train_v3.py

#!/usr/bin/env python3
# v3 max-throughput: yolo26m on 6673-img dataset, batch=128, cache=ram
from ultralytics import YOLO
import torch
print('GPU:', torch.cuda.get_device_name(0), '|', round(torch.cuda.get_device_properties(0).total_memory/1e9), 'GB')

model = YOLO('yolo26m.pt')
model.train(
    data='/home/azureuser/merged_v3/data.yaml',
    epochs=200,
    imgsz=640,
    batch=128,          # 2x v2 — should hit ~70GB VRAM
    device=0,
    workers=16,         # feed data faster
    cache='ram',        # dataset is ~1GB, fits easily
    project='runs_v3',
    name='h100_3class_v3',
    exist_ok=True,
    amp=True,
    cos_lr=True,
    close_mosaic=20,
    mosaic=1.0, mixup=0.15, copy_paste=0.3,
    hsv_h=0.015, hsv_s=0.7, hsv_v=0.4,
    degrees=5.0, translate=0.1, scale=0.5, fliplr=0.5,
    cls=1.0, box=7.5, dfl=1.5,
    weight_decay=0.0005,
    optimizer='auto',
    patience=60,
    plots=True, verbose=True,
)
print('TRAIN DONE — running val + test')
m = YOLO('runs_v3/h100_3class_v3/weights/best.pt')
print('--- VAL ---'); m.val(data='/home/azureuser/merged_v3/data.yaml', split='val')
print('--- TEST ---'); m.val(data='/home/azureuser/merged_v3/data.yaml', split='test')
print('--- TEST + TTA ---'); m.val(data='/home/azureuser/merged_v3/data.yaml', split='test', augment=True)