Differences between Caffe Deconvolution and tf.nn.cnv2d_transpose

[sharma...@gmail.com]

Hi All,

I was trying to rewrite the caffe code for a problem in Tensorflow to use as a building block for my network pipeline. The Caffe network uses the Deconvolution operation. The tensorflow equivalent of the same is tf.nn.conv2d_transpose. However, when I use the Deconvolution filters from caffe with this conv2d_transpose layer then I get very different results than what I get when I use the same filters with the Deconvolution layer in caffe.

I feel there could be one of these 2 issues which create a problem. Either the Deconvolution layer and conv2d_transpose have very different implementations or the caffe weights t deconvolution filter can't directly be used for the tensorflow layer.

If its case 1, then can you please suggest an alternate way to replicate the results of the Deconvolution operation and if it's case 2, what modifications should I make to the caffe filter for it to produce the same results as the Deconvolution layer.

Thanks a lot in advance 

[toney...@gmail.com]

hi, I also want to convert tensorflow model to caffe, which has contained deconv layer. Could I know your answer of this problem and your solution?

在 2016年9月2日星期五 UTC+8下午3:34:26，sharma...@gmail.com写道：

[samp...@gmail.com]

ping, any word on this? Can you directly convert Caffe weights for deconv to TF?