dcnv2-training

What problem does it solve?

This Skill automates the configuration, training, and evaluation of DCNv2 models within the FuxiCTR framework for click-through-rate prediction, reducing manual setup, configuration errors, and common runtime issues.

Core Features & Use Cases

• End-to-end training: Provides guidance for model_config.yaml and dataset_config.yaml to run experiments end-to-end.
• Multi-environment support: Supports CPU and single- or multi-GPU training options and DataParallel execution.
• Configuration templates & tuning: Exposes key hyperparameters such as embedding_dim, num_cross_layers, batch_size, and learning_rate for benchmarking and hyperparameter sweeps.
• Troubleshooting: Documents common fixes for NumPy compatibility, OOM errors, and invalid data paths.
• Use Case: Run reproducible DCNv2 experiments across MovielensLatest_x1 and other CTR datasets to compare AUC and logloss across model variants.

Quick Start

Run the dcnv2-training skill to train DCNv2 on a prepared dataset by pointing model_config.yaml and dataset_config.yaml to your data directory and executing run_expid with the desired GPU device.