3

I'm converting an autoencoder model from Caffe to TensorFlow. After transferring the weights, the network performs as expected through the convolutions until it reaches the deconvolutions. The output feature maps are different than expected with the given inputs and weights from Caffe. Does anyone know why I cannot transfer the weights from Caffe to TF for deconv layers?

The only thing I've been able to find online is this dead discussion with the exact same issue: https://groups.google.com/a/tensorflow.org/forum/#!topic/discuss/Dw4PmlRfaaM

Thanks!

Sam P
  • 1,821
  • 13
  • 26
  • I'm not familiar with the internal implementation of caffe/TF but is it possible you need to transpose the weights of deconv layer when transferring model? – Shai Nov 07 '17 at 08:04

0 Answers0