YOLO → SAM Benchmark

About 381 wordsAbout 1 min

YoloSamBench composes object detection and segmentation as a real multi-model graph. YOLO and SAM live in separate persistent Ray actor pools, allowing different images to occupy different stages concurrently while avoiding model reload for every batch.

Topology

Stages preserve per-image alignment across multiple output Ports. In CUDA mode, yolo_replicas=2 and sam_replicas=2 reserve four GPUs concurrently by default, so replica counts should reflect both cluster capacity and the relative speed of the two models.

Run

Prepare input images, YOLO weights, a SAM checkpoint, and the dependencies listed in env.json:

from rayorch.benchmark import YoloSamBench

bench = YoloSamBench(
    input_path="/shared/images",
    output_dir="/shared/yolo-sam-output",
    yolo_model="/shared/models/yolo.pt",
    sam_checkpoint="/shared/models/sam_vit_b.pth",
    input_limit=100,
    yolo_replicas=2,
    sam_replicas=2,
    batch_size=4,
    input_batch_size=16,
    max_active_input_batches=2,
)

report = bench.run(ray_address="auto")

For a CPU smoke test, use device="cpu" and normally one replica for each model. To place model stages in separate environments or on selected nodes, override runtime_env, custom resources, or GPU counts through stage_options["yolo"] and stage_options["sam"].

Outputs and metrics

Each image produces <stem>.overlay.jpg and <stem>.json. report.outputs includes input path, output path, yolo_boxes, and sam_masks; added workload metrics are images, detections, and masks.

What performance it demonstrates

Observation	Where to read it	Question answered
end-to-end image throughput	`images / measured_wall_s`	images per second through detection, segmentation, rendering, and storage
model-pool balance	compare YOLO and SAM Call time, RPCs, and average batch	which model limits throughput and whether replica ratios are appropriate
pipeline parallelism	increase input count and active input batches	whether overlapping YOLO, SAM, and CPU work across images improves throughput
batching efficiency	model Call batch statistics	whether configured batches fill or the workload is too small
output complexity	`detections` and `masks`	explain work variation caused by image content; these counts are not throughput by themselves

Scan (yolo_replicas, sam_replicas) on the same image set instead of assuming equal counts are optimal. Storage bandwidth and automatic-mask count can dominate, so reports should state storage type, model variants, image resolution, and thresholds.