Add operators for Deformable ConvNets/DFF #6298
Conversation
|
code looks in good shape. two general comments: |
|
Hi @piiswrong , sorry for the delay. Now test code for psroipoling, deformable convolution/psroipooling has been added. |
|
Could you also add a check_numerical_gradient test in addition to consistency checks? Its generally a good habit to check gradient. We have had people discovering their ops were outputing wrong gradient on corner cases when they submit PR. The model still trains with slightly wrong gradient but you often get better acc by correcting it. |
YuwenXiong
changed the title
|
It seems that building check always reports |
|
@orpine, @piiswrong: we would like to use this PR. What is still preventing it from being merged? |
| } | ||
|
|
||
| MXNET_REGISTER_OP_PROPERTY(_contrib_DeformableConvolution, DeformableConvolutionProp) | ||
| .describe(R"code(Compute *N*-D convolution on *(N+2)*-D input. |
There was a problem hiding this comment.
These docs seems to be copied from standard conv. Could you updated it to describe DeformableConvolution or remove inaccurate doc and insert a link to the original paper?
| DMLC_REGISTER_PARAMETER(DeformablePSROIPoolingParam); | ||
|
|
||
| MXNET_REGISTER_OP_PROPERTY(_contrib_DeformablePSROIPooling, DeformablePSROIPoolingProp) | ||
| .describe("Performs region-of-interest pooling on inputs. Resize bounding box coordinates by " |
There was a problem hiding this comment.
same. updated/remove docs and link to paper
|
@sbodenstein @orpine Will try to get it merged soon. |
| const TShape& pad, const TShape& stride, const TShape& dilation, | ||
| const uint32_t deformable_group, DType* data_col) { | ||
| if (2 == kernel_shape.ndim()) { | ||
| LOG(FATAL) << "not implemented"; |
There was a problem hiding this comment.
say "only implemented in GPU"
|
Updated, please have a check. @piiswrong |
|
merged. Thanks. |
|
Are you planning to submit the algorithms that use these ops as examples? |
|
Alternatively you can put a link to your repos in example/readme |
|
@piiswrong Thanks for your advice, I've added a link #6754 |
* Add operators for Deformable ConvNets/FCIS/DFF * fix programming rule to meet pr rule * fix programming rule to meet pr rule * fix programming rule to meet pr rule * minor fix channel operator * minor fix deformable conv * fix a stupid error * add test code * add test code * remove channel operator and add gradient check * remove redundant print * fix unittest code * dummy commit to trigger building check * dummy commit to trigger building check * Update deformable_convolution-inl.h * Update deformable_im2col.h * Update deformable_convolution.cc * Update deformable_psroi_pooling.cc
implementation from apache/mxnet#6298 for reproducibility.
No description provided.