In the evaluation function for QM9, you are using a rescaled loss, which calls this function. The argument is denormalize=FLAGS.task, so it will report the denormalized result.
If so, then this rescaled loss is unit * std * MAE(y, \hat y), while our goal is to evaluate unit * MAE(y, \hat y). Can you help confirm this?
==========
Follow-ups.
I figured out that the normalizing step is in the get_item function. No questions then.
Hi there,
Thank you for sharing the code.
I have one minor question on the QM9 experiments.
In the evaluation function for QM9, you are using a rescaled loss, which calls this function. The argument is
denormalize=FLAGS.task
, so it will report the denormalized result.If so, then this rescaled loss is
unit * std * MAE(y, \hat y)
, while our goal is to evaluateunit * MAE(y, \hat y)
. Can you help confirm this?==========
Follow-ups.
I figured out that the normalizing step is in the
get_item
function. No questions then.