def InceptionLayer(incoming, param_dict, block_name):
branch = [0] * len(param_dict)
# Loop across branches
for i, dict in enumerate(param_dict):
for j, style in enumerate(dict['style']): # Loop up branch
branch[i] = C2D(
incoming = branch[i] if j else incoming,
num_filters = dict['num_filters'][j],
filter_size = dict['filter_size'][j],
pad = dict['pad'][j] if 'pad' in dict else None,
stride = dict['stride'][j],
W = initmethod('relu'),
nonlinearity = dict['nonlinearity'][j],
name = block_name+'_'+str(i)+'_'+str(j)) if style=='convolutional'\
else NL(lasagne.layers.dnn.Pool2DDNNLayer(
incoming=incoming if j == 0 else branch[i],
pool_size = dict['filter_size'][j],
mode = dict['mode'][j],
stride = dict['stride'][j],
pad = dict['pad'][j],
name = block_name+'_'+str(i)+'_'+str(j)),
nonlinearity = dict['nonlinearity'][j]) if style=='pool'\
else lasagne.layers.DilatedConv2DLayer(
incoming = lasagne.layers.PadLayer(incoming = incoming if j==0 else branch[i],width = dict['pad'][j]) if 'pad' in dict else incoming if j==0 else branch[i],
num_filters = dict['num_filters'][j],
filter_size = dict['filter_size'][j],
dilation = dict['dilation'][j],
# pad = dict['pad'][j] if 'pad' in dict else None,
W = initmethod('relu'),
nonlinearity = dict['nonlinearity'][j],
name = block_name+'_'+str(i)+'_'+str(j)) if style== 'dilation'\
else DL(
incoming = incoming if j==0 else branch[i],
num_units = dict['num_filters'][j],
W = initmethod('relu'),
b = None,
nonlinearity = dict['nonlinearity'][j],
name = block_name+'_'+str(i)+'_'+str(j))
# Apply Batchnorm
branch[i] = BN(branch[i],
name=block_name + '_bnorm_' + str(i) + '_' +
str(j)) if dict['bnorm'][j] else branch[i]
# Concatenate Sublayers
return CL(incomings=branch, name=block_name)
def get_model(interp=False):
dims, n_channels = tuple(cfg['dims']), cfg['n_channels']
shape = (None, n_channels) + dims
l_in = lasagne.layers.InputLayer(shape=shape)
l_enc_conv1 = C2D(incoming=l_in,
num_filters=128,
filter_size=[5, 5],
stride=[2, 2],
pad=(2, 2),
W=initmethod(0.02),
nonlinearity=lrelu(0.2),
name='enc_conv1')
l_enc_conv2 = BN(C2D(incoming=l_enc_conv1,
num_filters=256,
filter_size=[5, 5],
stride=[2, 2],
pad=(2, 2),
W=initmethod(0.02),
nonlinearity=lrelu(0.2),
name='enc_conv2'),
name='bnorm2')
l_enc_conv3 = BN(C2D(incoming=l_enc_conv2,
num_filters=512,
filter_size=[5, 5],
stride=[2, 2],
pad=(2, 2),
W=initmethod(0.02),
nonlinearity=lrelu(0.2),
name='enc_conv3'),
name='bnorm3')
l_enc_conv4 = BN(C2D(incoming=l_enc_conv3,
num_filters=1024,
filter_size=[5, 5],
stride=[2, 2],
pad=(2, 2),
W=initmethod(0.02),
nonlinearity=lrelu(0.2),
name='enc_conv4'),
name='bnorm4')
print(lasagne.layers.get_output_shape(l_enc_conv4, (196, 3, 64, 64)))
l_enc_fc1 = BN(DL(incoming=l_enc_conv4,
num_units=1000,
W=initmethod(0.02),
nonlinearity=relu,
name='enc_fc1'),
name='bnorm_enc_fc1')
# Define latent values
l_enc_mu, l_enc_logsigma = [
BN(DL(incoming=l_enc_fc1,
num_units=cfg['num_latents'],
nonlinearity=None,
name='enc_mu'),
name='mu_bnorm'),
BN(DL(incoming=l_enc_fc1,
num_units=cfg['num_latents'],
nonlinearity=None,
name='enc_logsigma'),
name='ls_bnorm')
]
l_Z_IAF = GaussianSampleLayer(l_enc_mu, l_enc_logsigma, name='l_Z_IAF')
l_IAF_mu, l_IAF_logsigma = [
MADE(l_Z_IAF, [cfg['num_latents']], 'l_IAF_mu'),
MADE(l_Z_IAF, [cfg['num_latents']], 'l_IAF_ls')
]
l_Z = IAFLayer(l_Z_IAF, l_IAF_mu, l_IAF_logsigma, name='l_Z')
l_dec_fc2 = DL(incoming=l_Z,
num_units=512 * 16,
nonlinearity=lrelu(0.2),
W=initmethod(0.02),
name='l_dec_fc2')
l_unflatten = lasagne.layers.ReshapeLayer(
incoming=l_dec_fc2,
shape=([0], 512, 4, 4),
)
l_dec_conv1 = DeconvLayer(incoming=l_unflatten,
num_filters=512,
filter_size=[5, 5],
stride=[2, 2],
crop=(2, 2),
W=initmethod(0.02),
nonlinearity=None,
name='dec_conv1')
l_dec_conv2a = MDBLOCK(incoming=l_dec_conv1,
num_filters=512,
scales=[0, 2],
name='dec_conv2a',
nonlinearity=lrelu(0.2))
l_dec_conv2 = DeconvLayer(incoming=l_dec_conv2a,
num_filters=256,
filter_size=[5, 5],
stride=[2, 2],
crop=(2, 2),
W=initmethod(0.02),
nonlinearity=None,
name='dec_conv2')
l_dec_conv3a = MDBLOCK(incoming=l_dec_conv2,
num_filters=256,
scales=[0, 2, 3],
name='dec_conv3a',
nonlinearity=lrelu(0.2))
l_dec_conv3 = DeconvLayer(incoming=l_dec_conv3a,
num_filters=128,
filter_size=[5, 5],
stride=[2, 2],
crop=(2, 2),
W=initmethod(0.02),
nonlinearity=None,
name='dec_conv3')
l_dec_conv4a = MDBLOCK(incoming=l_dec_conv3,
num_filters=128,
scales=[0, 2, 3],
name='dec_conv4a',
nonlinearity=lrelu(0.2))
l_dec_conv4 = BN(DeconvLayer(incoming=l_dec_conv4a,
num_filters=128,
filter_size=[5, 5],
stride=[2, 2],
crop=(2, 2),
W=initmethod(0.02),
nonlinearity=lrelu(0.2),
name='dec_conv4'),
name='bnorm_dc4')
R = NL(MDCL(l_dec_conv4, num_filters=2, scales=[2, 3, 4], name='R'),
sigmoid)
G = NL(
ESL([
MDCL(l_dec_conv4, num_filters=2, scales=[2, 3, 4], name='G_a'),
MDCL(R, num_filters=2, scales=[2, 3, 4], name='G_b')
]), sigmoid)
B = NL(
ESL([
MDCL(l_dec_conv4, num_filters=2, scales=[2, 3, 4], name='B_a'),
MDCL(CL([R, G]), num_filters=2, scales=[2, 3, 4], name='B_b')
]), sigmoid)
l_out = CL([
beta_layer(SL(R, slice(0, 1), 1), SL(R, slice(1, 2), 1)),
beta_layer(SL(G, slice(0, 1), 1), SL(G, slice(1, 2), 1)),
beta_layer(SL(B, slice(0, 1), 1), SL(B, slice(1, 2), 1))
])
minibatch_discrim = MinibatchLayer(
lasagne.layers.GlobalPoolLayer(l_enc_conv4),
num_kernels=500,
name='minibatch_discrim')
l_discrim = DL(incoming=minibatch_discrim,
num_units=3,
nonlinearity=lasagne.nonlinearities.softmax,
b=None,
W=initmethod(0.02),
name='discrimi')
return {
'l_in': l_in,
'l_out': l_out,
'l_mu': l_enc_mu,
'l_ls': l_enc_logsigma,
'l_Z': l_Z,
'l_IAF_mu': l_IAF_mu,
'l_IAF_ls': l_IAF_logsigma,
'l_Z_IAF': l_Z_IAF,
'l_introspect': [l_enc_conv1, l_enc_conv2, l_enc_conv3, l_enc_conv4],
'l_discrim': l_discrim
}
def get_model(interp=False):
dims, n_channels, n_classes = tuple(
cfg['dims']), cfg['n_channels'], cfg['n_classes']
shape = (None, n_channels) + dims
l_in = lasagne.layers.InputLayer(shape=shape)
l_enc_conv1 = C2D(incoming=l_in,
num_filters=128,
filter_size=[5, 5],
stride=[2, 2],
pad=(2, 2),
W=initmethod(0.02),
nonlinearity=lrelu(0.2),
name='enc_conv1')
l_enc_conv2 = BN(C2D(incoming=l_enc_conv1,
num_filters=256,
filter_size=[5, 5],
stride=[2, 2],
pad=(2, 2),
W=initmethod(0.02),
nonlinearity=lrelu(0.2),
name='enc_conv2'),
name='bnorm2')
l_enc_conv3 = BN(C2D(incoming=l_enc_conv2,
num_filters=512,
filter_size=[5, 5],
stride=[2, 2],
pad=(2, 2),
W=initmethod(0.02),
nonlinearity=lrelu(0.2),
name='enc_conv3'),
name='bnorm3')
l_enc_conv4 = BN(C2D(incoming=l_enc_conv3,
num_filters=1024,
filter_size=[5, 5],
stride=[2, 2],
pad=(2, 2),
W=initmethod(0.02),
nonlinearity=lrelu(0.2),
name='enc_conv4'),
name='bnorm4')
l_enc_fc1 = BN(DL(incoming=l_enc_conv4,
num_units=1000,
W=initmethod(0.02),
nonlinearity=elu,
name='enc_fc1'),
name='bnorm_enc_fc1')
l_enc_mu, l_enc_logsigma = [
BN(DL(incoming=l_enc_fc1,
num_units=cfg['num_latents'],
nonlinearity=None,
name='enc_mu'),
name='mu_bnorm'),
BN(DL(incoming=l_enc_fc1,
num_units=cfg['num_latents'],
nonlinearity=None,
name='enc_logsigma'),
name='ls_bnorm')
]
l_Z = GaussianSampleLayer(l_enc_mu, l_enc_logsigma, name='l_Z')
l_dec_fc2 = BN(DL(incoming=l_Z,
num_units=1024 * 16,
nonlinearity=relu,
W=initmethod(0.02),
name='l_dec_fc2'),
name='bnorm_dec_fc2')
l_unflatten = lasagne.layers.ReshapeLayer(
incoming=l_dec_fc2,
shape=([0], 1024, 4, 4),
)
l_dec_conv1 = BN(DeconvLayer(incoming=l_unflatten,
num_filters=512,
filter_size=[5, 5],
stride=[2, 2],
crop=(2, 2),
W=initmethod(0.02),
nonlinearity=relu,
name='dec_conv1'),
name='bnorm_dc1')
l_dec_conv2 = BN(DeconvLayer(incoming=l_dec_conv1,
num_filters=256,
filter_size=[5, 5],
stride=[2, 2],
crop=(2, 2),
W=initmethod(0.02),
nonlinearity=relu,
name='dec_conv2'),
name='bnorm_dc2')
l_dec_conv3 = BN(DeconvLayer(incoming=l_dec_conv2,
num_filters=128,
filter_size=[5, 5],
stride=[2, 2],
crop=(2, 2),
W=initmethod(0.02),
nonlinearity=relu,
name='dec_conv3'),
name='bnorm_dc3')
l_out = DeconvLayer(incoming=l_dec_conv3,
num_filters=3,
filter_size=[5, 5],
stride=[2, 2],
crop=(2, 2),
W=initmethod(0.02),
b=None,
nonlinearity=lasagne.nonlinearities.tanh,
name='dec_out')
minibatch_discrim = MinibatchLayer(
lasagne.layers.GlobalPoolLayer(l_enc_conv4),
num_kernels=500,
name='minibatch_discrim')
l_discrim = DL(incoming=minibatch_discrim,
num_units=1,
nonlinearity=lasagne.nonlinearities.sigmoid,
b=None,
W=initmethod(),
name='discrimi')
return {
'l_in': l_in,
'l_out': l_out,
'l_mu': l_enc_mu,
'l_ls': l_enc_logsigma,
'l_latents': l_Z,
'l_introspect': [l_enc_conv1, l_enc_conv2, l_enc_conv3, l_enc_conv4],
'l_discrim': l_discrim
}
def get_model(dnn=True):
if dnn:
import lasagne.layers.dnn
from lasagne.layers.dnn import Conv2DDNNLayer as C2D
from theano.sandbox.cuda.basic_ops import (as_cuda_ndarray_variable,
host_from_gpu,
gpu_contiguous, HostFromGpu,
gpu_alloc_empty)
from theano.sandbox.cuda.dnn import GpuDnnConvDesc, GpuDnnConv, GpuDnnConvGradI, dnn_conv, dnn_pool
from layers import DeconvLayer
else:
import lasagne.layers
from lasagne.layers import Conv2DLayer as C2D
dims, n_channels, n_classes = tuple(cfg['dims']), cfg['n_channels'], cfg['n_classes']
shape = (None, n_channels)+dims
l_in = lasagne.layers.InputLayer(shape=shape)
l_enc_conv1 = C2D(
incoming = l_in,
num_filters = 128,
filter_size = [5,5],
stride = [2,2],
pad = (2,2),
W = initmethod(0.02),
nonlinearity = lrelu(0.2),
flip_filters=False,
name = 'enc_conv1'
)
l_enc_conv2 = BN(C2D(
incoming = l_enc_conv1,
num_filters = 256,
filter_size = [5,5],
stride = [2,2],
pad = (2,2),
W = initmethod(0.02),
nonlinearity = lrelu(0.2),
flip_filters=False,
name = 'enc_conv2'
),name = 'bnorm2')
l_enc_conv3 = BN(C2D(
incoming = l_enc_conv2,
num_filters = 512,
filter_size = [5,5],
stride = [2,2],
pad = (2,2),
W = initmethod(0.02),
nonlinearity = lrelu(0.2),
flip_filters=False,
name = 'enc_conv3'
),name = 'bnorm3')
l_enc_conv4 = BN(C2D(
incoming = l_enc_conv3,
num_filters = 1024,
filter_size = [5,5],
stride = [2,2],
pad = (2,2),
W = initmethod(0.02),
nonlinearity = lrelu(0.2),
flip_filters=False,
name = 'enc_conv4'
),name = 'bnorm4')
l_enc_fc1 = BN(DL(
incoming = l_enc_conv4,
num_units = 1000,
W = initmethod(0.02),
nonlinearity = elu,
name = 'enc_fc1'
),
name = 'bnorm_enc_fc1')
l_enc_mu,l_enc_logsigma = [BN(DL(incoming = l_enc_fc1,num_units=cfg['num_latents'],nonlinearity = None,name='enc_mu'),name='mu_bnorm'),
BN(DL(incoming = l_enc_fc1,num_units=cfg['num_latents'],nonlinearity = None,name='enc_logsigma'),name='ls_bnorm')]
l_Z = GaussianSampleLayer(l_enc_mu, l_enc_logsigma, name='l_Z')
l_dec_fc2 = BN(DL(
incoming = l_Z,
num_units = 1024*16,
nonlinearity = relu,
W=initmethod(0.02),
name='l_dec_fc2'),
name = 'bnorm_dec_fc2')
l_unflatten = lasagne.layers.ReshapeLayer(
incoming = l_dec_fc2,
shape = ([0],1024,4,4),
)
if dnn:
l_dec_conv1 = BN(DeconvLayer(
incoming = l_unflatten,
num_filters = 512,
filter_size = [5,5],
stride = [2,2],
crop = (2,2),
W = initmethod(0.02),
nonlinearity = relu,
name = 'dec_conv1'
),name = 'bnorm_dc1')
l_dec_conv2 = BN(DeconvLayer(
incoming = l_dec_conv1,
num_filters = 256,
filter_size = [5,5],
stride = [2,2],
crop = (2,2),
W = initmethod(0.02),
nonlinearity = relu,
name = 'dec_conv2'
),name = 'bnorm_dc2')
l_dec_conv3 = BN(DeconvLayer(
incoming = l_dec_conv2,
num_filters = 128,
filter_size = [5,5],
stride = [2,2],
crop = (2,2),
W = initmethod(0.02),
nonlinearity = relu,
name = 'dec_conv3'
),name = 'bnorm_dc3')
l_out = DeconvLayer(
incoming = l_dec_conv3,
num_filters = 3,
filter_size = [5,5],
stride = [2,2],
crop = (2,2),
W = initmethod(0.02),
b = None,
nonlinearity = lasagne.nonlinearities.tanh,
name = 'dec_out'
)
else:
l_dec_conv1 = SL(SL(BN(TC2D(
incoming = l_unflatten,
num_filters = 512,
filter_size = [5,5],
stride = [2,2],
crop = (1,1),
W = initmethod(0.02),
nonlinearity = relu,
name = 'dec_conv1'
),name = 'bnorm_dc1'),indices=slice(1,None),axis=2),indices=slice(1,None),axis=3)
l_dec_conv2 = SL(SL(BN(TC2D(
incoming = l_dec_conv1,
num_filters = 256,
filter_size = [5,5],
stride = [2,2],
crop = (1,1),
W = initmethod(0.02),
nonlinearity = relu,
name = 'dec_conv2'
),name = 'bnorm_dc2'),indices=slice(1,None),axis=2),indices=slice(1,None),axis=3)
l_dec_conv3 = SL(SL(BN(TC2D(
incoming = l_dec_conv2,
num_filters = 128,
filter_size = [5,5],
stride = [2,2],
crop = (1,1),
W = initmethod(0.02),
nonlinearity = relu,
name = 'dec_conv3'
),name = 'bnorm_dc3'),indices=slice(1,None),axis=2),indices=slice(1,None),axis=3)
l_out = SL(SL(TC2D(
incoming = l_dec_conv3,
num_filters = 3,
filter_size = [5,5],
stride = [2,2],
crop = (1,1),
W = initmethod(0.02),
b = None,
nonlinearity = lasagne.nonlinearities.tanh,
name = 'dec_out'
),indices=slice(1,None),axis=2),indices=slice(1,None),axis=3)
# l_in,num_filters=1,filter_size=[5,5],stride=[2,2],crop=[1,1],W=dc.W,b=None,nonlinearity=None)
minibatch_discrim = MinibatchLayer(lasagne.layers.GlobalPoolLayer(l_enc_conv4), num_kernels=500,name='minibatch_discrim')
l_discrim = DL(incoming = minibatch_discrim,
num_units = 1,
nonlinearity = lasagne.nonlinearities.sigmoid,
b = None,
W=initmethod(),
name = 'discrimi')
return {'l_in':l_in,
'l_out':l_out,
'l_mu':l_enc_mu,
'l_ls':l_enc_logsigma,
'l_Z':l_Z,
'l_introspect':[l_enc_conv1, l_enc_conv2,l_enc_conv3,l_enc_conv4],
'l_discrim' : l_discrim}