I am using a neural network with a data type MPSDataTypeFloat32 but it seems to be that the final MPSCNNSoftMaxNode for the inference run returns an image using MPSDataTypeFloat16. Is this always the case? How do the nodes determine the output data type? Can I change them?
I am going to answer my own question: when using graphs (of MPSNNGraph type), setting the graph's format to MPSImageFeatureChannelFormatFloat32 or to any other suitable type does the job.