atft-training

What problem does it solve?

Manually managing complex deep learning training loops, hyper-parameter optimization, and GPU resource allocation for the ATFT-GAT-FAN model is error-prone and inefficient. This Skill automates these critical tasks, ensuring optimal model performance.

Core Features & Use Cases

• Production-Grade Training: Launch and monitor optimized training runs for the ATFT-GAT-FAN forecaster, ensuring correct dataset and version parity.
• Hyper-Parameter Optimization: Tune critical parameters like learning rate and batch size, leveraging 80GB GPU headroom for efficient exploration.
• Use Case: Initiate a new production training run, automatically compiling with TorchInductor and FlashAttention2, then monitor its progress and GPU utilization to ensure optimal performance and timely completion.

Quick Start

Example: Run optimized training and monitor

make train-optimized DATASET=output/ml_dataset_latest_full.parquet make train-monitor