Add basic ranking example. #7
Conversation
examples/basic_ranking.py
Outdated
|
|
||
| def __init__( | ||
| self, | ||
| user_embeddings_count, |
There was a problem hiding this comment.
Done, although I'm keeping the term candidate instead of item everywhere for consistency.
| return ( | ||
| # Inputs are user ids and movie ids | ||
| { | ||
| "user_id": tf.strings.to_number(x["user_id"], out_type=tf.int32), |
There was a problem hiding this comment.
Are these IDs guaranteed to be sequential? If they're just random int-like, we would want to reindex them from 0 to num_users
There was a problem hiding this comment.
I cheated because I know the dataset and they are sequential.
The original example was more general because it wasn't making any assumption on the IDs, not even that they were numbers, because it's using a StringLookup layer as part of the model. StringLookup is not multi-backend so we can't do that as part of the model, it would have to be part of preprocessing. But then it gets more complicated to demonstrate inference (in the retrieval example), because the reverse lookup is needed, so we would need:
- model.predict
- convert to Tensorflow
- reverse lookup
- convert to Python
- lookup movie title
We have 3 options:
- make it fully generic and use
StringLookupoutside of the model - assume IDs are ints and use an int to int mapping
- leave it as is (assume sequential IDs)
There was a problem hiding this comment.
leave it as is (assume sequential IDs)
That's ok if you clearly document this gotcha
There was a problem hiding this comment.
Added some detailed language about this shortcut.
Example was run with all 3 backends (tf, jax and torch).
hertschuh
force-pushed
the
basic_ranking
branch
from
48878ad to
f4db280
Compare
Example was run with all 3 backends (tf, jax and torch).