mindspore/model_zoo/research/recommend/autodis

• Contents
• AutoDis Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• Script Description
• • Script and Sample Code
  • Script Parameters
  • Training Process
  • • Training
  • Evaluation Process
  • • Evaluation
• Model Description
• • Performance
  • • Training Performance
    • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

Contents

• AutoDis Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• Script Description
  • Script and Sample Code
  • Script Parameters
  • Training Process
    • Training
  • Evaluation Process
    • Evaluation
• Model Description
  • Performance
    • Evaluation Performance
    • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

AutoDis Description

The common methods for numerical feature embedding are Normalization and Discretization. The former shares a single embedding for intra-field features and the latter transforms the features into categorical form through various discretization approaches. However, the first approach surfers from low capacity and the second one limits performance as well because the discretization rule cannot be optimized with the ultimate goal of CTR model.
To fill the gap of representing numerical features, in this paper, we propose AutoDis, a framework that discretizes features in numerical fields automatically and is optimized with CTR models in an end-to-end manner. Specifically, we introduce a set of meta-embeddings for each numerical field to model the relationship among the intra-field features and propose an automatic differentiable discretization and aggregation approach to capture the correlations between the numerical features and meta-embeddings. AutoDis is a valid framework to work with various popular deep CTR models and is able to improve the recommendation performance significantly.

Paper: Huifeng Guo*, Bo Chen*, Ruiming Tang, Zhenguo Li, Xiuqiang He. AutoDis: Automatic Discretization for Embedding Numerical Features in CTR Prediction

Model Architecture

AutoDis leverages a set of meta-embeddings for each numerical field, which are shared among all the intra-field feature values. Meta-embeddings learn the relationship across different feature values in this field with a manageable number of embedding parameters. Utilizing meta-embedding is able to avoid explosive embedding parameters introduced by assigning each numerical feature with an independent embedding simply. Besides, the embedding of a numerical feature is designed as a differentiable aggregation over the shared meta-embeddings, so that the discretization of numerical features can be optimized with the ultimate goal of deep CTR models in an end-to-end manner.

Dataset

• [1] A dataset Criteo used in Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, Xiuqiang He. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction[J]. 2017.

Environment Requirements

• Hardware（Ascend/GPU）
  • Prepare hardware environment with Ascend or GPU processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Quick Start

After installing MindSpore via the official website, you can start training and evaluation as follows:

• running on Ascend

  # run training example
  python train.py \
    --dataset_path='dataset/train' \
    --ckpt_path='./checkpoint' \
    --eval_file_name='auc.log' \
    --loss_file_name='loss.log' \
    --device_target='Ascend' \
    --do_eval=True > ms_log/output.log 2>&1 &
  
  # run evaluation example
  python eval.py \
    --dataset_path='dataset/test' \
    --checkpoint_path='./checkpoint/autodis.ckpt' \
    --device_target='Ascend' > ms_log/eval_output.log 2>&1 &
  OR
  sh scripts/run_eval.sh 0 Ascend /dataset_path /checkpoint_path/autodis.ckpt
  

  For distributed training, a hccl configuration file with JSON format needs to be created in advance.

  Please follow the instructions in the link below:

  https://gitee.com/mindspore/mindspore/tree/master/model_zoo/utils/hccl_tools.

Script Description

Script and Sample Code

.
└─autodis
  ├─README.md
  ├─mindspore_hub_conf.md             # config for mindspore hub
  ├─scripts
    ├─run_standalone_train.sh         # launch standalone training(1p) in Ascend or GPU
    └─run_eval.sh                     # launch evaluating in Ascend or GPU
  ├─src
    ├─__init__.py                     # python init file
    ├─config.py                       # parameter configuration
    ├─callback.py                     # define callback function
    ├─autodis.py                      # AutoDis network
    ├─dataset.py                      # create dataset for AutoDis
  ├─eval.py                           # eval net
  └─train.py                          # train net

Script Parameters

Parameters for both training and evaluation can be set in config.py

• train parameters

  optional arguments:
  -h, --help            show this help message and exit
  --dataset_path DATASET_PATH
                        Dataset path
  --ckpt_path CKPT_PATH
                        Checkpoint path
  --eval_file_name EVAL_FILE_NAME
                        Auc log file path. Default: "./auc.log"
  --loss_file_name LOSS_FILE_NAME
                        Loss log file path. Default: "./loss.log"
  --do_eval DO_EVAL     Do evaluation or not. Default: True
  --device_target DEVICE_TARGET
                        Ascend or GPU. Default: Ascend
  

• eval parameters

  optional arguments:
  -h, --help            show this help message and exit
  --checkpoint_path CHECKPOINT_PATH
                        Checkpoint file path
  --dataset_path DATASET_PATH
                        Dataset path
  --device_target DEVICE_TARGET
                        Ascend or GPU. Default: Ascend
  

Training Process

Training

• running on Ascend

  python train.py \
    --dataset_path='dataset/train' \
    --ckpt_path='./checkpoint' \
    --eval_file_name='auc.log' \
    --loss_file_name='loss.log' \
    --device_target='Ascend' \
    --do_eval=True > ms_log/output.log 2>&1 &
  

  The python command above will run in the background, you can view the results through the file ms_log/output.log.

  After training, you'll get some checkpoint files under ./checkpoint folder by default. The loss value are saved in loss.log file.

  2020-12-10 14:58:04 epoch: 1 step: 41257, loss is 0.44559600949287415
  2020-12-10 15:06:59 epoch: 2 step: 41257, loss is 0.4370603561401367
  ...
  

  The model checkpoint will be saved in the current directory.

Evaluation Process

Evaluation

• evaluation on dataset when running on Ascend

  Before running the command below, please check the checkpoint path used for evaluation.

  python eval.py \
    --dataset_path='dataset/test' \
    --checkpoint_path='./checkpoint/autodis.ckpt' \
    --device_target='Ascend' > ms_log/eval_output.log 2>&1 &
  OR
  sh scripts/run_eval.sh 0 Ascend /dataset_path /checkpoint_path/autodis.ckpt
  

  The above python command will run in the background. You can view the results through the file "eval_output.log". The accuracy is saved in auc.log file.

  {'result': {'AUC': 0.8109881454077731, 'eval_time': 27.72783327102661s}}
  

Model Description

Performance

Training Performance

Parameters	Ascend
Model Version	AutoDis
Resource	Ascend 910; CPU 2.60GHz, 192cores; Memory 755G
uploaded Date	12/12/2020 (month/day/year)
MindSpore Version	1.1.0
Dataset	[1]
Training Parameters	epoch=15, batch_size=1000, lr=1e-5
Optimizer	Adam
Loss Function	Sigmoid Cross Entropy With Logits
outputs	Accuracy
Loss	0.42
Speed	1pc: 8.16 ms/step;
Total time	1pc: 90 mins;
Parameters (M)	16.5
Checkpoint for Fine tuning	191M (.ckpt file)
Scripts	AutoDis script

Inference Performance

Parameters	Ascend
Model Version	AutoDis
Resource	Ascend 910
Uploaded Date	12/12/2020 (month/day/year)
MindSpore Version	0.3.0-alpha
Dataset	[1]
batch_size	1000
outputs	accuracy
AUC	1pc: 0.8112;
Model for inference	191M (.ckpt file)

Description of Random Situation

We set the random seed before training in train.py.

ModelZoo Homepage

Please check the official homepage.