How can I fix the scale to 50% (0.5) and learn only the translation parameters?

[abhisheksgumadi]

Thanks for this wonderful implementation. I want to learn just the translation parameters and fix the scale to 50% (0.5) and keep the scale a constant during training.

How can we do this please with this implementation? I couldn't think of a way to keep the scale as a constant with an initial value.

[albanD]

Hi,

Unfortunately, this is not supported right now.

• You can modify the AffineTransformMatrixGenerator to be able to work with constants.
• You can also create a custom module, that you will place between the AffineTransformMatrixGenerator and the AffineGridGeneratorBHWD, that will apply this scaling. The forward pass of this module would be very close to this and the backward pass to this where scaleFactors is the constant value 0.5.

[abhisheksgumadi]

Thanks Alban, I got it working finally. I wrote a new Torch layer that takes the Affine matrix as input and just subtracts (0.5) from the initialised scale parameters that are not learned.