Yolo training on google colab crashes with error buffer over flow

Issue

I am trying to train a standard unmodified model using yolov3-tiny.cfg . I am using Google Colab because i can’t afford a good GPU to train my data on. I am training on 2 classes. and this error occurred..

The command i pass to Colab:

!./darknet detector train build/darknet/x64/data/obj.data build/darknet/x64/cfg/yolov3-tiny.cfg yolov3-tiny.weights -dont_show

After the model loads i see this output:

* buffer overflow detected *: ./darknet terminated

All I modified in the yolov3-tiny.cfg was classes = 2, anchors, filters = 21
I tried with a different cfg file, yolov3.cfg with no weights and got the same buffer overflow error.
I opened an issue on git: https://github.com/pjreddie/darknet/issues/1472

Here is my colab file:
https://colab.research.google.com/drive/1kdwOl9AJ7A7Bn37YcrSVmwGH8U2EI_GC

I changed to yolov2.cfg with the same training data, same buffer over flow error, so i think it might be the train data or the obj.data file.

Solution

It was all about the formatting of the text files.

The text files were generated on a Windows OS and the Google Colab runs an Ubuntu VM machine, so I formatted the .txt with the line.rstrip('\r').

I got some additional errors like can’t load the .jpg file, it needs relative path from the darknet directory, so in the train.txt the path: /data/obj/1.jpg it isn’t enough it needs /build/darknet/x64/data/obj/1.jpg.

For the moment, I can train yolo using Google Colab on my personal data set.

Answered By – Cristi.Fatu

Answer Checked By – Jay B. (AngularFixing Admin)

Leave a Reply

Your email address will not be published.