[TOPI] implement pool3d op

[optima2005]

This is a part of attempts to implement #4009
The implemation in this PR is for max and avg 3d pooling.

@masahi

[masahi]

@optima2005 I am traveling for another few days, will review afte I am back.

[masahi]

thanks @optima2005 this is merged.

[deepakbabel]

Hi @optima2005 , @masahi ,
I am new to tvm community. I was trying to implement maxpool3d operator based on reference implementation from tensorflow(r=1.14) into tvm codebase. For testing the same, I have written relay and topi test cases. I am also writing test cases in test_forward.py(for tensorflow) which compares tvm output with tensorflow output. I want to know is there any specific reason why you have not added any changes in test_forward.py(comparing tvm with tf) for unit testing?? Should i ignore this file altogether?

Best Regards,
Deepak Babel

[optima2005]

Hi, @deepakbabel
This PR is just to implement a 3d pooling(max&avg) operation for relay, not including the TF frontend, which is to transform the 3d pooling operation in TF model into TVM relay operation.

[deepakbabel]

Hi @optima2005 ,
Thank you for clarifying this. I want to support TF frontend in TVM for this operator(maxpool3d), so that I can compile a TF model using TVM. I am trying to ascertain changes required in TVM front-end specific files like test_forward.py and tensorflow.py file. As of now these files have support only for 2D image data format("NCHW"). Looks like there would be some small design change to include 3D("NCDHW") or nD dimension data. Do you happen to have any inputs/suggestions on the same.

[deepakbabel]

@optima2005 -
Here are the path of the files that I am referring to:
tvm/tests/python/frontend/tensorflow/test_forward.py
tvm/python/tvm/relay/frontend/tensorflow.py

Best Regards,
Deepak Babel

[optima2005]

@deepakbabel, If I would implement 3D pooling TF frontend, I would create a new function by following the current 2D pooling function. And add it to the converting map for TF 3D pooling transformation. By this approach, the original 2d pooling won't be impacted and we won't handle too much conditionals in one function.

[deepakbabel]

@optima2005 : Thank you for the suggestion. I actually have done the same(seperate function for 3dpooling) in tensorflow.py file(Frontend support). Also added test cases in test_forward.py to test it. Thank you for your inputs. In test_forward.py(for Tensorflow Front-end), I could see that while creating the TVM Graph, they use "NCHW" layout by-default and currently there is no option to pass the "NCDHW" layout information in the same file. So, I guess we might have to set the layout from somewhere else(may be tensorflow.py in the new function for 3d pooling?), otherwise it will require design changes in existing test_forward.py file.
Anyways really appreciate your inputs on this.

Best Regards,
Deepak

[apivovarov]

pad_d and pad_h are not used. The fix is here #4738 @optima2005 @masahi