Filter dilation for pooling layer and how to set the output size

[XiaobingSuper]

Hi, I have some questions about mkldnn pooling, mainly about max_pooling.

1. Is there support filter dilation for max_pooling ?

2. What is the model used when setting the output size, ceil model or floor model? for example, we can chose the model by the parameter ceil_model in pytorch-maxpool2d, which can get different output size using different mode. But in mkldnn, there will be error when using floor model and ceil model is ok.

Thanks.

[vpirogov]

Intel MKL-DNN supports dilation in convolutions only, not in pooling. We are also not aware of any cases where dilation is used in pooling layers. Do you have any example models or a paper that describes how it's used?

The output shape for pooling in Intel MKL-DNN can be controlled using padding parameters. The library does not have ceil or floor flags, so I'm not exactly sure what errors you are referring to.

[XiaobingSuper]

For the first question, I didn't find related papers now, but you can see this issue saying why we should add dilation in pooling, I have found that many frameworks supporting dilation pooling, such as pytorch and tensorflow.

For last question, it is my fault because I didn't consider to use the padding parameters to controll the output shape. Thanks!

[vpirogov]

Thanks! I'm adding 'enhancement' label to track the dilated pooling request.

[jignparm]

Any update or ETA on supporting dilations in the pooling operations?

The latest version of the ONNX operators have added ceil_mode and dilations to the MaxPool operator (see below)

https://github.com/onnx/onnx/blob/master/docs/Operators.md#MaxPool

[XiaobingSuper]

@jignparm, the ceil model can be support by changing the padding size of MKLDNN MaxPooling padding to get the given output size. The dilations to MaxPool is not supported now. Thanks！

[XiaobingSuper]

@vpirogov, Any progress? PyTorch Community pytorch/pytorch#34675 said Max_pool2d is slow in native CPU path, so I think if MKLDNN can support it, may be can get a good performance.

[vpirogov]

Hi @XiaobingSuper,

We did not get to evaluation of this functionality yet. Considering that there's interest from Pytorch community I'm bumping priority of this request.

[vpirogov]

Filter dilation for pooling is introduced in dc68611.