ctr-training

What problem does it solve?

This Skill streamlines configuring, running, and evaluating CTR prediction experiments so practitioners can compare models and reproduce results without manually stitching together dataset, model, and training steps.

Core Features & Use Cases

• End-to-end training orchestration: Prepare model and dataset YAMLs, launch experiments, and manage GPU/CPU execution for models such as DeepFM, DCNv2, DIN, xDeepFM and AutoInt.
• Config templates and tuning guidance: Provides dataset_config and model_config templates, recommended hyperparameters, and best-practice settings for different dataset scales.
• Troubleshooting and evaluation: Includes common error fixes (NumPy compatibility, CUDA OOM), logging conventions, and guidance for monitoring AUC/logloss and saving best models.
• Use Case: Run a unified benchmark to compare multiple CTR models across MovieLens, Frappe, TaobaoAd, and Criteo datasets and collect metrics for model selection.

Quick Start

Run the ctr-training workflow by preparing dataset and model config YAMLs in the appropriate FuxiCTR model_zoo/{MODEL}/config directory and executing the run_expid.py command with your desired expid and GPU flags.