def __init__(self, in_channels, out_channels, kernel_size=1,
padding=1, stride=1, use_dcn=0, num_convs=1):
super(ConvBlockSeq, self).__init__()
if isinstance(kernel_size, int):
kernel_size = [kernel_size]*num_convs
else:
num_convs = len(kernel_size)
if isinstance(padding, int):
padding = [padding]*num_convs
if isinstance(stride, int):
stride = [stride]*num_convs
if isinstance(use_dcn, (int, bool)):
use_dcn = [use_dcn]*num_convs
channels1 = [in_channels] + [out_channels]*(num_convs-1)
channels2 = [out_channels]*num_convs
self.conv_blocks = nn.Sequential(*[
ConvBlock(c1, c2, k, p, s, dcn)
for c1, c2, k, p, s, dcn in zip(channels1, channels2, kernel_size, padding, stride, use_dcn)
])
def __init__(self, in_channels, m, n):
super(MontageFeatLayer1234, self).__init__()
self.feat_expasions = nn.ModuleList([NoopLayer()] * m + [
ConvBlock(
in_channels, in_channels, kernel_size=3, stride=1, padding=1)
for _ in range(n)
])
def __init__(self, cfg, in_channels):
super(FEModule, self).__init__()
num_convs = cfg.MODEL.RETINAPACK.FE.NUM_CONVS
self.out_channels = cfg.MODEL.RETINAPACK.FE.OUT_CHANNELS
self.reduction = ConvBlock(in_channels,
self.out_channels,
kernel_size=1,
stride=1,
padding=0)
self.conv_blocks = nn.ModuleList([
ConvBlock(self.out_channels,
self.out_channels,
kernel_size=3,
use_dcn=False,
use_gn=True,
stride=1,
padding=1,
use_act=True) for _ in range(num_convs)
])
def __init__(self, in_channels, m, n):
super(MontageFeatLayer567_3x3conv, self).__init__()
module_list = [NoopLayer()] * m
for i in range(n):
module_list.extend([
NoopLayer(),
ConvBlock(in_channels,
in_channels,
kernel_size=3,
stride=1,
padding=1),
ConvBlock(in_channels,
in_channels,
kernel_size=3,
stride=1,
padding=1),
])
module_list.append(
ConvBlock(in_channels,
in_channels,
kernel_size=3,
stride=1,
padding=1))
self.feat_expasions = nn.ModuleList(module_list)
def __init__(self, in_channels, levels, mode):
super(MontageFeatLayer, self).__init__()
self.mode = mode
self.levels = levels
assert mode in ['recursive', 'separate']
uniques, scales_cnt = torch.unique(torch.tensor(levels).int(),
sorted=True,
return_counts=True)
for i, n in zip(uniques, scales_cnt):
for j in range(n):
if j == 0:
layer = NoopLayer()
else:
layer = ConvBlock(in_channels,
in_channels,
kernel_size=3,
stride=1,
padding=1)
setattr(self, "mf_s%d_%d" % (i, j), layer)
print(self)