Patrick Orlando

ScaNN is only used for efficient retrieval. It can only help with speeding up evaluation and has no impact on training step speed.

There's not much specific advice I can give if you are running on CPU only, other than the best practices around using the `tf.data.Dataset` API with parallelism. Moving your lookups...

@hkristof03, Your implementation of the Cross block is correct. The output of layer `x_(l+1)` of a Cross stack is the output of the previous layer, `x_l`, crossed with `x_0`. In...

Hi @zeroonesfas, A tip for formatting your issue. Your code is hard to read. I'd suggest you include your code in a code block with syntax highlighting, using the following...

Great @zeroonesfas! This is much easier. It looks like an issue when your building the index. If we look at the call signature below https://github.com/tensorflow/recommenders/blob/main/tensorflow_recommenders/layers/factorized_top_k.py#L174-L195 The function expects a tuple...

Since your model depends on the `user_id` and has no additional user features, you won't be able to make predictions for users who are not in your training set. Cold-start...

@Lulu20220 The custom model defined in the tutorial overrides the default keras train and test steps so passing a metric to `compile` will not work. You need to pass the...

Hi @houghtonweihu, @rlcauvin is correct. In the case of the Movielens dataset, you would expect the ranking model to result in better performance on your chosen metrics. Like ROC-AUC, Hitrate@K...

Keras automatically casts input to layers depending on their dtype. The issue is that the `Retrieval` task is a keras layer for which the dtype is set as the global...

I haven't tried label smoothing myself @OmarMAmin and I haven't seen any literature on label smoothing in two tower models. However, you can write your own implementations of `RemoveAccidentalHits` and...