the different result on tiny-yolo and yolov2

[CuiChangh]

I used the same data to train on tiny-yolo and yolov2. on tiny-yolo,I got a good detection result(recall about 0.8). But on yolov2, even I have trained 40000 batch, the obj is still lower than 0.1, and the result is that I can't detect any object if I set thresh 0.24 I try to change different learn-rate, but no effects. my project: samples : 5000 image, train to detect car and pedestrian

[CuiChangh]

If anyone knows the reason, please tell me. Thanks!

[EnricoBeltramo]

Strange, from my side I had opposite result: yolo 2.0 work well after 2000 iteraction, tiny yolo never.. I attach my version of cfg and names (is a custom version for 4 classe: person, car, cat, dog) yolo2.0.zip

[AlexeyAB]

@CuiChangh Do you use batch=64 subdivision=8 in your cfg-files? I can successfully train both tiny-yolo-voc.cfg and yolo-voc.2.0.cfg on my own data.

[CuiChangh]

@EnricoBeltramo It was strange, I trained several times in the same data and same yolo-detector, I obtained a converged suddenly in the last training.

[CuiChangh]

@AlexeyAB It was strange, I trained several times in the same data and same yolo-detector, I obtained a converged suddenly in the last training.

[AlexeyAB]

@CuiChangh Training is a partially random process - training uses random sequence of images from dataset, and uses random weights initialization for conv-layers whose weights were not loaded from the pre-trained file darknet19_448.conv.23. So training the same model on the same dataset can give you different result for the same number of iteraions.