model training issues

[Christcool]

Good morning !
I am contacting you to report a serious problem that I encountered in the process of detecting plant disease with yolo. The metrics are almost average, I tried to apply several techniques such as augmentation, contrast adjustment, resizing but despite everything the result remains critical.

I am contacting you because you certainly have the experience and if your time permits you will be able to help me! Thanks a lot

[UltralyticsAssistant]

👋 Hello @Christcool, thank you for your interest in Ultralytics 🚀! This is an automated response to help get you moving quickly, and an Ultralytics engineer will also assist you soon.
We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.
Since this appears to be a custom training ❓ question, please provide as much information as possible, including dataset image examples, your dataset YAML, model used, training command, key metrics, confusion matrix/results plots, and training logs, and verify you are following our Tips for Best Training Results 🌱
If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.
Join the Ultralytics community where it suits you best. For real-time chat, head to Discord 🎧. Prefer in-depth discussions? Check out Discourse. Or dive into threads on our Subreddit to share knowledge with the community.

Upgrade

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[Christcool]

these are a sample of my result

[Christcool]

@glenn-jocher

[glenn-jocher]

Thanks for the ping — the plots are useful, but to diagnose this I still need the last 10 rows of results.csv pasted as text, your exact train args or args.yaml, and 3–5 matching val_batch*_labels.jpg/val_batch*_pred.jpg pairs; if the confusion is mostly between similar disease classes, that usually points to class-definition or label-consistency issues more than augmentation, and the performance metrics guide may help interpret the curves.

[Christcool]

@glenn-jocher
epoch,time,train/box_loss,train/cls_loss,train/dfl_loss,metrics/precision(B),metrics/recall(B),metrics/mAP50(B),metrics/mAP50-95(B),val/box_loss,val/cls_loss,val/dfl_loss,lr/pg0,lr/pg1,lr/pg2
1,1655.43,2.07852,2.85817,1.87781,0.32124,0.28629,0.21971,0.09519,2.0689,2.79163,1.86273,6.66496e-05,6.66496e-05,0.0667418
2,3262.33,1.98568,2.35214,1.76649,0.38211,0.34154,0.31997,0.13664,2.03934,2.27207,1.81993,0.000128917,0.000128917,0.0334708
3,4797.32,1.94298,2.18297,1.7279,0.41738,0.37279,0.35332,0.15914,1.97662,2.2111,1.75798,0.000186784,0.000186784,0.000195298
4,6348.03,1.91631,2.08472,1.70012,0.43147,0.39351,0.37882,0.17269,1.94539,2.1379,1.73359,0.0001802,0.0001802,0.0001802
5,7892.32,1.89635,2.02524,1.68051,0.44057,0.40406,0.38669,0.17661,1.94129,2.14161,1.72083,0.0001736,0.0001736,0.0001736
6,9442.48,1.88378,1.982,1.66583,0.45165,0.40078,0.39295,0.18182,1.92696,2.08837,1.70303,0.000167,0.000167,0.000167
7,10996.4,1.87101,1.9543,1.65304,0.45425,0.41339,0.4022,0.1866,1.91878,2.05356,1.69792,0.0001604,0.0001604,0.0001604
8,12548.8,1.8651,1.91983,1.64492,0.46587,0.41841,0.40873,0.18939,1.91566,2.02223,1.69028,0.0001538,0.0001538,0.0001538
9,14098.2,1.85629,1.8997,1.63642,0.46267,0.41922,0.40811,0.19019,1.91088,2.03281,1.68434,0.0001472,0.0001472,0.0001472
10,15649.2,1.84854,1.87871,1.62835,0.46817,0.42005,0.41148,0.19294,1.90672,2.02075,1.67666,0.0001406,0.0001406,0.0001406
11,17189.7,1.8437,1.86179,1.62263,0.47143,0.42226,0.41413,0.19418,1.90683,2.00369,1.67477,0.000134,0.000134,0.000134
12,18726.3,1.83707,1.84191,1.6161,0.47334,0.42149,0.41514,0.19523,1.90367,2.00604,1.67183,0.0001274,0.0001274,0.0001274
13,20269.7,1.82566,1.82624,1.61116,0.47848,0.42278,0.41676,0.19604,1.90215,1.99655,1.67031,0.0001208,0.0001208,0.0001208
14,21817.7,1.82439,1.81518,1.60888,0.47767,0.42586,0.41828,0.1964,1.90105,1.99369,1.66864,0.0001142,0.0001142,0.0001142
15,23368.8,1.82064,1.80426,1.60067,0.47929,0.42667,0.4186,0.19656,1.90045,1.99347,1.66676,0.0001076,0.0001076,0.0001076
16,24915.5,1.8163,1.79293,1.59585,0.48018,0.42588,0.41945,0.19721,1.89949,1.99012,1.6662,0.000101,0.000101,0.000101
17,26470,1.8123,1.78097,1.59331,0.47672,0.42879,0.42051,0.19767,1.89867,1.98519,1.66507,9.44e-05,9.44e-05,9.44e-05
18,28009.4,1.80579,1.76938,1.58936,0.47905,0.4289,0.42155,0.19825,1.89832,1.98172,1.66455,8.78e-05,8.78e-05,8.78e-05
19,29557,1.80335,1.75999,1.58495,0.47909,0.43136,0.4221,0.19847,1.89788,1.97699,1.66399,8.12e-05,8.12e-05,8.12e-05
20,31124.4,1.80078,1.74947,1.58169,0.47982,0.43136,0.42291,0.19877,1.89796,1.97239,1.66404,7.46e-05,7.46e-05,7.46e-05
21,32603.4,1.82458,1.67555,1.57912,0.48046,0.43031,0.42391,0.19932,1.89713,1.96833,1.66412,6.8e-05,6.8e-05,6.8e-05
22,34083.2,1.81651,1.65738,1.56938,0.48206,0.42892,0.42492,0.19987,1.89701,1.96345,1.66432,6.14e-05,6.14e-05,6.14e-05
23,35563.4,1.81284,1.64011,1.56501,0.48499,0.42621,0.42576,0.20035,1.89677,1.96106,1.66426,5.48e-05,5.48e-05,5.48e-05
24,37042.5,1.81023,1.62658,1.55947,0.48839,0.42535,0.42615,0.20078,1.89674,1.95729,1.66411,4.82e-05,4.82e-05,4.82e-05

[glenn-jocher]

Thanks — from these logs metrics/mAP50-95(B) plateaus around 0.20 by epochs 20-24 while train/* and val/* losses flatten together, so this looks more like a data/task ceiling than an optimizer issue; the fastest next check is to overfit 20–50 carefully labeled images, and if it still cannot fit them nearly perfectly, relabel the hardest classes or switch to classification or segmentation if the disease affects most of the leaf rather than forming tight localized lesions.