CornerNet
CornerNet copied to clipboard
princeton-vl
Could you please explain why nstack == 2?
Dear contributor, from kp.py, I see lines( I attached it at the end) which make me confused.
Can It be interpreted as "The backbone network has a small PRE(two layers) component, followed by 2 paralleled branches, each of which predicts corners respectively?"
Cheers, Ross
` class kp(nn.Module): def init( self, n, nstack, dims, modules, out_dim, pre=None, cnv_dim=256, make_tl_layer=make_tl_layer, make_br_layer=make_br_layer, make_cnv_layer=make_cnv_layer, make_heat_layer=make_kp_layer, make_tag_layer=make_kp_layer, make_regr_layer=make_kp_layer, make_up_layer=make_layer, make_low_layer=make_layer, make_hg_layer=make_layer, make_hg_layer_revr=make_layer_revr, make_pool_layer=make_pool_layer, make_unpool_layer=make_unpool_layer, make_merge_layer=make_merge_layer, make_inter_layer=make_inter_layer, kp_layer=residual ): super(kp, self).init()
self.nstack = nstack
self._decode = _decode
curr_dim = dims[0]
self.pre = nn.Sequential(
convolution(7, 3, 128, stride=2),
residual(3, 128, 256, stride=2)
) if pre is None else pre
self.kps = nn.ModuleList([
kp_module(
n, dims, modules, layer=kp_layer,
make_up_layer=make_up_layer,
make_low_layer=make_low_layer,
make_hg_layer=make_hg_layer,
make_hg_layer_revr=make_hg_layer_revr,
make_pool_layer=make_pool_layer,
make_unpool_layer=make_unpool_layer,
make_merge_layer=make_merge_layer
) for _ in range(nstack) # nstack : 2, so from here, we get two branches
])
self.cnvs = nn.ModuleList([
make_cnv_layer(curr_dim, cnv_dim) for _ in range(nstack)
])
self.tl_cnvs = nn.ModuleList([
make_tl_layer(cnv_dim) for _ in range(nstack)
])
self.br_cnvs = nn.ModuleList([
make_br_layer(cnv_dim) for _ in range(nstack)
])
## keypoint heatmaps
self.tl_heats = nn.ModuleList([
make_heat_layer(cnv_dim, curr_dim, out_dim) for _ in range(nstack)
])
self.br_heats = nn.ModuleList([
make_heat_layer(cnv_dim, curr_dim, out_dim) for _ in range(nstack)
])
## tags
self.tl_tags = nn.ModuleList([
make_tag_layer(cnv_dim, curr_dim, 1) for _ in range(nstack)
])
self.br_tags = nn.ModuleList([
make_tag_layer(cnv_dim, curr_dim, 1) for _ in range(nstack)
])
for tl_heat, br_heat in zip(self.tl_heats, self.br_heats):
tl_heat[-1].bias.data.fill_(-2.19)
br_heat[-1].bias.data.fill_(-2.19)
self.inters = nn.ModuleList([
make_inter_layer(curr_dim) for _ in range(nstack - 1)
])
self.inters_ = nn.ModuleList([
nn.Sequential(
nn.Conv2d(curr_dim, curr_dim, (1, 1), bias=False),
nn.BatchNorm2d(curr_dim)
) for _ in range(nstack - 1)
])
self.cnvs_ = nn.ModuleList([
nn.Sequential(
nn.Conv2d(cnv_dim, curr_dim, (1, 1), bias=False),
nn.BatchNorm2d(curr_dim)
) for _ in range(nstack - 1)
])
self.tl_regrs = nn.ModuleList([
make_regr_layer(cnv_dim, curr_dim, 2) for _ in range(nstack)
])
self.br_regrs = nn.ModuleList([
make_regr_layer(cnv_dim, curr_dim, 2) for _ in range(nstack)
])
self.relu = nn.ReLU(inplace=True)
`
nstack==2 means there are two top_left branch and two bottom right branch.
@tanhangkai Why CornerNet need two top_left branch and two bottom right branch？？Why do you do this? What are the benefits of doing so?
