Wrong custom YoloV5 labels about deepstream-yolo HOT 5 CLOSED

Now I understand what is happening, and this comment is just to register the solution (thanks to @marcoslucianops).

In your config_infer_primary_yolo.txt, the model-engine-file always prevails over onnx-file. And in fact, despite my previous ignorance, it makes sense. The .engine file set by model-engine-file is built from the onnx-file, and the former is the one ran by Deepstream. Therefore, if .engine file exists, its is reasonable to Deepstream that it doesn't have to build it at all, just need to run it. I makes model loading pretty much faster and onnx-file is ignored.

from deepstream-yolo.

You need to remove the engine file and generate the new one for the custom model.

from deepstream-yolo.

OK, just in case I removed all engine files and generated a new one, but the problem persists.

Take a look at this extract from label files of yolov5s.pt and my custom model epi.pt

yolov5s.pt labels   epi.pt labels
person                    person
bicycle                   helmet
car                       suit
motorcycle                boot
airplane                  gloves
bus                       mask
train                     glasses
truck                     earplug
boat
.....

And this is what I obtain when exporting epi.pt to epi.onnx. The DeepStream-Yolo folder now contains just the epi.onnx and label.txt. This is what I get:

Clearly, it is using the original weights, the ones from yolov5s. For instance, it detects an object whose class are bicycles ("1") and by label file says it is helmet, or an object belonging to trucks ("7") and my label file says it is an earplug. It look like it is ignoring my training and using just the original weights from yolov5s.

When I trained my custom model (epi.pt) I made a transfer learning from yolov5s. This is what I did for training my custom model (epi.pt): python3 train.py --img 640 --batch 16 --epochs 1000 --data epi.yaml --weights yolov5s.pt --cache --save-period 10

And this is what my custom model epi.pt returns when I run: python3 ~/yolov5/detect.py --weights epi.pt --source tegra.mp4.

What's going wrong? Is the export script creating an incomplete onnx file?

from deepstream-yolo.

It's a very strange case, if you are using the correct onnx model and changing the labels.txt file, it should work without any issues.

from deepstream-yolo.

As I said, you need to take care about the engine file, because it could run the old file if you don't delete it, even changing the onnx model in the config_infer file.

from deepstream-yolo.