What are wrong with my custom model?

I have trained my custom object detection model on Colab pro with some images, the training process seems good. But when I deploy the model to DeepStack, the result always are {‘success’: True, ‘predictions’: [], ‘duration’: 0} even I use the train image.
How I can check the mistake.

Thank you for your advice.

What does your data set look like? How many pictures in the train and validation folders? And how many different classes are you trying to identify? Just the three in the sample picture? What model did you use for the training? Yolov5? Which one? What are the training results?

I want to train the model of coral type, this is the second time after the first train almost 7000 pictures with 26 class. I use Yolo5s. It same results, not thing report so I try the second time with 4 classes! and just 30 pictures of each class.

The picture of result.txt attaches to this post.

I don’t know what you expect. The results say that the trained model got almost nothing right:

The best result is only scoring 0.09379 mAP_0.5/0.95on your 30 epocs.

You must have some issues with the quality of the training data, or the classes you are trying to predict are hard for the AI.

When i train my model for 6 classes (different cats) these are my results after 29 epocs:

Have you read all the recommended settings, data preparation guides and so on here:
https://github.com/ultralytics/yolov5 ?

I still thing something is wrong with the dataset, bad data in = bad data out. If you have complex pictures you could probably benefit from trying the Yolov5m or even Yolov5l models. The training is slower but you should give better results.

Also before you begin with another training, have you tried any of the example custom models available to verify that your deepstack detection setup are actually running propperly?
Here are some example models:

Thank you very MissMusic for your suggestion and your example result. I check the setup with the logo example ready. It works properly.
I will back to check the train data again and can check the result of the trained model. Thank you again.