Yolov5: Training and validation accuracy and loss from result and graph

[John12Reaper]

❔Question

Sir,

1. I wanted to know which are the training accuracy and validation accuracy
   and also training loss and validation loss in the results.txt?

1. How can I plot the training and validation accuracy in a single graph
   and training and validation loss in another graph?

Here is the result

Please it would be really helpful.
Thank You

Additional context

[glenn-jocher]

@John12Reaper accuracy is a classification metric. YOLOv5 does object detection. The metrics available are displayed to screen, saved to results.txt, plotted to results.png after training completion and logged to Tensorboard.

Suggest you start with the custom training tutorial, where all of this is explained.
https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data

[glenn-jocher]

@John12Reaper this explains the difference:

[John12Reaper]

@glenn-jocher Thank You Sir
So this means there is no training accuracy or validation accuracy in results.txt
Then what is the mAP@ : .5 and mAP@ .5 : .95 ?
And Sir What does all the matrix mean in the results.txt kindly if you tell or suggest any?
Sorry Sir I am new to object detection.

[dongjuns]

Yes, you are right.
There is no training accuracy or validation accuracy metric, but an mAP metric on your validation dataset.
And I think you need some detailed information about mAP,
https://jonathan-hui.medium.com/map-mean-average-precision-for-object-detection-45c121a31173

[John12Reaper]

@dongjuns Thank You, Sir
I really appreciate your kind help.
Sir, please could you tell me how to print of val AP and train AP in a single graph
or val Loss and train Loss in a single graph?
It will be quite helpful.

[dongjuns]

@John12Reaper You're welcome, look into these functions.
https://github.com/ultralytics/yolov5/blob/c8c5ef36c9a19c7843993ee8d51aebb685467eca/utils/general.py#L1263
https://github.com/ultralytics/yolov5/blob/c8c5ef36c9a19c7843993ee8d51aebb685467eca/utils/general.py#L1240

And you can draw training loss and validation loss in a single graph like this.

Move your results.txt file into your YOLOv5 directory, I'm using docker and in my case, YOLOv5 directory path is /usr/src/app.
Then you can get your results.png with this script.

root@21e424cd396c:/usr/src/app# ls
Dockerfile   detect.py   requirements.txt  road_detect.py  train.py
LICENSE      hubconf.py    runs     yolov5x.pt
README.md    inference     sotabench.py     tutorial.ipynb
__pycache__  merge.py     test.py       utils
data         models      results.txt    weights

root@21e424cd396c:/usr/src/app# python
Python 3.6.10 |Anaconda, Inc.| (default, May  8 2020, 02:54:21) 
[GCC 7.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from utils.general import *
>>> plot_results_overlay()

root@21e424cd396c:/usr/src/app# ls
Dockerfile   detect.py    requirements.txt    results.txt    weights
LICENSE      hubconf.py    road_detect.py    train.py       
README.md    inference     runs    yolov5x.pt
__pycache__  merge.py      sotabench.py    tutorial.ipynb
data         models      results.png       test.py        utils

[John12Reaper]

@dongjuns Sir, You are Great.
I have got the results.png as given below:

Sir, Thank You again.
I have also seen the functions and made some changes to get other relational graphs.
but Sir, One last favor, could you please help?
Is there any way to make these graphs in Tensorboard? just the above graphs.
Cause in TensorBoard it is quite smooth for the evaluation of the performance.

[dongjuns]

@John12Reaper Don't say 'Sir' please!!! :sob:
Anyway, I don't use tensorboard but maybe you could get an idea from this link and train.py file.
https://github.com/ultralytics/yolov5/blob/master/tutorial.ipynb

Good luck-!

[glenn-jocher]

@John12Reaper tensorboard automatically logs all trainings. Tensorboard Instructions are displayed at the beginning of every training run:

Start Tensorboard with "tensorboard --logdir runs/", view at http://localhost:6006/

[ZwNSW]

@dongjuns I watched the conversation between you and@John12Reaper, but I don’t think I understand what you said, what should I do to get the images of val and train. Or how to modify the code to get orange lines and blue lines? Mine only has blue lines, and no validation set lines.

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.