Calling Predictions from Models

[phillipshaong]

For all the co-lab examples (except for basic_retrieval), there isn't a clear way to get predictions from the models. For basic_retrieval, the use of the FacorizedTopK metric tfrs.tasks.Retrieval(metrics=tfrs.metrics.FactorizedTopK(songs.batch(64).map(song_model)) , and use of brute force to get the index tfrs.layers.factorized_top_k.BruteForce(model.user_model) is presented and works. However, for basic_ranking, context_features, deep_recommenders, multitask, and dcn, the code example ends at evaluation of test/dev sets.

So, the question is how does one do predictions? How does one infer from the model?

An option I have tried is calling model.predict() or model(), but that gives an error. The error codes below are from code I have run after running through the entire basic_ranking model co-lab notebook.

for i in train.take(1):
  print(i)
  model.predict(i)

{'movie_title': <tf.Tensor: shape=(), dtype=string, numpy=b'Postman, The (1997)'>, 'user_id': <tf.Tensor: shape=(), dtype=string, numpy=b'681'>, 'user_rating': <tf.Tensor: shape=(), dtype=float32, numpy=4.0>}
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-68-be1677278cba> in <module>()
      1 for i in train.take(1):
      2   print(i)
----> 3   model.predict(i)
      4   model(i)

4 frames
/usr/local/lib/python3.7/dist-packages/tensorflow/python/framework/tensor_shape.py in __getitem__(self, key)
    887       else:
    888         if self._v2_behavior:
--> 889           return self._dims[key].value
    890         else:
    891           return self._dims[key]

IndexError: list index out of range

And,

for i in train.take(1):
  print(i)
  model(i)

{'movie_title': <tf.Tensor: shape=(), dtype=string, numpy=b'Postman, The (1997)'>, 'user_id': <tf.Tensor: shape=(), dtype=string, numpy=b'681'>, 'user_rating': <tf.Tensor: shape=(), dtype=float32, numpy=4.0>}
---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
<ipython-input-69-fb4f109a1076> in <module>()
      1 for i in train.take(1):
      2   print(i)
----> 3   model(i)

1 frames
/usr/local/lib/python3.7/dist-packages/tensorflow/python/keras/engine/training.py in call(self, inputs, training, mask)
    444         a list of tensors if there are more than one outputs.
    445     """
--> 446     raise NotImplementedError('When subclassing the `Model` class, you should '
    447                               'implement a `call` method.')
    448 

NotImplementedError: When subclassing the `Model` class, you should implement a `call` method.

Thanks for your help!

[phillipshaong]

I recognize that there are related issues #248 and #215, but haven't been able to implement either solutions. Is there a way to do predictions without adding additional code within the *Model classes?

[maciejkula]

The actual part of the model that's responsible for generating predictions is model.ranking_model. Have you tried calling model.ranking_model(batch)?

[pemoriarty]

@maciejkula Could you clarify what format/structure batch should be in the call model.ranking_model(batch)? I'm having the same challenge as @phillipshaon.

[Niouscha91]

@pemoriarty @phillipshaong I have the same issue as yours. I want to predict the ranking in the ranking part. Have you found any solution for that?

[pemoriarty]

I did actually get it working, but unfortunately I don't have access to the code anymore. @felixDulys, maybe you could help answer this now?

[xiaoyaoyang]

ranking_model

I think what you suggest would work and it is similar to what I did (by random gues..), I would like to know more what exactly the model call after I call model.predict (sorry I am totally new to TF keras..), seems it calls the call function..?

I did make it work by adding this line into the MovielensModel , but not sure this is a good solution though

def call(self, features: Dict\[Text, tf.Tensor\]): 
    rating_predictions = self.ranking_model(
        (features["user_id"], features["movie_title"]))
    return rating_predictions

[windmaple]

Check out this section: https://www.tensorflow.org/recommenders/examples/basic_ranking#testing_the_ranking_model

[felixDulys]

I second @xiaoyaoyang -- it looks like @pemoriarty (hi! hope you are well :-) ) used a call() method inside of the ranking model class (which is a subclass of tf.keras.Model) like:

def call(self, inputs):
      user_id, movie_id = inputs
      user_embedding = self.user_embeddings(user_id)
      movie_embedding = self.movie_embeddings(movie_id)
      return self.ratings(tf.concat([user_embedding, movie_embedding], axis=1))

Note that the initialization function in that class holds the architecture of the model. self.movie_embedding and the other embedding are defined in the initialization function in the same class and are keras.Sequential objects. Also note that the inputs are a tuple.

Then, when you have trained your model and you want to make predictions, it looks like she did some specific transformations / typing:

import itertools
import pandas as pd
import numpy as np

prediction_df = pd.DataFrame(list(itertools.product(guests, movies)))
prediction_cols = ["guest_id", "movie_id"]

prediction_dict = {
    "guest_id": np.array(tuple(prediction_df["guest_id"])),
    "movie_id": np.array(tuple(prediction_df["movie_id"]))
}

prediction_scores = model.call((prediction_dict["guest_id"], prediction_dict["movie_id"]))

Hopefully this will work for you! Apologies for the delay. Good luck!