Text-to-image semantic retrieval demo based on Unsplash Lite 2.5K image dataset. It enables user search image by natural language text.

The demo including the following parts:

• online retrieval pipeline service
• offline model export, data fetch and index

All negative samples have the same category(genre) as the corresponding positive samples. The category should correspond to the item id (movie_id):

Spark requires extra setting of entrypoint and user inside container.

rename the loan overdue project to loan default and fix some runtime errors.

In a lambda submitted to co_spawn, the captured by copy in lambda would be released twice under gcc 11 build.
Possibly related to https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100611 .
Use capture by ref as a workaround.

Add DCN V2 to CTR Demo and give benchmark on MovieLens and Criteo datasets.

DCN V2
criteo_d5:
Train AUC: 0.7487 Test AUC: 0.7290
1m uid mid:
Train AUC: 0.8901 Test AUC: 0.8611
25m uid mid:
Train AUC: 0.8888 Test AUC: 0.8323

Implement DSSM model and demonstrate the effect for recall stage on MovieLens datasets. DSSM paper.

Add DCN to CTR Demo and give benchmark on MovieLens and Criteo datasets.

25m uid mid:
Train AUC: 0.8972 Test AUC: 0.8430
1m uid mid:
Train AUC: 0.9021 Test AUC: 0.8746
criteo_d5:
Train AUC: 0.7413 Test AUC: 0.7304

1. Add diversification algorithms. We implemented a diversification model, named "Maximize Marginal Relevance Disperser" which refers to the paper "The Use of MMR, Diversity-Based Reranking for Reordering The dispersing method mentioned in Documents and Producing Summaries". Compared with SimpleDiversifier, MaximalMarginalRelevanceDiversifier can take into account information in multiple dimensions.
2. Integrating the MaximalMarginalRelevanceDiversifier into the pipeline after we completed the unit test of MaximalMarginalRelevanceDiversifier. In addition, we have updated the configuration information of the diversify method in the Consul file.
   .

Currently, metaspore-serving-bin init and load use different directory structure, where the latter contains model version in the path, which makes it impossible for metaspore-serving-bin to init from a previously loaded model directory.

Some files are already outdated, e.g. old compile/build scripts.

Implement scorecard demo based on loan default probability.

For the MovieLens Demo, we would better add python.zip before we submit submit fg_movielens.py PySpark job.

def init_spark():
    ## add a line of code here
    subprocess.run(['zip', '-r', './python.zip', 'fg_neg_sampler.py', 'fg_sparse_features_extractor.py', 'fg_gbm_features_extractor.py' ], cwd='./')
    spark = (SparkSession.builder
        .appName('MovieLens Demo')
        .config("spark.executor.memory","10G")
        .config("spark.submit.pyFiles", "python.zip")
        .config("spark.executor.instances","4")
        .config("spark.network.timeout","500")
        .getOrCreate())
    ... 

Moreover, please change the function name generate_spare_features to generate_sparse_features in fg_sparse_features_extractor.py

1. Warnings about lambda with implicit this pointer capture;
2. multiple definition of `absl::lts_20211102::Status::kMovedFromString

Update the implementation of MaximalMarginalRelevanceDiversifier, from the original List based implementation to the LinkList based implementation.

Arrow plan build and execution in feature compute is not threadsafe and may lead to crash under concurrent requests

Add a general export specification and NLP/CV examples for HuggingFace pre-trained model inference.

Add DeepFM to CTR Demo and give benchmark on MovieLens and Criteo datasets.

25m uid mid:
Train AUC: 0.8908 Test AUC: 0.8359
1m uid mid:
Train AUC: 0.8891 Test AUC: 0.8658
criteo_d5:
Train AUC: 0.7531 Test AUC: 0.7271

Kubernetes deployment with helm chart:

1. serving-chart
2. movielens-chart

Currently, the Python preprocessor uses stdout to pass input_names and output_names back to metaspore-serving-bin, which will make exceptions thrown in the Python preprocessor invisible in the log of metaspore-serving-bin. Another mechanism is needed.

update multimodal retrieval demo docs, add reference for online part's guide.

Please fix the fg.yaml error, screen shot:

Serving now just loads tabular model. We need to support loading Onnx model with tensors as inputs.

Implement dense/semantic retrieval dual-encoder models based on offline and online negatives sampling strategy.

Implement ESMM model to estimate CVR:

• Implement ESMM model
• Add Alibaba CPP dataset to test the model performance
• Test AUC: 0.6296, Test CTR AUC: 0.5731, Test CVR AUC: 0.6429

1. Support wrapping recurring run schedule time parameters and training components' parameters
2. Help fill experiment specific parameters for Estimator
3. Decouple experiment, scheduling and publish from model runners

Add PNN to CTR Demo and give benchmark on MovieLens and Criteo datasets.

iPNN
criteo_d5:
Train AUC: 0.7544 Test AUC: 0.7292
1m uid mid:
Train AUC: 0.8914 Test AUC: 0.8649
25m uid mid:
Train AUC: 0.8916 Test AUC: 0.8362

oPNN
criteo_d5:
Train AUC: 0.7533 Test AUC: 0.7287
1m uid mid:
Train AUC: 0.8896 Test AUC: 0.8633
25m uid mid:
Train AUC: 0.8905 Test AUC: 0.8353

build one online pipeline for multi-model retrieval scenes.

Add Wide&Deep to CTR Demo and give benchmark on MovieLens and Criteo datasets.

25m uid mid:
Train AUC: 0.8898 Test AUC: 0.8343
1m uid mid:
Train AUC: 0.8937 Test AUC: 0.8682
criteo_d5:
Train AUC: 0.7394 Test AUC: 0.7294

Create demo notebooks for MetaSpore usage in Alpha IDE. Users can run it on MovieLens-1M dataset.

Unify the data processing for movieLens-1m, movielens-25m, criteo-5d and other datasets, including feature generation, match dataset generation, ranking dataset generation, negative sampling, etc.

QA is a text-to-text semantic retrieval demo based on 1M Baike-Question-Answer database.

The demo including the following parts

1. online system: an end-to-end online retrieval services.
2. offline system: model training and export, data fetch and index.

1. Refactor code organization with seperate algo, runner, component and pipeline definitions.
2. Auto export kubeflow components of built-in algo runners and also a python decorator for customized use.
3. Load components by name to construct kubeflow pipeline and upload it automatically.

Add Kolmogorov–Smirnov Test metric:

• A util of KS-Test to calculate its metric.
• A notebook to show how to get statistic of KS-Test and how to plot the KS-Test curve.

Use ms.nn.Normalization instead of torch.nn.BatchNorm1d:
In parameter server architecture, we should use ms.nn.Normalization which handled global exponentially weighted moving average already.

Add AutoInt to CTR Demo and give benchmark on MovieLens and Criteo datasets.

AutoInt
criteo_d5:
Train AUC: 0.7558 Test AUC: 0.7361
1m uid mid:
Train AUC: 0.9028 Test AUC: 0.8741
25m uid mid:
Train AUC: 0.8968 Test AUC: 0.8421

Implement Multi-Task MMoE model:

1. MMoE net.
2. MMoE training pipeline (demo).
3. Preprocessing of Census dataset.

metaspore-serving-bin tries each model load type sequentially, which will leave extra error logs. A better way would be checking model load type according to directory structure in advance.

We will use dataset to train a LightGBM model for loan overdue rate estimation.

Add xDeepFM to CTR Demo and give benchmark on MovieLens and Criteo datasets.

xDeepFM
criteo_d5:
Train AUC: 0.7541 Test AUC: 0.7300
1m uid mid:
Train AUC: 0.8892 Test AUC: 0.8641
25m uid mid:
Train AUC: 0.8911 Test AUC: 0.8367

Goal

To provide a framework for calling python method in user custom scripts. Python code is executed in a separate process rather than embeded in cpp process. The CPython interpreters are run on a per-thread basis to avoid GIL contention.

Design

1. Control plane via gRPC and unix domain socket between cpp and python;
2. Data plane via either gRPC for small data and shared memory for large data;
3. Model packaged with customized python venv and user scripts;
4. For each cpp compute thread, run a CPython interpreter process with user entry script;
5. Provide an async iterator style interface for python.