def __init__(self, nlatentdim, nClasses, gpu_ids, **kwargs):
super(EncD, self).__init__()
self.nfc = 1024
self.noise_std = 0
for key in ('noise_std', 'nfc'):
if key in kwargs:
setattr(self, key, kwargs[key])
self.noise = torch.zeros(0)
self.nClasses = nClasses;
self.gpu_ids = gpu_ids
nfc = self.nfc
self.main = nn.Sequential(
spectral_norm(nn.Linear(nlatentdim, nfc)),
# nn.BatchNorm1d(nfc),
nn.LeakyReLU(0.2, inplace=True),
spectral_norm(nn.Linear(nfc, nfc)),
nn.BatchNorm1d(nfc),
nn.LeakyReLU(0.2, inplace=True),
spectral_norm(nn.Linear(nfc, 512)),
nn.BatchNorm1d(512),
nn.LeakyReLU(0.2, inplace=True),
spectral_norm(nn.Linear(512, nClasses))
)
def __init__(self, nout, nch, gpu_ids, **kwargs):
super(DecD, self).__init__()
self.noise_std = 0
for key in ('noise_std'):
if key in kwargs:
setattr(self, key, kwargs[key])
self.noise = torch.zeros(0)
self.gpu_ids = gpu_ids
self.fcsize = 2
self.noise = torch.zeros(0)
self.start = spectral_norm(nn.Conv2d(nch, 64, ksize, dstep, 1))
layer_sizes = (2**np.arange(6, 12))
layer_sizes[layer_sizes>1024] = 1024
self.poolBlocks = nn.ModuleList([])
self.transitionBlocks = nn.ModuleList([])
for i in range(1, len(layer_sizes)):
self.poolBlocks.append(nn.AvgPool2d(2**i, stride=2**i))
self.transitionBlocks.append(nn.Sequential(
nn.LeakyReLU(0.2, inplace=True),
spectral_norm(nn.Conv2d(int(layer_sizes[i-1]+nch), int(layer_sizes[i]), ksize, dstep, 1)),
nn.BatchNorm2d(int(layer_sizes[i])),
))
self.end = nn.LeakyReLU(0.2, inplace=True)
self.fc = nn.Linear(1024*int(self.fcsize), nout)
def __init__(self, nLatentDim, nClasses, nRef, nch, gpu_ids):
super(Enc, self).__init__()
self.gpu_ids = gpu_ids
self.fcsize = 2
self.nLatentDim = nLatentDim
self.nClasses = nClasses
self.nRef = nRef
self.main = nn.Sequential(
spectral_norm(nn.Conv3d(nch, 64, ksize, dstep, 1)),
nn.BatchNorm3d(64),
nn.ReLU(inplace=True),
spectral_norm(nn.Conv3d(64, 128, ksize, dstep, 1)),
nn.BatchNorm3d(128),
nn.ReLU(inplace=True),
spectral_norm(nn.Conv3d(128, 256, ksize, dstep, 1)),
nn.BatchNorm3d(256),
nn.ReLU(inplace=True),
spectral_norm(nn.Conv3d(256, 512, ksize, dstep, 1)),
nn.BatchNorm3d(512),
nn.ReLU(inplace=True),
spectral_norm(nn.Conv3d(512, 1024, ksize, dstep, 1)),
nn.BatchNorm3d(1024),
nn.ReLU(inplace=True),
spectral_norm(nn.Conv3d(1024, 1024, ksize, dstep, 1)),
nn.BatchNorm3d(1024),
nn.ReLU(inplace=True)
)
if self.nClasses > 0:
self.classOut = nn.Sequential(
spectral_norm(nn.Linear(1024*int(self.fcsize*1*1), self.nClasses)),
# nn.BatchNorm1d(self.nClasses),
nn.LogSoftmax()
)
if self.nRef > 0:
self.refOut = nn.Sequential(
spectral_norm(nn.Linear(1024*int(self.fcsize*1*1), self.nRef)),
# nn.BatchNorm1d(self.nRef)
)
if self.nLatentDim > 0:
self.latentOut = nn.Sequential(
spectral_norm(nn.Linear(1024*int(self.fcsize*1*1), self.nLatentDim)),
# nn.BatchNorm1d(self.nLatentDim)
)
def __init__(self, nLatentDim, nClasses, nRef, nch, gpu_ids):
super(Enc, self).__init__()
self.gpu_ids = gpu_ids
self.fcsize = 2
self.nLatentDim = nLatentDim
self.nClasses = nClasses
self.nRef = nRef
self.first = nn.Sequential(nn.Conv3d(nch, 64, ksize, dstep, 1),
nn.BatchNorm3d(64), nn.ReLU(inplace=True))
layer_sizes = (2**np.arange(6, 12))
layer_sizes[layer_sizes > 1024] = 1024
self.poolBlocks = nn.ModuleList([])
self.transitionBlocks = nn.ModuleList([])
for i in range(1, len(layer_sizes)):
self.poolBlocks.append(nn.AvgPool3d(2**i, stride=2**i))
self.transitionBlocks.append(
nn.Sequential(
spectral_norm(
nn.Conv3d(int(layer_sizes[i - 1] + nch),
int(layer_sizes[i]), ksize, dstep, 1)),
nn.BatchNorm3d(int(layer_sizes[i])),
nn.ReLU(inplace=True),
))
if self.nClasses > 0:
self.classOut = nn.Sequential(
nn.Linear(1024 * int(self.fcsize * 1 * 1), self.nClasses),
# nn.BatchNorm1d(self.nClasses),
nn.LogSoftmax())
if self.nRef > 0:
self.refOut = nn.Sequential(
nn.Linear(1024 * int(self.fcsize * 1 * 1), self.nRef),
# nn.BatchNorm1d(self.nRef)
)
if self.nLatentDim > 0:
self.latentOutMu = nn.Sequential(
nn.Linear(1024 * int(self.fcsize * 1 * 1), self.nLatentDim))
self.latentOutLogSigma = nn.Sequential(
nn.Linear(1024 * int(self.fcsize * 1 * 1), self.nLatentDim))
def __init__(self,
nLatentDim,
nClasses,
nRef,
nch,
gpu_ids,
output_padding=(0, 1, 0)):
super(Dec, self).__init__()
self.gpu_ids = gpu_ids
self.fcsize = 2
self.nLatentDim = nLatentDim
self.nClasses = nClasses
self.nRef = nRef
self.fc = spectral_norm(
nn.Linear(self.nLatentDim + self.nClasses + self.nRef,
1024 * int(self.fcsize * 1 * 1)))
self.main = nn.Sequential(
nn.BatchNorm3d(1024),
nn.ReLU(inplace=True),
spectral_norm(
nn.ConvTranspose3d(1024,
1024,
ksize,
dstep,
1,
output_padding=output_padding)),
nn.BatchNorm3d(1024),
nn.ReLU(inplace=True),
spectral_norm(nn.ConvTranspose3d(1024, 512, ksize, dstep, 1)),
nn.BatchNorm3d(512),
nn.ReLU(inplace=True),
spectral_norm(nn.ConvTranspose3d(512, 256, ksize, dstep, 1)),
nn.BatchNorm3d(256),
nn.ReLU(inplace=True),
spectral_norm(nn.ConvTranspose3d(256, 128, ksize, dstep, 1)),
nn.BatchNorm3d(128),
nn.ReLU(inplace=True),
spectral_norm(nn.ConvTranspose3d(128, 64, ksize, dstep, 1)),
nn.BatchNorm3d(64),
nn.ReLU(inplace=True),
spectral_norm(nn.ConvTranspose3d(64, nch, ksize, dstep, 1)),
# nn.BatchNorm3d(nch),
nn.Sigmoid())
def __init__(self, nout, nch, gpu_ids, **kwargs):
super(DecD, self).__init__()
self.noise_std = 0
for key in (['noise_std']):
if key in kwargs:
setattr(self, key, kwargs[key])
self.noise = torch.zeros(0)
self.gpu_ids = gpu_ids
self.fcsize = 2
self.noise = torch.zeros(0)
self.main = nn.Sequential(
spectral_norm(nn.Conv3d(nch, 64, ksize, dstep, 1)),
# nn.BatchNorm3d(64),
nn.LeakyReLU(0.2, inplace=True),
spectral_norm(nn.Conv3d(64, 128, ksize, dstep, 1)),
nn.BatchNorm3d(128),
nn.LeakyReLU(0.2, inplace=True),
spectral_norm(nn.Conv3d(128, 256, ksize, dstep, 1)),
nn.BatchNorm3d(256),
nn.LeakyReLU(0.2, inplace=True),
spectral_norm(nn.Conv3d(256, 512, ksize, dstep, 1)),
nn.BatchNorm3d(512),
nn.LeakyReLU(0.2, inplace=True),
spectral_norm(nn.Conv3d(512, 1024, ksize, dstep, 1)),
nn.BatchNorm3d(1024),
nn.LeakyReLU(0.2, inplace=True),
spectral_norm(nn.Conv3d(1024, 1024, ksize, dstep, 1)),
nn.BatchNorm3d(1024),
nn.LeakyReLU(0.2, inplace=True)
)
self.fc = spectral_norm(nn.Linear(1024*int(self.fcsize), nout))