DNN random seed (NCF, xDeepFM, W&D) #785
Conversation
…. Update wide-deep
|
Check out this pull request on ReviewNB: https://app.reviewnb.com/microsoft/recommenders/pull/785 Visit www.reviewnb.com to know how we simplify your Jupyter Notebook workflows. |
reco_utils/common/tf_utils.py
Outdated
| import tensorflow as tf | ||
| from tensorflow.python.estimator.inputs.queues import feeding_functions | ||
| from tensorflow.python.estimator.inputs.numpy_io import ( | ||
| _get_unique_target_key, |
There was a problem hiding this comment.
this hurts a little bit lol, but I guess it's not our fault. I guess there are no public functions that we can use right?
There was a problem hiding this comment.
ah... yeah. Not sure why they did not expose the seed variable to numpy input function.
If we really don't want to do this, alternative way is to shuffle by ourselves like:
def numpy_input_fn(x, ...):
...
if shuffle==True and seed is not None:
# set random seed and shuffle x
....
# then call tf's numpy_input_fn with the shuffled x
tf.estimator.inputs.numpy_input_fn(x)
I actually like this solution. Much cleaner.
There was a problem hiding this comment.
Ah... the thing is, w/ this, we cannot shuffle for every epoch... hmmm
There was a problem hiding this comment.
I think ‘tf.set_random_seed’ will work for this case. Let me try and will update codes if works.
There was a problem hiding this comment.
Refactored to use tf.data.dataset (newer API) within our pandas_input_fn and now much cleaner.
@loomlike, hope you don't mind me commiting directly to your branch, I think it was faster to just remove the line than to tell you to do it :-)
DNN random seed (NCF, xDeepFM, W&D)
Description
Refactor NCF (dataset and model) and W&D (input_fns) to accept random seed.
Update tests accordingly
Related Issues
#736
Checklist: