I believe that the variance reduction and biais correction placeholders (in the "scsg.py" file) are mixed up... Instead of : g+= self._bias_correction_ph[i] - self._variance_reduction_ph[i] , I tried g -= self._bias_correction_ph[i] - self._variance_reduction_ph[i] in the _compute_single_update method; and it yielded better results (but not by much).
It seems that the wrong value is assoicadted with the wrong placeholder in the batch_updatemethod when you do the _feed_dict.update.
Also I have fixed what needed to be fixed to make it work on python3, I'd be happy to share it with you if you'd like.
I believe that the variance reduction and biais correction placeholders (in the "scsg.py" file) are mixed up... Instead of :
g+= self._bias_correction_ph[i] - self._variance_reduction_ph[i]
, I triedg -= self._bias_correction_ph[i] - self._variance_reduction_ph[i]
in the_compute_single_update
method; and it yielded better results (but not by much).It seems that the wrong value is assoicadted with the wrong placeholder in the
batch_update
method when you do the_feed_dict.update
.Also I have fixed what needed to be fixed to make it work on python3, I'd be happy to share it with you if you'd like.
Thanks 👍