STN for only Attention Mechanism (Isotropic Scaling)

[akshath123]

I was searching for STN implementations in GitHub and came across yours. I have a few queries regarding STN implementation only for the Attention mechanism which have fixed isotropic scaling say 0.5 and the localization network predicts only translation parameters (tx and ty).

Queries:

1. If you use the same Spatial Transformer module written by you will it work?
2. Should I use the localization network to predict only two parameters (tx and ty)?
3. I bring the theta of shape (2, ) to (2, 3) by following the steps below,
   a. [tx, ty] * [0, 0, 1] --> [[0, 0, tx], [0, 0, ty]]
   b. [[0, 0, tx], [0, 0, ty]] + [[0.5, 0, 0], [0, 0.5, 0]] --> [[0.5, 0, tx], [0, 0.5, ty]]
   will it still be differentiable?
   Followed by a spatial transformer network.
   Will this work?