def __init__(self, lin_arc = lin_arc, res_arc = res_arc, depths = ([1]*(len(res_arc)+1)),
block = resnet.basic_block, *args, **kwargs):
super().__init__()
self.lin_arc = lin_arc[::-1]
self.linear = nn.ModuleList([
nn.Linear(self.lin_arc[0], self.lin_arc[1]),
*[nn.Linear(input, output)
for (input, output) in zip(self.lin_arc[1:], self.lin_arc[2:])]
])
self.res_arc = [int(res_arc[-1] / 2)] + res_arc[::-1]
self.in_out_block_sizes = list(zip(self.res_arc, self.res_arc[1:]))
self.conv = nn.ModuleList([
resnet.layer(self.res_arc[0], self.res_arc[0], n=depths[0],
block=block, conv=conv,
sampling=1, *args, **kwargs),
*[resnet.layer(in_channels * block.expansion,
out_channels, n=n, conv=conv,
block=block,
sampling=1, *args, **kwargs)
for (in_channels, out_channels), n in zip(self.in_out_block_sizes, depths[1:])],
])
self.gate = nn.Sequential(
nn.Conv1d(self.res_arc[-1], 20, kernel_size=1, stride=1, padding=0, bias=False),
nn.ReLU(),
)
def __init__(self, conv, in_channels=20, block_arch=[1024, 256, 64, 3], deepths=[2, 2, 2, 2],
block = resnet.basic_block, *args, **kwargs):
super().__init__()
self.block_arch = block_arch
self.in_out_block_sizes = list(zip(block_arch, block_arch[1:]))
self.blocks = nn.ModuleList([
resnet.layer(in_channels, block_arch[0], n=deepths[0],
block=block, conv=conv, *args, **kwargs),
*[resnet.layer(in_channels * block.expansion,
out_channels, n=n, conv=conv,
block=block, *args, **kwargs)
for (in_channels, out_channels), n in zip(self.in_out_block_sizes, deepths[1:])],
])
self.linit = nn.Linear(standard_length, bottleneck)
def __init__(self, lin_arc = lin_arc, res_arc = res_arc, depths =([1]*len(res_arc)),
block = resnet.basic_block, *args, **kwargs):
super().__init__()
self.linear = nn.ModuleList([
nn.Linear(lin_arc[0], lin_arc[1]),
*[nn.Linear(input, output)
for (input, output) in zip(lin_arc[1:], lin_arc[2:])]
])
self.in_out_block_sizes = list(zip(res_arc, res_arc[1:]))
self.conv = nn.ModuleList([
resnet.layer(in_channels, res_arc[0], n=depths[0],
block=block, conv=conv,
sampling=1, *args, **kwargs),
*[resnet.layer(in_channels * block.expansion,
out_channels, n=n, conv=conv,
sampling=1,
block=block, *args, **kwargs)
for (in_channels, out_channels), n in zip(self.in_out_block_sizes, depths[1:])],
])
def __init__(self, conv, block_arch=[3, 64, 256, 1024], deepths=[2, 2, 2, 2],
block=resnet.basic_block, *args, **kwargs):
super().__init__()
self.block_arch = block_arch
self.linit = nn.Linear(bottleneck, standard_length)
self.in_out_block_sizes = list(zip(block_arch, block_arch[1:]))
self.blocks = nn.ModuleList([
resnet.layer(block_arch[0], block_arch[0], n=deepths[0],
block=block, conv=conv, *args, **kwargs),
*[resnet.layer(in_channels * block.expansion,
out_channels, n=n, conv=conv,
block=block, *args, **kwargs)
for (in_channels, out_channels), n in zip(self.in_out_block_sizes, deepths[1:])],
])
self.gate = nn.Sequential(
nn.Conv1d(block_arch[-1], 20, kernel_size=1, stride=1, padding=0, bias=False),
nn.ReLU(),
)
def __init__(self,
lin_arc=lin_arc,
res_arc=res_arc,
block=resnet.basic_block,
*args,
**kwargs):
super().__init__()
#linear layers
self.lin_arc = lin_arc[::-1]
self.linear = nn.ModuleList([
nn.Linear(self.lin_arc[0], self.lin_arc[1]), *[
nn.Linear(input, output)
for (input, output) in zip(self.lin_arc[1:], self.lin_arc[2:])
]
])
#convolutional layers
self.res_arc = res_arc[::-1] + [in_channels]
self.conv = nn.ModuleList([
resnet.layer(in_channels * block.expansion,
out_channels,
conv=conv,
sampling=1,
block=block,
*args,
**kwargs)
for (in_channels,
out_channels) in zip(self.res_arc[0:], self.res_arc[1:])
] + [
resnet.layer(self.res_arc[-1],
in_channels,
block=block,
conv=conv,
sampling=1,
*args,
**kwargs)
])