I’m attempting to implement object detection using the data_block APIs. I need to transform the activations from three separate layers into actual predictions. I’m trying to use analyze_pred() to do this, but somewhere along the way it’s assuming that the output of my model is a tensor with dims (batch_size, …, …, …). Why can’t analyze_pred get passed arbitrary outputs from the model like a list of tensors? Any pointers on where I could implement this? Keeping the tensors for each layer separate makes the implementation of the loss function much simpler.