Error performing object detection with yolov5 and converted roboflow dataset

Question

in this notebook:

https://colab.research.google.com/drive/1omLGQarT6-WOyQgoXK7d98rj01Ietphm#scrollTo=412Aa8kIs0mi

I am trying to run yolov5 to operate object detection. First I download a dataset (originally created in coco format, then converted in yolov5 format). However when I run the code I got an exception (and before a warning)

AssertionError: train: No labels in /content/datasets/roboflow/train/labels.cache. Can not train without labels.

I think normally folders train/test/validation should contains a "images" and a "labels" (I got from roboflow subfolders with "labelTxt" name), so I created renaming from "labelTxt" to "labels" (see second notebook link).

Then instead I got warnings like

train: WARNING: /content/datasets/roboflow/train/images/hybrid_rgb_0297__volcano_39_ero1_ecc1_rcrat1_png.rf.ab11b2a7ee4ef956fa1439b144057bc9.jpg: ignoring corrupt image/label: could not convert string to float: 'volcano'
train: WARNING: /content/datasets/roboflow/train/images/hybrid_rgb_0305__mountain_0_ero0_ecc0_rcrat0_png.rf.ece26ced043322846c804623ce4a44af.jpg: ignoring corrupt image/label: could not convert string to float: 'mountain'

Now the strings 'volcano' and 'mountain' are my class names, as from data.yaml content:

path: ../datasets/roboflow

train: train/images
val: valid/images
test: test/images

nc: 3
names: ['done', 'mountain', 'volcano']

Any idea? Thank you.

---

EDIT: yolov5 can operate also on instance segmentation, as explained in https://blog.roboflow.com/train-yolov5-instance-segmentation-custom-dataset/ (right now I changed approach).

Answer 1

YOLOv5 is an object detection model. May I ask why you’re using it for segmentation?

You’d want to create an instance segmentation project on Roboflow and then train it. For example: https://blog.roboflow.com/detectron2-custom-instance-segmentation/

Additionally, if you were attempting object detection and said segmentation by mistake, here’s a tutorial on training with YOLOv5 - be sure all images with labels were added to your dataset when you generate the version for model training: Starting Your First Project on Roboflow

Training a YOLOv5 model: https://help.roboflow.com/en_US/roboflow-train/model-training-yolov5-video-tutorial

Answer 2

I assume that the dataset that you are using is corrupted. The file name consists of .png and .jpg extension as well,which is weird. I'll suggest you to download your dataset as zip file and then upload it to the contents/ which will not compromise the image quality also.

And also why your data.yaml shows different classes instead of mask? Is that also an error? Kindly provide more info for more precise answers.

Answer 3

This happens when your labels are not valid for training or they are not in the path. If the path is the problem, try giving absolute paths for them:

#path:  
train: /content/drive/MyDrive/yolov8/images  # train 
val: /content/drive/MyDrive/yolov8/images/val # val
# test:  # test images (optional)

For me, it was because of using labelImg for annotations, and it had some predefined classes. Therefore, I changed the dataset.yaml classes and therefore the config would go wrong.