Open Philharmy-Wang opened 3 years ago
In my test, without pre-training weights, it is 62.5% mAP
In my test, without pre-training weights, it is 62.5% mAP
thank you for your reply! The mAP I tested is 62.57% which is basically consistent with your results
@Philharmy-Wang hi, i also train voc07+12,use yolov4-tiny-3l.cfg, if i directly use official yolov4-tiny.weights as trained weights, is this work? command like this: ./darknet detector train mytrain.data yolov4-tiny-3l.cfg yolov4-tiny.weights -map
@Philharmy-Wang hi, i also train voc07+12,use yolov4-tiny-3l.cfg, if i directly use official yolov4-tiny.weights as trained weights, is this work? command like this: ./darknet detector train mytrain.data yolov4-tiny-3l.cfg yolov4-tiny.weights -map
- This can work if the yolov4-tiny.weights are simply used as pre-training weights. To some extent it allows the convolutional layers in the backbone part of the model to learn the real-world features. However, yolov4-tiny-3l and yolov4-tiny do not have the same model structure in the NECK and HEAD parts, so yolov4-tiny-3l may get wrong initial weights in the deeper part of the model. If the number of iterations of the model is large enough, the model will automatically correct the wrong weights. Overall, it is beneficial to do so.
- This will not work if yolov4-tiny.weights is used as the weight file and yolov4-tiny-3l.cfg is used as the model file to perform detection on images or videos. In the prediction and application phase, the model .cfg file must correspond to the .weights file one by one.
Who knows what map YOLOV4 tested on the VOC dataset?