embedding dimension change.

[xiexbing]

hello, I finished the first run impl of the tzrec based rec training. thanks for the help along the way. one thing I don't understand is, I tried to use embedding group to manage the tables, and I tried build the features as IdFeature and lookupfeature, but no matter how I configure my embedding_dim in the config, the table dimension will be 4 when I check the sharding plan. is this correct? please see my embedding config below.

feature_configs = []
# Initialize RawFeature instances for each feature
for emb_bag_name, emb_bag_config in embedding_bag_configs.items():
    feature_names = emb_bag_config.feature_names
    for feature_name in feature_names:
        feature_configs.append(
                feature_pb2.FeatureConfig(
                    lookup_feature=feature_pb2.LookupFeature(
                        feature_name=feature_name,
                        embedding_name=emb_bag_config.name,
                        embedding_dim=128,
                        pooling='mean',
                        num_buckets=emb_bag_config.num_embeddings
                        )
                    )
                )

    # Create features
features = create_features(feature_configs)
# Define feature groups
feature_groups = [
    model_pb2.FeatureGroupConfig(
        group_name="wide",
        feature_names=embedding_features,
        group_type=model_pb2.FeatureGroupType.WIDE,
    ),
]

# Initialize EmbeddingGroup
embedding_group = EmbeddingGroup(features, feature_groups, device=device)
return embedding_group, embedding_features

[tiankongdeguiji]

If you want to use embedding_dim in the feature config, the group_type should be specified as model_pb2.FeatureGroupType.DEEP. If group_type is set to model_pb2.FeatureGroupType.WIDE, the table dimension will be 4, whereas we aim to set it to 1. However, it's important to note that there's currently a limitation in torchrec (as documented in this GitHub issue) that restricts this functionality.