I added some args for dilation. Therefore, the generated kernel could compute with dilated pixels. It might be helpful to expand receptive field in RedNet backbone without losing feature map size in some tasks like detection and segmentation.
For example, dilations=(1, 1, 2, 4), strides=(1, 2, 1, 1) instead of 'strides=(1, 2, 2, 2)' in configs.
I added some args for dilation. Therefore, the generated kernel could compute with dilated pixels. It might be helpful to expand receptive field in RedNet backbone without losing feature map size in some tasks like detection and segmentation. For example,
dilations=(1, 1, 2, 4), strides=(1, 2, 1, 1)instead of 'strides=(1, 2, 2, 2)' in configs.