Converting caffe2 model from float32 to float16


I’d like to convert an existing caffe2 model to float16 compute (existing in the sense that I have the .pb files in hand) … it seems like the kind of thing that should be simple to do by setting a switch in the workspace or when loading the file (analogous to .set_floatx in Keras), but I’m not getting very far trying to set float16_compute in places like NetDef.ParseFromString or workspace.CreateNet… is there an easy way to do this, or do I need to explicitly step through the layers of the old model and set float16_compute on each one?