tfrs.metrics.FactorizedTopK

[soheil-asgari]

hi when i used this code i seen this error

    self.task = tfrs.tasks.Retrieval(
        metrics=tfrs.metrics.FactorizedTopK(
            candidates=(
                movies.batch(128)
                .cache()
                .map(lambda title: (title, self.movie_model(title)))
            )
        )
    )

error:

ValueError: Cannot convert '('c', 'o', 'u', 'n', 't', 'e', 'r')' to a shape. Found invalid entry 'c' of type '<class 'str'>'. Exception ignored in: <function AtomicFunction.del at 0x7d6fc6688d60>

[Lopera47]

Hello @soheil-asgari I got the same error and solved it. It is a dependency issue with the latest version of TensorFlow. In my case, in order to solve it, I installed TFRS without dependencies (pip install tensorflow-recommenders --no-deps), because I already had the needed version of tensorflow (2.11.0) and other dependencies.

Here's another similar approach: https://stackoverflow.com/questions/78144515/error-initializing-factorizedtopk-in-tensorflow-recommenders-on-sagemaker-cann

[soheil-asgari]

Hello @Lopera47 I also installed the version (2.11.0) of TensorFlow, but when I install scann, it automatically updates the version of TensorFlow to the latest version.

[Lopera47]

@soheil-asgari are you explicitly using Scann?, in my case I'm not. Maybe, you can try installing Scann without dependencies to not make it update TensorFlow.

[rlcauvin]

This is a bug in the implementation of the tfrs.layers.factorized_top_k module.

In recommenders/tensorflow_recommenders/layers/factorized_top_k.py, we find:

self._counter = self.add_weight("counter", dtype=tf.int32, trainable=False)

Note that the first argument passed to add_weight is the “counter” string.

In the Keras 2.15.0 implementation of add_weight, we find:

def add_weight(
    self,
    name=None,
    shape=None,
    dtype=None,
    initializer=None,
    regularizer=None,
    trainable=None,
    constraint=None,
    use_resource=None,
    synchronization=tf.VariableSynchronization.AUTO,
    aggregation=tf.VariableAggregation.NONE,
    **kwargs,
):

Note that first argument is the name of the weight variable. Passing “counter” as the name in this manner works fine in Keras 2.15.0, and we won’t get the ValueError: Cannot convert '('c', 'o', 'u', 'n', 't', 'e', 'r')' to a shape error.

But the Keras 3.1.1 implementation of add_weight expects shape as the first argument:

def add_weight(
    self,
    shape=None,
    initializer=None,
    dtype=None,
    trainable=True,
    regularizer=None,
    constraint=None,
    name=None,
):

So when it goes on to check for a valid shape, the shape is set to “counter”, which is not valid, and it results in ValueError: Cannot convert '('c', 'o', 'u', 'n', 't', 'e', 'r')' to a shape.

The bug, therefore, is in TensorFlow Recommenders’ implementation of the tfrs.layers.factorized_top_k module.

[Chm-vinicius]

I can confirm that is a bug with tensorflow2.16.0 and tensorflow-recommenders0.7.3 with the output error: Cannot convert '('c', 'o', 'u', 'n', 't', 'e', 'r')' to a shape

I only was able to perform a training by downgrading tensorflow to 2.15.1 version, but i can't save model with signature_inputs, the kernel get in infinty loop, this behavior is refered on this issue

[rlcauvin]

Seems to me that this issue should be marked as a bug in TensorFlow Recommenders (see my diagnosis of exactly where the bug in the code is) and fixed ASAP. Maybe @soheil-asgari can add the Bug label?

[soheil-asgari]

Hello @rlcauvin, unfortunately, I cannot add label bugs

[celha]

I'm facing the same error. Also with tensorflow version 2.15. But only for tensorflow-recommenders version 0.7.2. Any ideas?

[rlcauvin]

For now, I've worked around the TensorFlow Recommenders bug by including the following code in my notebook before installing any TensorFlow related packages:

import os
os.environ['TF_USE_LEGACY_KERAS'] = '1'

My notebook installs tensorflow-recommenders version 0.7.3 and tensorflow version 2.16.1.

I've been unable to install and import tensorflow-ranking, however, without reverting to a prior version of tensorflow.

[celha]

Alright! Thank you @rlcauvin :)

[y-71]

what worked for me is

python3.7

and my requirements.txt is

Tensorflow==2.10.0
tensorflow_recommenders==0.7.2
tensorflow-datasets
scann