wxcstc opened a new issue #11629: deconvolution result different with caffe result URL: https://github.com/apache/incubator-mxnet/issues/11629 ## Description same params, deconvolution result different with caffe ## Environment info (Required) windows 10 python3 mxnet version "1.3.0" Package used (Python/R/Scala/Julia): I'm using Python ## Steps to reproduce 1. python code as below: `import mxnet as mx` `import mxnet.ndarray as nd` `a = nd.reshape(nd.arange(9), shape = (1, 1, 3, 3))` `b = nd.reshape(nd.arange(4), shape=(1, 1, 2, 2))` `c=mx.nd.Deconvolution(data=a, weight=b, kernel=(2,2), num_filter=1, dilate=(2, 2), stride=(2,2))` `print(c)` 2. result is [[[[ 0. 0. 0. 0. 1. 0. 2.] [ 0. 0. 0. 0. 0. 0. 0.] [ 0. 0. 0. 2. 3. 4. 6.] [ 0. 0. 3. 0. 4. 0. 5.] [ 6. 0. 9. 8. 12. 10. 15.] [ 0. 0. 6. 0. 7. 0. 8.] [12. 0. 18. 14. 21. 16. 24.]]]] <NDArray 1x1x7x7 @cpu(0)> 3. same params in caffe result is: [[[[ 0. 0. 0. 0. 1. 0. 2.] [ 0. 0. 0. 0. 0. 0. 0.] [ 0. 0. 5. 0. 11. 0. 11.] [ 0. 0. 0. 0. 0. 0. 0.] [ 6. 0. 23. 0. 29. 0. 23.] [ 0. 0. 0. 0. 0. 0. 0.] [ 12. 0. 32. 0. 37. 0. 24.]]]] 4. and caffe prototxt as below： input_shape { dim: 1 dim: 1 dim: 3 dim: 3 } layer { name: "deconv1" type: "Deconvolution" bottom: "data" top: "deconv1" convolution_param { num_output: 1 kernel_size: 2 stride_h: 2 stride_w: 2 dilation: 2 pad_h: 0 pad_w: 0 group: 1 } } ## What have you tried to solve it? 1.I calculated deconvoluiton manually，result was same with caffe 2.So，I think implemnet of mxnet maybe have bug.Or I misunderstand the behavior of deconvolution?
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
