fully-convolution layer settings

[KissyZhou]

hi,
why the transformed model, named deploy_fc.prototxt, set the 'deepid-conv' layer to be 'num_output :160 kernel_size : 1' ?
according to this, i think the kernel size in 'deepid-conv' layer should be 3. cause the output size of 'pool4' layer is 80x3x3 and the parameter size of 'deepid' layer is 160 * 720(115200), however parameter size in the 'deepid-conv' layer is 160 x 80 x 1 x 1 (12800 not equal to 115200)... any suggestions about this??
thx in advance.

[chenriwei]

hi, @KissyZhou, IMO, we should only care the full-connect layer itself, as you know, full-connect layer can be seen as a convolution layer , which kernel_size be 1.

[kli-casia]

同意 @KissyZhou 的说法，第一个全连接层改成conv层kernel大小应该是3x3，而第二个全连接层改成conv层kernel大小是1x1