def __init__(self, n_ch=2, nf=64, ks=3, nc=5, nd=5):
"""
:param n_ch: number of channels
:param nf: number of filters
:param ks: kernel size
:param nc: number of iterations
:param nd: number of CRNN/BCRNN/CNN layers in each iteration
"""
super(CRNN_MRI, self).__init__()
self.nc = nc
self.nd = nd
self.nf = nf
self.ks = ks
self.bcrnn = BCRNNlayer(n_ch, nf, ks)
self.conv1_x = nn.Conv2d(nf, nf, ks, padding=ks // 2)
self.conv1_h = nn.Conv2d(nf, nf, ks, padding=ks // 2)
self.conv2_x = nn.Conv2d(nf, nf, ks, padding=ks // 2)
self.conv2_h = nn.Conv2d(nf, nf, ks, padding=ks // 2)
self.conv3_x = nn.Conv2d(nf, nf, ks, padding=ks // 2)
self.conv3_h = nn.Conv2d(nf, nf, ks, padding=ks // 2)
self.conv4_x = nn.Conv2d(nf, n_ch, ks, padding=ks // 2)
self.relu = nn.ReLU(inplace=True)
dcs = []
for i in range(nc):
dcs.append(cl.DataConsistencyInKspace(norm='ortho'))
self.dcs = dcs
def __init__(self, n_channels=2, nc=5, nd=5, **kwargs):
super(DnCn3DShared, self).__init__()
self.nc = nc
self.nd = nd
print('Creating D{}C{}-S (3D)'.format(nd, nc))
self.conv_block = conv_block(n_channels, nd, **kwargs)
self.dc = cl.DataConsistencyInKspace(norm='ortho')
def __init__(self, n_channels=2, nc=5, nd=5, **kwargs):
super(DnCn, self).__init__()
self.nc = nc
self.nd = nd
print('Creating D{}C{}'.format(nd, nc))
conv_blocks = []
dcs = []
conv_layer = conv_block
for i in range(nc):
conv_blocks.append(conv_layer(n_channels, nd, **kwargs))
dcs.append(cl.DataConsistencyInKspace(norm='ortho'))
self.conv_blocks = nn.ModuleList(conv_blocks)
self.dcs = dcs
def __init__(self,
n_channels=2,
nc=5,
nd=5,
fr_d=None,
clipped=False,
mode='pytorch',
**kwargs):
"""
Parameters
----------
fr_d: frame distance for data sharing layer. e.g. [1, 3, 5]
"""
super(DnCn3DDS, self).__init__()
self.nc = nc
self.nd = nd
self.mode = mode
print('Creating D{}C{}-DS (3D)'.format(nd, nc))
if self.mode == 'theano':
print('Initialised with theano mode (backward-compatibility)')
conv_blocks = []
dcs = []
kavgs = []
if not fr_d:
fr_d = list(range(10))
self.fr_d = fr_d
conv_layer = conv_block
# update input-output channels for data sharing
n_channels = 2 * len(fr_d)
n_out = 2
kwargs.update({'n_out': 2})
for i in range(nc):
kavgs.append(
cl.AveragingInKspace(fr_d, i > 0, clipped, norm='ortho'))
conv_blocks.append(conv_layer(n_channels, nd, **kwargs))
dcs.append(cl.DataConsistencyInKspace(norm='ortho'))
self.conv_blocks = nn.ModuleList(conv_blocks)
self.dcs = nn.ModuleList(dcs)
self.kavgs = nn.ModuleList(kavgs)
