Closed alykhantejani closed 3 months ago
The issue appears to be in the default_partition_fn
sometimes it returns partitions shape != batch size
Ok I think I've found why this happens. so with_unique
is defaulted to True
in the Embedding
layer, this then will filter out duplicate keys and the partition function will receive less keys than is batch size, however it doesn't then project this back.
Solved as issue was with with_unique=True
System information OS Platform: Debian TF version: 2.15.0 Python version: 3.10 TFRA: built from master GPU: no
I am using a local in-process cluster to simulate a PS strategy. I am trying to train a very simple model with one embedding layer and fake data. I am getting the following error:
which is coming form
/tensorflow_recommenders_addons/dynamic_embedding/python/ops/data_flow_ops.py", line 46, in dynamic_partition
Please see below snippet for a locally reproducible example: