1) Could WIT just default to grabbing the first tensor in case predict_output_tensor isn't set? That will match many Keras models.
2) In the case there are multiples, could we have a smarter guess? Keras Sequential will use "output_1", or "dense_1" in many cases. Not sure what estimators will do.
3) Finally, if it's set incorrectly, could we have a better error message? Right now it's a black box called "probabilities"
I noticed in https://github.com/tensorflow/tensorboard/blob/master/tensorboard/plugins/interactive_inference/witwidget/notebook/base.py#L510 that it expects a few hard-coded tensor names for output. A few issues here:
1) Could WIT just default to grabbing the first tensor in case predict_output_tensor isn't set? That will match many Keras models. 2) In the case there are multiples, could we have a smarter guess? Keras Sequential will use "output_1", or "dense_1" in many cases. Not sure what estimators will do. 3) Finally, if it's set incorrectly, could we have a better error message? Right now it's a black box called "probabilities"
We may want to make this a required param, too.