search

sanghoon / pva-faster-rcnn

Demo code for PVANet
https://arxiv.org/abs/1611.08588
Other
651 stars 241 forks source link

Ask about deconvolution #6

Closed ghost closed 7 years ago

ghost commented 7 years ago

layer { name: "upsample" type: "Deconvolution" bottom: "conv5_4" top: "upsample" param { lr_mult: 0 decay_mult: 0} convolution_param { num_output: 384 kernel_size: 4 pad: 1 stride: 2 group: 384 weight_filler: {type: "bilinear" } bias_term: false }
}

My question is why the layer's lr_mult is 0?

And If the lr_mult is 0, why not use "reshape" to get upsample.

thanks !

sanghoon commented 7 years ago

Hi,

Here, the deconv. layer is initialized to be a bilinear upsampler and works just like it.

As far as I know, 'reshape' lets us manipulate blob dimensions without increasing or reducing the amount of data. (e.g. 32x32x3 image blob => a vector with 3072 values) Therefore, it can't handle image or feature re-sizing.

Any corrections are welcome

ghost commented 7 years ago

@sanghoon Thanks，I forgot the "reshape" layer 's usuage. You are right!

good job!