Scripts/train_parallel.py

#!/usr/bin/env python3

# Example:
# source ~/envs/yolo/bin/activate
# python3 train_parallel.py \
#     --dataset /media/rtx5090/IRIS/Synthetic_Train_Sets/4K_Physics_Intrinsics_RGB_Exp/yolo/dataset.yaml \
#     --run-start 1 \
#     --run-end 6

import multiprocessing as mp
import subprocess
import os
from pathlib import Path
import socket
import argparse


HOSTNAME = socket.gethostname()

# Minimum free VRAM required for one training job.
# GPUs with less memory will automatically be grouped together.
VRAM_PER_TRAINING_MB = 22500


def get_gpu_memory():

    result = subprocess.check_output(
        [
            "nvidia-smi",
            "--query-gpu=memory.total,memory.used",
            "--format=csv,noheader,nounits",
        ]
    )

    gpu_info = []

    for line in result.decode().strip().split("\n"):

        total, used = line.split(",")

        gpu_info.append(
            {
                "total": int(total),
                "used": int(used),
                "free": int(total) - int(used),
            }
        )

    return gpu_info


def create_gpu_slots():

    gpu_info = get_gpu_memory()

    print("\nGPU memory status:")

    for gpu_id, gpu in enumerate(gpu_info):

        print(
            f"GPU {gpu_id}: "
            f"{gpu['free']} MB free / "
            f"{gpu['total']} MB total"
        )

    slots = []
    used = set()

    #
    # First create single-GPU slots.
    #
    for gpu_id, gpu in enumerate(gpu_info):

        if gpu["free"] >= VRAM_PER_TRAINING_MB:

            slots.append(gpu_id)
            used.add(gpu_id)

    #
    # Group remaining GPUs until enough memory is available.
    #
    remaining = [
        (gpu_id, gpu_info[gpu_id])
        for gpu_id in range(len(gpu_info))
        if gpu_id not in used
    ]

    current_group = []
    current_memory = 0

    for gpu_id, gpu in remaining:

        current_group.append(gpu_id)
        current_memory += gpu["free"]

        if current_memory >= VRAM_PER_TRAINING_MB:

            slots.append(current_group)

            current_group = []
            current_memory = 0

    if current_group:

        print(
            "\nWarning: Remaining GPUs "
            f"{current_group} do not have enough combined free memory "
            "for another training job."
        )

    if not slots:
        raise RuntimeError(
            "No GPU (or GPU group) has enough free memory to start training."
        )

    return slots


def train_worker(run_id, gpu, seed, dataset_path):

    from ultralytics import YOLO

    model = YOLO("yolo11m.pt")

    dataset_name = Path(dataset_path).parents[1].name

    project_folder = os.path.join(
        "training_stats",
        dataset_name,
        HOSTNAME,
    )

    print(
        f"Starting run {run_id} "
        f"on GPU(s) {gpu} "
        f"with seed {seed}"
    )

    model.train(
        data=dataset_path,
        epochs=500,
        patience=30,
        batch=16,
        imgsz=1024,
        lr0=0.01,
        device=gpu,
        project=project_folder,
        name=f"run_{run_id}",
        seed=seed,
    )


def run_job(args):

    run_id, gpu, seed, dataset_path = args

    train_worker(
        run_id,
        gpu,
        seed,
        dataset_path,
    )


if __name__ == "__main__":

    parser = argparse.ArgumentParser()

    parser.add_argument(
        "--dataset",
        type=str,
        required=True,
        help="Path to YOLO dataset.yaml",
    )

    parser.add_argument(
        "--run-start",
        type=int,
        default=1,
        help="First run number.",
    )

    parser.add_argument(
        "--run-end",
        type=int,
        default=6,
        help="Last run number (inclusive).",
    )

    args = parser.parse_args()

    if args.run_start > args.run_end:
        raise ValueError("--run-start must be <= --run-end")

    mp.set_start_method(
        "spawn",
        force=True,
    )

    slots = create_gpu_slots()

    print("\nAvailable training slots:")

    for i, slot in enumerate(slots):

        if isinstance(slot, list):
            print(f"Slot {i}: GPUs {slot}")
        else:
            print(f"Slot {i}: GPU {slot}")

    jobs = []

    for run_id in range(
        args.run_start,
        args.run_end + 1,
    ):

        gpu = slots[
            (run_id - args.run_start) % len(slots)
        ]

        seed = run_id - 1

        jobs.append(
            (
                run_id,
                gpu,
                seed,
                args.dataset,
            )
        )

    processes = []

    for job in jobs:

        p = mp.Process(
            target=run_job,
            args=(job,),
        )

        p.start()

        processes.append(p)

        #
        # Run at most one job per available slot.
        #
        if len(processes) >= len(slots):

            for p in processes:
                p.join()

            processes = []

    for p in processes:
        p.join()

    print("\nAll training runs completed.")