y78h11b09
y78h11b09
one(former) is to filter some anchors for those cls-score less than threshold , anthor(latter) is to reduce some anchors for overlap
my model has 57.1m parameters on a d0 basis with 1601 classes and 401 attributes
dataset VG , network efficientdet-d0, batch_size 2
I have the same problem, I got the hightest mAp just 0.19 after trained 186 epoches
Thank you very much. Yesternight, I also found the code updated. and I found the eval code before updated had a problem, which is red, i.e, for label in range(valid_dataset.num_classes()):...
I think the error is caused for model not trained enough, i got only one box when I load this trained weight to do inference by demo.py
the new avg mAP: 0.43364948014325755 , the model is trained 247 epoaches on VOC0712
yeah, there are some problems, run demo.py, still happen no boxes to NMS, but can got mAP 0.5439399966. and training, the CPU cost is very hight, even over 85%
I also meet the same question, in validating , i changed the valid to 'test', and so, i get the mAp is 0.197 on VOC2012 Val. The model was trained...
> > I meet the same question, and I have printed out all the feature map sizes. The problem I think is miss of resizing the input image. I also...