Guide to Using Tencent TI-ONE Platform for the 2020 Advertising Algorithm Competition

This guide introduces the Tencent TI-ONE platform as the designated machine learning platform for the 2020 Tencent Advertising Algorithm Competition, outlining the steps to leverage its built-in frameworks for model development and submission.

The platform provides deep learning frameworks including TensorFlow 1.12, PyTorch 1.1.0, and PyCaffe 1.0.0-rc3-ssd, as well as machine learning frameworks such as Spark 2.4, PySpark 2.4, and Analytics Zoo 0.7.0.

Users begin by logging into the TI-ONE console, setting the region to Shanghai, creating a new project, and specifying a COS bucket for storing training data and intermediate results.

Within the project, a custom workflow is created; the input component “公共数据集 → 算法大赛数据集” is dragged onto the canvas to obtain dataset paths, which are accessed via program parameters ${ai_dataset_lib} and ${cos} .

Example paths are ${ai_dataset_lib}/contest/demo/iris_training.csv for the training set and ${ai_dataset_lib}/contest/demo/iris_test.csv for the test set, with results directed to ${cos}/contest_result .

For model training, a framework (e.g., TensorFlow) is selected from the algorithm bar and added to the canvas; note that connections only indicate execution order, and data must be passed via program parameters.

Algorithm parameters are configured in the framework’s settings panel, where users upload custom scripts, dependency packages, and set program parameters such as --train_path ${ai_dataset_lib}/contest/demo/iris_training.csv , --test_path ${ai_dataset_lib}/contest/demo/iris_test.csv , and --result_dir ${cos}/contest_result .

Additional settings include TensorBoard directory, program dependencies, Python version (chosen as 3.5), and resource type (TI.MEDIUM4.2core4g).

After saving the workflow, clicking “Run” executes the pipeline; upon completion, result files are located in the specified COS bucket folder, where users can retrieve the object address for submission to the competition website.

The tutorial concludes by reminding readers that the demonstrated paths and data are illustrative and not the official competition datasets.