I managed to introspect the input keys that a model expects by checking model.dummy_inputs dict (do all models define this reliably? is there a better way?)
But I am currently struggling to find a way to determine the output keys.
The distilbert model I am experimenting with returns "logits", but I understand that not all models may output that key? And also there are config flags (e.g. model.config.output_hidden_states) that would cause additional keys to be output.
I’m not sure if this is specifically a transformers question or a PyTorch question…