Conv parameters

[Bestupidas]

code in mmdet/models/utils/res_layer.pys shows below(line 52~58):
class ResLayer(Sequential):
...
if stride != 1 or inplanes != planes * block.expansion:
downsample = []
conv_stride = stride
...
downsample.extend([
build_conv_layer(
conv_cfg,
inplanes,
planes * block.expansion,
kernel_size=1,
stride=conv_stride,
bias=False),
build_norm_layer(norm_cfg, planes * block.expansion)[1]
])
downsample = nn.Sequential(*downsample)

my question:
the conv layer is built with kernel_size=1, stride=2 when conv_stride is 2. why conv stride is bigger than 1 when kener size is 1. is it right or has something carefully designed for specified purpose ? hope for reply

[ZwwWayne]

It is a convention when using the official implementation of ResNet in acedamia. In practice, using a larger kernel size (e.g., 3) or pooling layer could yield better performance.

[Bestupidas]

thanks ! i will do an training experiment to check the modification from kernel size 1 to 3 to see its affects for map in dections.