def phi(x):
_ = x / 255.0
# Nature structure
with O.argscope(O.conv2d, nonlin=O.relu):
_ = O.conv2d('conv1', _, 32, 8, stride=4)
_ = O.conv2d('conv2', _, 64, 4, stride=2)
_ = O.conv2d('conv3', _, 64, 3, stride=1)
return _
def forward_conv(x):
_ = x / 255.0
with O.argscope(O.conv2d, nonlin=O.relu):
_ = O.conv2d('conv0', _, 32, 5)
_ = O.max_pooling2d('pool0', _, 2)
_ = O.conv2d('conv1', _, 32, 5)
_ = O.max_pooling2d('pool1', _, 2)
_ = O.conv2d('conv2', _, 64, 4)
_ = O.max_pooling2d('pool2', _, 2)
_ = O.conv2d('conv3', _, 64, 3)
return _
def forward(x):
_ = x / 255.0
with O.argscope(O.conv2d, nonlin=O.relu):
_ = O.conv2d('conv0', _, 32, 5)
_ = O.max_pooling2d('pool0', _, 2)
_ = O.conv2d('conv1', _, 32, 5)
_ = O.max_pooling2d('pool1', _, 2)
_ = O.conv2d('conv2', _, 64, 4)
_ = O.max_pooling2d('pool2', _, 2)
_ = O.conv2d('conv3', _, 64, 3)
dpc.add_output(_, name='feature')
def encoder(x, nonlin):
w_init = O.truncated_normal_initializer(stddev=0.02)
with O.argscope(O.conv2d, O.deconv2d, kernel=4, stride=2, W=w_init),\
O.argscope(O.leaky_relu, alpha=0.2):
_ = x
_ = O.conv2d('conv1', _, 64, nonlin=O.leaky_relu)
_ = O.conv2d('conv2', _, 128, nonlin=nonlin, use_bias=False)
_ = O.conv2d('conv3', _, 256, nonlin=nonlin, use_bias=False)
_ = O.conv2d('conv4', _, 512, nonlin=nonlin, use_bias=False)
z = _
return z
def discriminator(img):
w_init = O.truncated_normal_initializer(stddev=0.02)
with O.argscope(O.conv2d, O.deconv2d, kernel=4, stride=2, W=w_init),\
O.argscope(O.fc, W=w_init),\
O.argscope(O.leaky_relu, alpha=0.2):
_ = img
_ = O.conv2d('conv1', _, 64, nonlin=O.leaky_relu)
_ = O.conv2d('conv2', _, 128, nonlin=O.bn_nonlin)
_ = O.leaky_relu(_)
_ = O.fc('fc1', _, 1024, nonlin=O.bn_nonlin)
_ = O.leaky_relu(_)
_ = O.fc('fct', _, 1)
return _
def forward(img):
_ = img
_ = O.conv2d('conv1',
_,
16, (3, 3),
padding='SAME',
nonlin=O.relu)
_ = O.pooling2d('pool1', _, kernel=2)
_ = O.conv2d('conv2',
_,
32, (3, 3),
padding='SAME',
nonlin=O.relu)
_ = O.pooling2d('pool2', _, kernel=2)
dpc.add_output(_, name='feature')
def forward(img):
_ = img
_ = O.conv2d('conv1',
_,
16, (3, 3),
padding='SAME',
nonlin=O.identity)
_ = O.batch_norm('bn1', _)
_ = O.relu(_)
_ = O.pooling2d('pool1', _, kernel=2)
_ = O.conv2d('conv2',
_,
32, (3, 3),
padding='SAME',
nonlin=O.identity)
_ = O.batch_norm('bn2', _)
_ = O.relu(_)
_ = O.pooling2d('pool2', _, kernel=2)
dpc.add_output(_, name='feature')
def discriminator(img):
w_init = O.truncated_normal_initializer(stddev=0.02)
with O.argscope(O.conv2d, O.deconv2d, kernel=4, stride=2, W=w_init),\
O.argscope(O.fc, W=w_init),\
O.argscope(O.leaky_relu, alpha=0.2):
_ = img
_ = O.conv2d('conv1', _, 64, nonlin=O.leaky_relu)
_ = O.conv2d('conv2', _, 128, nonlin=O.bn_nonlin)
_ = O.leaky_relu(_)
_ = O.fc('fc1', _, 1024, nonlin=O.bn_nonlin)
_ = O.leaky_relu(_)
with env.variable_scope('score'):
logits = O.fc('fct', _, 1)
with env.variable_scope('code'):
_ = O.fc('fc1', _, 128, nonlin=O.bn_nonlin)
_ = O.leaky_relu(_)
code = O.fc('fc2', _, zc_distrib.param_size)
return logits, code
def forward(img):
_ = img
_ = O.conv2d('conv1',
_,
4, (3, 3),
stride=2,
padding='SAME',
nonlin=O.relu)
# shape = (14, 14)
_ = O.conv2d('conv2',
_,
8, (3, 3),
stride=2,
padding='SAME',
nonlin=O.relu)
# shape = (7, 7)
_ = O.fc('fc', _, 392)
_ = _.reshape([-1, 7, 7, 8])
_ = O.deconv2d('deconv1',
_,
4, (3, 3),
stride=2,
padding='SAME',
nonlin=O.relu)
# shape = (14, 14)
_ = O.deconv2d('deconv2',
_,
1, (3, 3),
stride=2,
padding='SAME',
nonlin=O.sigmoid)
# shape = (28, 28)
out = _
loss = O.raw_cross_entropy_prob('raw_loss', out, img)
loss = O.get_pn_balanced_loss('loss', loss, img)
dpc.add_output(out, name='output')
dpc.add_output(loss, name='loss', reduce_method='sum')
def stacked_conv(prefix,
nr_convs,
in_,
channel,
kernel=(3, 3),
padding='SAME',
nonlin=O.relu):
for i in range(1, nr_convs + 1):
in_ = O.conv2d('{}_{}'.format(prefix, i),
in_,
channel,
kernel,
padding=padding,
nonlin=nonlin)
return in_
