When the input maps are float16, it's possible for the output to contain nan and inf values. This subtracts the max and converts to float32 before performing the possibly unstable calculations.
Adds a debug output statement that shows the unscaled min/max rates.
Initialize GaussianSmoothing in scfg_params to avoid recreating it every step.
Fix adding fields to wrong module (to_k/to_q instead of parent module)
When the input maps are float16, it's possible for the output to contain nan and inf values. This subtracts the max and converts to float32 before performing the possibly unstable calculations.
Adds a debug output statement that shows the unscaled min/max rates.
Initialize GaussianSmoothing in scfg_params to avoid recreating it every step.
Fix adding fields to wrong module (to_k/to_q instead of parent module)