Shear Paramterization Fix

[ekagra-ranjan]

Fixes the bug mentioned in #1185 using the solution proposed by @jacobhinkle.

[ekagra-ranjan]

@fmassa @vfdev-5 @ptrblck @jacobhinkle I did as agreed upon in #1185 but the result doesn't seem to consistent with the original shear. For e.g., using

RSS = [ cos(a+shear_y)/cos(shear_y)  cos(a+shear_y)tan(shear_x)/cos(shear_y) - sin(a) ]
      [ sin(a+shear_y)/cos(shear_y)  sin(a+shear_y)tan(shear_x)/cos(shear_y) + cos(a) ]
  

leads to shearing in -ve x-axis sense. When I changed it to:

RSS = [ cos(a+shear_y)/cos(shear_y)  cos(a+shear_y)tan(-shear_x)/cos(shear_y) - sin(a) ]
      [ sin(a+shear_y)/cos(shear_y)  sin(a+shear_y)tan(shear_x)/cos(shear_y) + cos(a) ]
 

the shearing in x-axis becomes consistent with the original implementation. However, I am not sure whether it is the right thing to do and if RSS still needs to be modified to behave properly. What form of RSS should be chosen for this?

[ekagra-ranjan]

Also, from the code it seems shear in x-axis is denoted by shear[0] whereas shearing in y-axis is given by shear[1]. Please correct me if I am wrong in this regard.

[fmassa]

Thanks for the PR @ekagra-ranjan !

However, I am not sure whether it is the right thing to do and if RSS still needs to be modified to behave properly. What form of RSS should be chosen for this?

I didn't check the math yet, and I'd like to refer to @jacobhinkle to chime on that first.

Also, from the code it seems shear in x-axis is denoted by shear[0] whereas shearing in y-axis is given by shear[1]. Please correct me if I am wrong in this regard.

From Wolfram Alpha simplifications for the inverse of the RSS matrix (in the current format in torchvision), it seems indeed that shear[0] denotes the x axis, and shear[1] the y axis.

[fmassa]

Can you revert those space changes? It will make the documentation render incorrectly.

[fmassa]

Superseeded by #1529

Thanks @ekagra-ranjan for the original PR!