I have the following small RRDB module network:
class RRDB(nn.Module):
"""Residual in Residual Dense Block."""
def __init__(
self,
nc: int,
kernel_size: int = 3,
gc: int = 32,
stride: int = 1,
bias: bool = True,
pad_type: Optional[blocks.PAD_TYPES_T] = "zero",
norm_type: Optional[blocks.NORM_TYPES_T] = None,
act_type: Optional[blocks.ACT_TYPES_T] = "leakyrelu",
mode: blocks.CONV_MODE_T = "CNA",
plus: bool = False
):
super().__init__()
self.RDB1 = ResidualDenseBlock5C(nc, kernel_size, gc, stride, bias, pad_type, norm_type, act_type, mode, plus)
self.RDB2 = ResidualDenseBlock5C(nc, kernel_size, gc, stride, bias, pad_type, norm_type, act_type, mode, plus)
self.RDB3 = ResidualDenseBlock5C(nc, kernel_size, gc, stride, bias, pad_type, norm_type, act_type, mode, plus)
def forward(self, x):
out = self.RDB1(x)
out = self.RDB2(out)
out = self.RDB3(out)
# Empirically, we use 0.2 to scale the residual for better performance
return out.mul(0.2) + x
Within it, you see I create three RDB block instances within it. If I keep them as self.RDBn it works correctly. If I lowercase them as self.rdbn, it returns incorrect output data noticeably.
I have absolutely no idea what could be causing this. I looked and the initialized structure of the instances matches both ways, but the output in forward does not.
Here’s a comparison of both outputs. I’m consistently getting an increase in green, even on sources that are mute of color.
