What problem does it solve? Measuring training performance of object detection models on NVIDIA GPUs requires manual container setup, distributed training configuration, and log parsing. This Skill automates the full workflow: launching a Docker container, running mmdetection training benchmarks across models and precisions, and extracting AVG_ITER_TIME metrics into a structured JSON result. ## Core Features & Use Cases - Automated Benchmark Execution: Runs batch training benchmarks for faster_rcnn, mask_rcnn, cascade_rcnn, retinanet, yolov3, fcos, ssd300, centernet, solo, and swin_mask_rcnn on COCO with fixed fp16/fp32 precisions. - Structured Metric Collection: Parses MMEngine training logs for AVG_ITER_TIME, DATA, and OP metrics and writes all model/GPU/precision combinations into a single eval_result.json. - Flexible GPU Scaling: Supports single-GPU and multi-GPU distributed training via the CV_DET_NGPU environment variable. - Use Case: Ask the agent to test mask_rcnn training performance on 8 NVIDIA GPUs; it starts the container, runs both fp16 and fp32 passes, and returns per-iteration timing results. ## Quick Start Ask the agent to test faster_rcnn detection training performance on NVIDIA GPUs with the COCO dataset and pretrained weights mounted.
