def init_gen_model(state):
gen_model = nn.Sequential()
gen_model.add(
nn.Linear(state['noise_size'],
state['g_num_filters'] * 4 * 7 * 7,
weight=state['g_init'],
use_bias=False))
gen_model.add(nn.BatchNorm(state['g_num_filters'] * 4 * 7 * 7))
gen_model.add(nn.Expression(T.nnet.relu))
gen_model.add(
nn.Expression(lambda x: T.reshape(x, (x.shape[0], state['g_num_filters'
] * 4, 7, 7))))
gen_model.add(
nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['g_num_filters'] * 4,
num_channels=state['g_num_filters'] * 2,
step=(2, 2),
border_mode=(1, 1),
use_bias=False,
weight=state['g_conv_init']))
gen_model.add(nn.BatchNorm(state['g_num_filters'] * 2))
gen_model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, num_filters, 14, 14)
gen_model.add(
nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['g_num_filters'] * 2,
num_channels=state['g_num_filters'],
step=(2, 2),
border_mode=(1, 1),
use_bias=False,
weight=state['g_conv_init']))
gen_model.add(nn.BatchNorm(state['g_num_filters']))
gen_model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, input_channels, 28, 28)
gen_model.add(
nn.Deconvolutional(filter_size=(3, 3),
num_filters=state['g_num_filters'],
num_channels=state['input_channels'],
step=(1, 1),
border_mode=(1, 1),
use_bias=True,
weight=state['g_conv_init']))
# gen_model.add(nn.Expression(T.nnet.sigmoid))
# out_shape == (b, input_channels, 28, 28)
return gen_model
def init_decoder(state):
# inp_shape == (b, d_num_filters*4, 8, 8)
model = nn.Sequential()
model.add(
nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['d_num_filters'] * 4,
num_channels=state['d_num_filters'] * 2,
step=(2, 2),
border_mode=(1, 1),
weight=state['d_conv_init'],
use_bias=False,
name='d_dec_conv1'))
model.add(nn.BatchNorm(state['d_num_filters'] * 2))
model.add(nn.LeakyRectify())
# model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, d_num_filters*2, 16, 16)
model.add(
nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['d_num_filters'] * 2,
num_channels=state['d_num_filters'],
step=(2, 2),
border_mode=(1, 1),
weight=state['d_conv_init'],
use_bias=False,
name='d_dec_conv2'))
model.add(nn.BatchNorm(state['d_num_filters']))
model.add(nn.LeakyRectify())
# model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, d_num_filters, 32, 32)
model.add(
nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['d_num_filters'],
num_channels=state['input_channels'],
step=(2, 2),
border_mode=(1, 1),
weight=state['d_conv_init'],
use_bias=False,
name='d_dec_conv3'))
model.add(nn.BatchNorm(state['input_channels']))
model.add(nn.Expression(T.tanh))
# out_shape == (b, input_channels, 64, 64)
return model
def init_gen_model(state):
gen_model = nn.Sequential()
gen_model.add(
nn.Linear(state['noise_size'],
state['g_num_filters'] * 4 * 4 * 4,
weight=state['g_init'],
use_bias=False))
gen_model.add(nn.BatchNorm(state['g_num_filters'] * 4 * 4 * 4))
gen_model.add(nn.Expression(T.nnet.relu))
gen_model.add(
nn.Expression(lambda x: T.reshape(x, (x.shape[0], state['g_num_filters'
] * 4, 4, 4))))
gen_model.add(
nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['g_num_filters'] * 4,
num_channels=state['g_num_filters'] * 4,
step=(2, 2),
border_mode=(1, 1),
use_bias=False,
weight=state['g_conv_init']))
gen_model.add(nn.BatchNorm(state['g_num_filters'] * 4))
gen_model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, g_num_filters*6, 8, 8)
gen_model.add(
nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['g_num_filters'] * 4,
num_channels=state['g_num_filters'] * 2,
step=(2, 2),
border_mode=(1, 1),
use_bias=False,
weight=state['g_conv_init']))
gen_model.add(nn.BatchNorm(state['g_num_filters'] * 2))
gen_model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, g_num_filters*2, 16, 16)
gen_model.add(
nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['g_num_filters'] * 2,
num_channels=state['g_num_filters'],
step=(2, 2),
border_mode=(1, 1),
use_bias=False,
weight=state['g_conv_init']))
gen_model.add(nn.BatchNorm(state['g_num_filters']))
gen_model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, g_num_filters, 32, 32)
gen_model.add(
nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['g_num_filters'],
num_channels=state['input_channels'],
step=(2, 2),
border_mode=(1, 1),
use_bias=True,
weight=state['g_conv_init']))
gen_model.add(nn.Expression(T.tanh))
# out_shape == (b, input_channels, 64, 64)
return gen_model
use_bias=True, name='q_fc5'))
# 2) Directed generator
gen_model = nn.Sequential()
gen_model.add(nn.Linear(state['noise_size'], state['g_num_filters']*4*4*4, weight=state['g_init'],
use_bias=False, name='g_fc1'))
gen_model.add(nn.BatchNorm(state['g_num_filters']*4*4*4, name='g_bn1'))
gen_model.add(nn.Expression(T.nnet.relu))
gen_model.add(nn.Expression(lambda x: T.reshape(x, (x.shape[0], state['g_num_filters']*4, 4, 4))))
gen_model.add(nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['g_num_filters']*4,
num_channels=state['g_num_filters']*4,
step=(2, 2), border_mode=(1, 1),
use_bias=False,
weight=state['g_init'],
name='g_deconv2'))
gen_model.add(nn.BatchNorm(state['g_num_filters']*4, name='g_bn2'))
gen_model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, num_filters, 8, 8)
gen_model.add(nn.Deconvolutional(filter_size=(4, 4),
num_filters=state['g_num_filters']*4,
num_channels=state['g_num_filters']*2,
step=(2, 2), border_mode=(1, 1),
use_bias=False,
weight=state['g_init'],
name='g_deconv3'))
gen_model.add(nn.BatchNorm(state['g_num_filters']*2, name='g_bn3'))