[BUG] categorical columns automatically sorted.

Describe the bug I put the

cats={'a_is_verified': (2, 16),
 'b_follows_a': (2, 16),
 'tw_first_word': (1411026, 16),
 'tw_second_word': (1111029, 16),
 'tw_last_word': (140149, 16),
 'tw_llast_word': (232494, 16),
 'language': (65, 16),
 'tweet_type': (3, 16),
 'media': (16, 16),
 'a_user_id_index': (41693, 16),
 'b_user_id_index': (10495, 16),
 'mfi_hashtags': (63165, 16),
 'mfg_hashtags': (77144, 16),
 'mfi_links': (29328, 16),
 'mfg_links': (29467, 16),
 'mfi_domains': (18090, 16),
 'mfg_domains': (18329, 16)}

TorcyAsyncItr Dataset, but,

batch = next(iter(train_loader))
cat_feats, cont_feats, labels = batch

features are returned ordered alphabetically.

    cat_feats.max(axis=0).values

this command yields a somewhat weird order.

tensor([     1,  19490,      1,   3080,     63,      8,   3886,  13118,   3410,
          3839,  11250,   3406, 123392,  24677,  38871, 112914,      2],
       device='cuda:0')

I think this is not intended behavior.

Steps/Code to reproduce bug

# External dependencies
import os
import cudf  # cuDF is an implementation of Pandas-like Dataframe on GPU
import shutil
import numpy as np

import nvtabular as nvt

from os import path
import pandas as pd
from nvtabular.loader.torch import TorchAsyncItr, DLDataLoader


x = pd.concat([pd.Series([0, 0, 0, 0, 0]), pd.Series([0,1,2,3,4]), pd.Series([1,1,1,1,1])], axis=1)
x.columns = ['z', 'a', 'c']
cat_cols = ['a', 'z']
x.to_csv("./ret.csv")


train_dataset = TorchAsyncItr(
    nvt.Dataset("./ret.csv"),
    batch_size=2,
    cats=cat_cols,
    conts=[],
    labels=["c"])
train_loader = DLDataLoader(
    train_dataset, batch_size=None, collate_fn=lambda x: x, pin_memory=False, num_workers=0
)

batch = next(iter(train_loader)); a,b,c = batch
print(a)


tensor([[0, 0],
        [1, 0]], device='cuda:0')

Expected behavior the first column is “a” and its maximum value will be 0.

**Environment details (please complete I’m using docker. I installed NVTabuler via conda commands in README.md (using CUDA 10.1)

conda env create -f=conda/environments/nvtabular_dev_cuda10.1.yml
conda activate nvtabular_dev_10.1
python -m ipykernel install --user --name=nvt
pip install -e .
jupyter notebook

[jperez999]

Sorry, I thought your question was as to why the max is 1 instead of zero. If you are asking whether we are sorting columns alphabetically, you are correct we are sorting. That code can be found here (https://github.com/NVIDIA/NVTabular/blob/739c6ef92915ef2493ec33ae5020e0978ab2d170/nvtabular/loader/backend.py#L455-L462). This is done to guarantee ordering in both tensor representation and tensor embeddings size table. We use the same ordering function to order the tensor embedding sizes (https://github.com/NVIDIA/NVTabular/blob/e02549a785dc3297b2065b44266226e0fe7af330/nvtabular/ops/categorify.py#L451). Let me know if this helps.

[jperez999]

We are modifying this for our upcoming release, currently in a PR (https://github.com/NVIDIA/NVTabular/pull/793). We are changing to mimic the tensorflow dataloader output which provides a dictionary based structure with the following possible key/value signatures “column name”: tensor “column name”: (tensor, offsets) “column_name”: sparse-tensor This will allow the user to dictate any ordering they may want to specify.