Which caffe version supports ceil_mode in Pooling?

[quietsmile]

Nice work!
However, I found that ceil_mode is not supported in the master version of caffe.
Is there an open-sourced caffe version that supports ceil_mode in pooing? Thanks.

[shicai]

@quietsmile pls see: https://github.com/BVLC/caffe/pull/3057/files
just add several lines of code.

[quietsmile]

Oh, this is interesting. I mean, is it compatible with cudnn_pooling implementation? Or, just to make the output a little bit larger, even though the elements in the last row/col may still be zeros?

Thanks.

[shicai]

As i know, cudnn does not support ceil mode, but you'd better check cudnn manual by yourself.
I use caffe's own pooling layer to support ceil mode.

[oujieww]

can anyone sent me your caffe? I don't know which version of caffe should be used. I have tried to use a order version without working.But the new caffe didn't have vision_layers.hpp . I don't know what to do.please help me.

[Sarah635]

New version doesnot have vision_layers.hpp. In new version the file pooling_layer.hpp is to be modified.

[JinmingZhao]

@Sarah635 I have the same problem ,how to modify on the new version?

[Sarah635]

https://github.com/BVLC/caffe/pull/3057/files

Follow this link. Add the + lines and delete the - lines.

Instead of file vision_layers.hpp,, modify the file pooling_layers.hpp. Modify these 3 files i.e. pooling_layers.hpp, pooling_layers.cpp and caffe.proto according to link above and you are done.

[JinmingZhao]

@Sarah635 Thank you very much, done!

[XUYUNYUN666]

@Sarah635 Thank you very much, done!

Need I change the code of pooling layer.cu? Thank U.

[keerthanss]

With the latest version of Caffe, you can also get around this by replacing "ceil_mode: false" with "round_mode: FLOOR". No editing of Caffe source code necessary. Refer BVLC/caffe#6282 .

[lijiansong]

@keerthanss
When I run DenseNet-201, I get the following error, it seems that caffe fail to parse the DenseNet_201.caffemodel

layer[1002].clip_param.max
F1118 11:10:19.095713 30557 upgrade_proto.cpp:97] Check failed: ReadProtoFromBinaryFile(param_file, param) Failed to parse NetParameter file: DenseNet_201.caffemodel
*** Check failure stack trace: ***
    @     0x7f3e5d2af5cd  google::LogMessage::Fail()
    @     0x7f3e5d2b1433  google::LogMessage::SendToLog()
    @     0x7f3e5d2af15b  google::LogMessage::Flush()
    @     0x7f3e5d2b1e1e  google::LogMessageFatal::~LogMessageFatal()
    @     0x7f3e5da70731  caffe::ReadNetParamsFromBinaryFileOrDie()
    @     0x7f3e5da9b4fa  caffe::Net<>::CopyTrainedLayersFromBinaryProto()
    @           0x4091ff  test()
    @           0x407310  main
    @     0x7f3e5c21f830  __libc_start_main
    @           0x407b39  _start
    @              (nil)  (unknown)

[mrgransky]

@keerthanss I tried both ceil_mode: false and round_mode: FLOOR for deploy.prototxt file in DenseNet161 in caffe-gpu installed via conda, I get the error explained here: #37 .