Experience Sharing on Using Tencent TI-ONE Platform for Advertising Algorithm Competition

The 2021 Tencent Advertising Algorithm Competition officially launched its preliminary round on April 30, 2021, hosted by Tencent Advertising with co‑organizers Tencent Cloud AI, Tencent Big Data, Tencent Recruitment, Tencent University Cooperation, and NVIDIA, and supported by TI‑ONE and the Angel machine‑learning platform.

The competition was selected as part of the ACM Multimedia Grand Challenge, featuring two tracks focused on video advertising and gaining international recognition.

To help participants, past top contestants shared their TI‑ONE usage experiences. The author recounts using TI‑ONE in the 2020 competition and the Huawei click‑through‑rate contest, noting the platform’s two development modes: a GUI‑based engineering mode with drag‑and‑drop algorithm components for rapid baseline construction, and a Notebook mode offering flexible instance resources, pre‑configured Conda environments for major frameworks, and lifecycle‑based environment installation.

In the engineering mode, users can quickly build baselines for classification, NLP, and unsupervised tasks, while the Notebook mode provides isolated containers with monitoring, logging, and the ability to allocate GPU memory, though instance crashes often stem from exhausted memory.

The author describes using a V100×4 configuration with MirroredStrategy in the advertising competition, loading entire datasets into 130 GB memory, and contrasts this with the Huawei contest where feature engineering caused memory bottlenecks, requiring generator‑based training and smaller batch sizes.

Finally, the piece reflects on the relentless demand for compute resources, praises the platform’s support staff, notes recent updates such as pre‑installed PyTorch/TensorFlow Conda environments, and encourages others to utilize TI‑ONE while valuing its precious computational power.