Resume and Iterate

Pick the Right Kind of Run#

Worked Example: Resume an Interrupted Run#

resume=True restores the model, optimizer state, learning-rate schedule, and epoch counter from last.pt, then continues the original plan:

yolo train resume=True model=runs/detect/pills_resume/weights/last.pt

Real output from a 5-epoch run stopped after epoch 2:

Resuming training runs/detect/pills_resume/weights/last.pt from epoch 3 to 5 total epochs
        3/5         0G      1.232      2.393   0.005478        206        320: 100% 6/6 4.5s
        4/5         0G      1.345      1.859   0.005833        208        320: 100% 6/6 4.5s
        5/5         0G       1.31      1.625   0.005854        191        320: 100% 6/6 4.4s
3 epochs completed in 0.004 hours.

• from epoch 3 to 5 total epochs: it picked up exactly after the last saved epoch, with the original target.
• You didn't pass data or epochs: every argument came from the checkpoint. A few runtime settings such as batch, device, and workers can be overridden on resume; anything else is ignored with a Resume ignores ... warning.

Worked Example: Warm-Start Round Two#

Most "next round" experiments aren't resumes; they're new runs that start from your last good model — plain transfer learning within your own project. Use a lower learning rate so the model refines rather than relearns:

from ultralytics import YOLO

model = YOLO("runs/detect/pills_v1/weights/best.pt")  # warm start from round 1
model.train(
    data="my_dataset/data.yaml",
    epochs=5,
    imgsz=320,
    optimizer="AdamW",  # set explicitly: optimizer="auto" ignores lr0
    lr0=0.0005,  # gentler than round 1's auto-selected 0.002
    name="pills_v2_warm",
)

The log confirms the setting took effect: optimizer: AdamW(lr=0.0005, momentum=0.937). The fine-tuning guide covers layer freezing and two-stage training; the hyperparameter tuning guide covers sweeps.

Worked Example: Compare Runs#

Name every run after what changed — pills_v2_warm, pills_v3_imgsz640, pills_v4_edge_frames — never train2 or final. Each run writes a results.csv, so a comparison takes a few lines:

from pathlib import Path

import pandas as pd

for csv in sorted(Path("runs/detect").glob("*/results.csv")):
    df = pd.read_csv(csv)
    best = df["metrics/mAP50-95(B)"].max()
    print(f"{csv.parent.name:<14} epochs={len(df):<3} best mAP50-95={best:.3f}")

pills_resume   epochs=5   best mAP50-95=0.579
pills_v1       epochs=10  best mAP50-95=0.659
pills_v2_warm  epochs=5   best mAP50-95=0.694

Five warm-started epochs lifted mAP50-95 from 0.659 to 0.694, while five epochs from COCO weights (pills_resume) reached only 0.579. Warm-starting is cheaper progress — as long as the classes and labels haven't changed underneath it. Past a handful of runs, a tracker pays for itself: Ultralytics Platform (see Compare Runs and Pick a Winner), or the Comet and Weights & Biases integrations, log every run automatically.

Keep Each Round Cheap#

Most rounds answer one question, such as "did the new frames help?". These arguments buy faster answers while you explore:

Change one thing per round, and confirm the winner with a full-size run before you ship it. When you do need a fresh start, it's the same call with the pretrained weights:

from ultralytics import YOLO

model = YOLO("yolo26n.pt")  # back to the COCO-pretrained weights
model.train(data="my_dataset/data.yaml", epochs=10, imgsz=320, name="pills_v3_relabeled")

Avoid These Mistakes#