upsampling: how to set kernel size in deconvolution layer with biliear filter?

[Qingze Wang]

for upsampling, caffe's doc list below:

layer {
 name: "upsample", type: "Deconvolution"
 bottom: "{{bottom_name}}" top: "{{top_name}}"
 convolution_param {
 kernel_size: {{2 * factor - factor % 2}} stride: {{factor}}
 num_output: {{C}} group: {{C}}
 pad: {{ceil((factor - 1) / 2.)}}
 weight_filler: { type: "bilinear" } bias_term: false
 }
 param { lr_mult: 0 decay_mult: 0 }
}

but there is no comment to explain why to set kernel_size to 2 * factor - factor % 2, stride to factor, and pad to ceil((factor - 1) / 2.

so someone know inference about this?

[saqib qamar]

Have you got your answer...can you share it..