Instead of initializing our own Metric implementation (metrics._Metric) with a given scaling and correction, initialise them without those values and just do scaling * _Metric + correction when appending the metric to the list of metrics used for the calculation of the loss. For now, the affected metrics are PoissonNLLLoss and VGGLoss.
The overall goal is to move away from creating our own metrics to having them implemented in torchmetrics but this will take some time.
Instead of initializing our own
Metricimplementation (metrics._Metric) with a givenscalingandcorrection, initialise them without those values and just doscaling * _Metric + correctionwhen appending the metric to the list of metrics used for the calculation of the loss. For now, the affected metrics arePoissonNLLLossandVGGLoss.The overall goal is to move away from creating our own metrics to having them implemented in
torchmetricsbut this will take some time.