代わりに適用する
d = deconv(h2ry, w3, subsample=(2, 2), border_mode=(2, 2))
printVal('d', d)
#h3 = relu(batchnorm(deconv(h2, w3, subsample=(2, 2), border_mode=(2, 2))))
h3 = relu(batchnorm(d))
h3 = conv_cond_concat(h3, yb)
x = sigmoid(deconv(h3, wx, subsample=(2, 2), border_mode=(2, 2)))
return x, h2
def predict(_x, _params, n_layers=3):
w = _params[0]
h0 = lrelu(dnn_conv(_x, w, subsample=(2, 2), border_mode=(2, 2)))
hs = [h0]
for n in range(n_layers):
hin = hs[-1]
w, g, b = _params[1 + 3 * n:1 + 3 * (n + 1)]
hout = lrelu(
batchnorm(dnn_conv(hin, w, subsample=(2, 2), border_mode=(2, 2)),
g=g,
b=b))
hs.append(hout)
h = T.flatten(hs[-1], 2)
y = tanh(T.dot(h, _params[-1]))
return y