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_dcgan_disc(state):
disc_model = disc_shared_structure(state)
disc_model.add(nn.Linear(state['d_num_filters']*4*7*7, 1, weight=state['d_init'], use_bias=True))
disc_model.add(nn.Expression(lambda x: T.flatten(x)))
return disc_model
def disc_shared_structure(state):
model = nn.Sequential()
model.add(nn.Convolutional(filter_size=(3, 3),
num_filters=state['d_num_filters'],
num_channels=state['input_channels'],
step=(1, 1), border_mode=(1, 1),
weight=state['d_conv_init'], use_bias=False))
model.add(nn.BatchNorm(state['d_num_filters']))
# model.add(nn.Expression(T.nnet.relu))
model.add(nn.LeakyRectify())
# out_shape == (b, num_filters, 28, 28)
model.add(nn.Convolutional(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))
model.add(nn.BatchNorm(state['d_num_filters']*2))
# model.add(nn.Expression(T.nnet.relu))
model.add(nn.LeakyRectify())
# out_shape == (b, num_filters, 14, 14)
model.add(nn.Convolutional(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))
model.add(nn.BatchNorm(state['d_num_filters']*4))
# model.add(nn.Expression(T.nnet.relu))
model.add(nn.LeakyRectify())
# out_shape == (b, num_filters, 7, 7)
model.add(nn.Expression(lambda x: T.flatten(x, 2)))
return model
def disc_shared_structure(state):
model = nn.Sequential()
if state.has_key('init_d') and state['init_d'] is not None:
model.add(
nn.Linear(state['input_size'],
state['hidden_size'],
weight=state['init_d']))
else:
model.add(nn.Linear(state['input_size'], state['hidden_size']))
model.add(nn.Expression(T.nnet.relu))
if state.has_key('init_d') and state['init_d'] is not None:
model.add(
nn.Linear(state['hidden_size'],
state['hidden_size'],
weight=state['init_d']))
else:
model.add(nn.Linear(state['hidden_size'], state['hidden_size']))
model.add(nn.Expression(T.nnet.relu))
return model
def init_gen_model(state):
gen_model = nn.Sequential()
gen_model.add(
nn.Linear(state['noise_size'],
state['hidden_size'],
weight=state['init_g']))
gen_model.add(nn.BatchNorm(state['hidden_size']))
gen_model.add(nn.Expression(T.nnet.relu))
gen_model.add(
nn.Linear(state['hidden_size'],
state['hidden_size'],
weight=state['init_g']))
gen_model.add(nn.BatchNorm(state['hidden_size']))
gen_model.add(nn.Expression(T.nnet.relu))
gen_model.add(
nn.Linear(state['hidden_size'],
state['input_size'],
weight=state['init_g']))
return gen_model
def init_dcgan_disc(state):
disc_model = disc_shared_structure(state)
disc_model.add(
nn.MinibatchDisc(state['d_num_filters'] * 4 * 4 * 4,
num_kernels=100,
dim_per_kernel=5,
concat=True))
disc_model.add(
nn.Linear(state['d_num_filters'] * 4 * 4 * 4 + 100,
1,
weight=state['d_init'],
use_bias=True))
disc_model.add(nn.Expression(lambda x: T.flatten(x)))
return disc_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
def disc_shared_structure(state):
# inp_shape == (b, input_channels, 64, 64)
model = nn.Sequential()
if state['dropout'] > 0:
model.add(nn.Dropout(state['dropout']))
model.add(
nn.Convolutional(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_conv1'))
model.add(nn.BatchNorm(state['d_num_filters']))
# model.add(nn.LeakyRectify())
model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, num_filters, 32, 32)
model.add(
nn.Convolutional(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_conv2'))
model.add(nn.BatchNorm(state['d_num_filters'] * 2))
# model.add(nn.LeakyRectify())
model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, num_filters, 16, 16)
model.add(
nn.Convolutional(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))
model.add(nn.BatchNorm(state['d_num_filters'] * 4))
# model.add(nn.LeakyRectify())
model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, num_filters, 8, 8)
model.add(
nn.Convolutional(filter_size=(4, 4),
num_filters=state['d_num_filters'] * 4,
num_channels=state['d_num_filters'] * 4,
step=(2, 2),
border_mode=(1, 1),
weight=state['d_conv_init'],
use_bias=False))
model.add(nn.BatchNorm(state['d_num_filters'] * 4))
# model.add(nn.LeakyRectify())
model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, num_filters, 4, 4)
model.add(nn.Expression(lambda x: T.flatten(x, 2)))
return model
np.random.seed(12345)
############################
# Init model & parameters
############################
# 1) Eneregy discriminator
disc_model = nn.Sequential()
disc_model.add(nn.Convolutional(filter_size=(3, 3),
num_filters=state['d_num_filters'],
num_channels=state['input_channels'],
step=(1, 1), border_mode=(1, 1),
weight=state['d_init'], use_bias=False,
name='d_conv1'))
disc_model.add(nn.BatchNorm(state['d_num_filters'], name='d_bn1'))
disc_model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, num_filters, 32, 32)
disc_model.add(nn.Convolutional(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_init'], use_bias=False,
name='d_conv2'))
disc_model.add(nn.BatchNorm(state['d_num_filters']*2, name='d_bn2'))
disc_model.add(nn.Expression(T.nnet.relu))
# out_shape == (b, num_filters, 16, 16)
disc_model.add(nn.Convolutional(filter_size=(4, 4),
num_filters=state['d_num_filters']*4,
num_channels=state['d_num_filters']*2,