def _decode(self, X, init_params):
params = []
regs = []
d4, pars = Ops.dense(X,
self.latent_size,
self.input_size[0] * self.input_size[1] * 5,
layer_name='d4',
init_params=get_params('d4', init_params))
params += pars
regs.append(pars[0])
d4 = T.reshape((d4), (-1, 5, self.input_size[0], self.input_size[1]))
d3, pars = Ops.conv_2d(activations.relu(d4), (8, 5, 5, 5),
layer_name='d3',
mode='half',
init_params=get_params('d3', init_params))
params += pars
regs.append(pars[0])
d2, pars = Ops.conv_2d(activations.relu(d3), (10, 8, 5, 5),
layer_name='d2',
mode='half',
init_params=get_params('d2', init_params))
params += pars
regs.append(pars[0])
d1, pars = Ops.conv_2d(activations.relu(d2), (1, 10, 5, 5),
layer_name='d1',
mode='half',
init_params=get_params('d1', init_params))
params += pars
regs.append(pars[0])
return activations.sigmoid(d1), params, regs
def _encode(self, X, init_params):
params = []
regs = []
e1, pars = Ops.conv_2d(X, (10, 1, 5, 5),
layer_name='e1',
mode='half',
init_params=get_params('e1', init_params))
params += pars
regs.append(pars[0])
e2, pars = Ops.conv_2d(activations.relu(e1), (8, 10, 5, 5),
layer_name='e2',
mode='half',
init_params=get_params('e2', init_params))
params += pars
regs.append(pars[0])
e3, pars = Ops.conv_2d(activations.relu(e2), (5, 8, 5, 5),
layer_name='e3',
mode='half',
init_params=get_params('e3', init_params))
params += pars
regs.append(pars[0])
e4, pars = Ops.dense(Ops.flatten(activations.relu(e3)),
self.input_size[0] * self.input_size[1] * 5,
self.latent_size,
layer_name='e4',
init_params=get_params('e4', init_params))
params += pars
regs.append(pars[0])
return e4, params, regs
def _conv_block(x, nb_filters, nb_channels, filter_size, block_idx, theano_model, double=True, init_params=None):
params = []
to_reg = []
layer, pars = conv_2d(
x, (nb_filters, nb_channels, filter_size,
filter_size), layer_name=str(block_idx) + 'l1',
init_params=theano_model.get_params(str(block_idx) + 'l1', param_list=init_params))
params += pars
to_reg.append(pars[0])
layer, pars = bn(layer, trainable=True, layer_name=str(block_idx) + 'bn1',
init_params=theano_model.get_params(
str(block_idx) + 'bn1',
param_list=init_params))
params += pars
if double:
layer, pars = conv_2d(
layer, (2 * nb_filters, nb_filters, filter_size, filter_size), layer_name=str(block_idx) + 'l2',
init_params=theano_model.get_params(str(block_idx) + 'l2', param_list=init_params))
params += pars
to_reg.append(pars[0])
else:
layer, pars = conv_2d(
layer, (nb_filters, nb_filters, filter_size,
filter_size), layer_name=str(block_idx) + 'l2',
init_params=theano_model.get_params(str(block_idx) + 'l2', param_list=init_params))
params += pars
to_reg.append(pars[0])
return relu(layer), params, to_reg
def accgan_deconv(x, theano_model, input_channels, nb_filters, layer_name, init_params, filter_size=3, stride=(1,1), batch_norm=True, dropout_prob=0.2):
params = []
regs = []
l = T.nnet.abstract_conv.bilinear_upsampling(x, ratio=2)
l, pars = conv_2d(l, (nb_filters, input_channels, filter_size, filter_size),
layer_name=layer_name,
mode='half',
stride=stride,
init_params=theano_model.get_params(layer_name, init_params))
params += pars
regs.append(pars[0])
if batch_norm:
l, pars = bn(inpt=l, trainable=True, layer_name=layer_name, init_params=theano_model.get_params(layer_name, init_params))
params += pars
l, pars = conv_2d(l, (nb_filters, nb_filters, filter_size, filter_size),
layer_name=layer_name+'second',
mode='half',
stride=stride,
init_params=theano_model.get_params(layer_name+'second', init_params))
params += pars
regs.append(pars[0])
if batch_norm:
l, pars = bn(inpt=l, trainable=True, layer_name=layer_name+'second', init_params=theano_model.get_params(layer_name+'second', init_params))
params += pars
l = dropout(relu(l), dropout_prob)
return l, params, regs
def label_encoder(self, x):
code, _par = dense(flatten(x), self.class_label_size, self.encoded_label_size,
layer_name='omega', init_params=self.get_params('omega', self.init_params))
code = relu(code)
self.params += _par
self.to_regularize.append(_par[0])
return code
def accgan_conv(x, theano_model, input_channels, nb_filters, layer_name, init_params, filter_size=3, stride=(1,1), batch_norm=True, dropout_prob=0.2):
params = []
regs = []
l, pars = conv_2d(x, (nb_filters, input_channels, filter_size, filter_size),
layer_name=layer_name,
mode='half',
stride=stride,
init_params=theano_model.get_params(layer_name, init_params))
params += pars
regs.append(pars[0])
if batch_norm:
l, pars = bn(inpt=l, trainable=True, layer_name=layer_name, init_params=theano_model.get_params(layer_name, init_params))
params += pars
l = dropout(relu(l), dropout_prob)
return l, params, regs