def resnet34():
class Dummy():
inplanes = n_features >> 3
r = Dummy()
layer = lambda *a, **b: ResNet._make_layer(r, BasicBlock, *a, **b)
return [
T('x', 'x', lambda x: x.permute(0, 3, 1, 2)),
T(
'x',
'x',
nn.Sequential(
Conv2d(n_ch,
n_features >> 3,
kernel_size=(7, 7),
padding=(3, 3),
bias=False),
layer(n_features >> 3, 3),
layer(n_features >> 2, 4, stride=2),
layer(n_features >> 1, 6, stride=2),
layer(n_features, 3, stride=2),
nn.AdaptiveAvgPool2d((None, 1)), # avgpool the freq. dim.
)),
T('x', 'x', lambda x: x.squeeze(3)),
T('x', 'x', lambda x: x.transpose(1, 2)),
]
def make_res_layer(block, inplanes, planes, blocks, norm_layer=nn.BatchNorm2d, base_width=64, groups=1, stride=1,
dilation=1, dilate=False, **kwargs) -> nn.Module:
"""
Args:
block: Module class. For example `BasicBlock` or `Bottleneck`.
inplanes: Number of in planes
planes: Number of planes
blocks: Number of blocks
norm_layer: Norm Module class
base_width: Base width. Acts as a factor of the bottleneck size of the Bottleneck block and is used with groups.
groups:
stride:
dilation:
dilate:
Returns:
"""
d = Dict(inplanes=inplanes, _norm_layer=norm_layer, base_width=base_width,
groups=groups, dilation=dilation) # almost a ResNet
return RN._make_layer(self=d, block=block, planes=planes, blocks=blocks, stride=stride, dilate=dilate)
