[BUG] get_embedding_sizes generates wrong embedding shape with encode_type = 'joint'

[rnyak]

Describe the bug

When we jointly encode categorical columns, nvt.ops.get_embedding_sizes(workflow) does not generate the correct embedding table.

Steps/Code to reproduce bug

df = cudf.DataFrame({'a_user_id': ["User_A","User_E","User_B","User_C","User_A","User_B","User_B","User_C","User_B","User_A"],
                    'b_user_id': ["User_B", "User_F", "User_D", "User_D", "User_B", "User_E", "User_E", "User_D", "User_D", "User_D"], 
                    'media':[3, 3, 12, 17, 3, 1, 1, 0, 1, 12], 'language': ['en', 'en', 'spn', 'fr', 'spn', 'en', 'fr', 'ch', 'ch', 'en']})
dataset = nvt.Dataset(df)
cat_users = ([['a_user_id','b_user_id']]) >> nvt.ops.Categorify(encode_type = 'joint')
cat_others = ['media', 'language'] >> nvt.ops.Categorify()
workflow = nvt.Workflow(cat_users + cat_others)
workflow.fit(dataset)
new_gdf = workflow.transform(dataset).to_ddf().compute()
new_gdf.head()

a_user_id b_user_id media language
0   1   2   3   2
1   5   6   3   2
2   2   4   4   4
3   3   4   5   3
4   1   2   3   4

nvt.ops.get_embedding_sizes(workflow)

{'media': (6, 16),
 'language': (5, 16),
 'a_user_id': (0, 16),
 'b_user_id': (0, 16)}

Expected behavior The following embedding table shapes are expected:

nvt.ops.get_embedding_sizes(workflow)

{'media': (6, 16),
 'language': (5, 16),
 'a_user_id_b_user_id': (6, 16)}

Environment details (please complete the following information):

• Environment location: Docker Pytorch-training image.
• Method of NVTabular install: Docker

[rnyak]

@benfred do you think this is a bug, or nvt.ops.get_embedding_sizes(workflow) gives the expected output?