Why init weights like that?

[leftthomas]

SparseInst/sparseinst/encoder.py

Lines 61 to 62 in fd6fb38

	c2_xavier_fill(lateral_conv)
	c2_xavier_fill(output_conv)

SparseInst/sparseinst/encoder.py

Line 71 in fd6fb38

c2_msra_fill(self.fusion)

SparseInst/sparseinst/decoder.py

Lines 52 to 63 in fd6fb38

	def _init_weights(self):
	for m in self.inst_convs.modules():
	if isinstance(m, nn.Conv2d):
	c2_msra_fill(m)
	bias_value = -math.log((1 - self.prior_prob) / self.prior_prob)
	for module in [self.iam_conv, self.cls_score]:
	init.constant_(module.bias, bias_value)
	init.normal_(self.iam_conv.weight, std=0.01)
	init.normal_(self.cls_score.weight, std=0.01)
	init.normal_(self.mask_kernel.weight, std=0.01)
	init.constant_(self.mask_kernel.bias, 0.0)

SparseInst/sparseinst/decoder.py

Lines 99 to 103 in fd6fb38

	def _init_weights(self):
	for m in self.mask_convs.modules():
	if isinstance(m, nn.Conv2d):
	c2_msra_fill(m)
	c2_msra_fill(self.projection)

can you explain why init weights like this? And I find if we remove these codes, the AP will be very bad. This means that your method relies heavily on parameter initialization, so I hope you can provide a reasonable explanation.