EfficientDet.Pytorch
EfficientDet.Pytorch copied to clipboard
toandaominh1997
No boxes to NMS!!!!!!!!!!!!!!!!!!!
The code can not be used at all. It can be trained and tested. However, demo.py prompt:"no boxes to NMS", and eval. Py prompt:"AVG map: 0.0"
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%
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%
hi,can u infer an image correctly?
I had solved this problem by downgrading the version of torchvision from 0.4.2 to 0.4.0
in the readme.md, it says like that: PyTorch 1.3+ Torchvision 0.4.0+ (We need high version because Torchvision support nms now.) while the torchvision==0.4.0 should match torch==1.2.0, so which version of you torch？
in the readme.md, it says like that: PyTorch 1.3+ Torchvision 0.4.0+ (We need high version because Torchvision support nms now.) while the torchvision==0.4.0 should match torch==1.2.0, so which version of you torch？
pytorch will downgrade to 1.2.0 meanwhile, but it still work.
in the readme.md, it says like that: PyTorch 1.3+ Torchvision 0.4.0+ (We need high version because Torchvision support nms now.) while the torchvision==0.4.0 should match torch==1.2.0, so which version of you torch？
pytorch will downgrade to 1.2.0 meanwhile, but it still work. I changed my torch=1.2.0 and torchvision=0.4.0, after trained 10 epoch, its ap is still 0.0 and no boxes to nms, do you have any ideas?
in the readme.md, it says like that: PyTorch 1.3+ Torchvision 0.4.0+ (We need high version because Torchvision support nms now.) while the torchvision==0.4.0 should match torch==1.2.0, so which version of you torch？
pytorch will downgrade to 1.2.0 meanwhile, but it still work. I changed my torch=1.2.0 and torchvision=0.4.0, after trained 10 epoch, its ap is still 0.0 and no boxes to nms, do you have any ideas?
I wrote a blog: https://www.cnblogs.com/xiaoyh/p/12201934.html, I hope to help you.
thanks, just like the blog you wrote, change this line parser.add_argument('--network', default='efficientdet-d0', type=str, help='efficientdet-[d0, d1, ..]') to: parser.add_argument('--network', default='efficientdet-d1', type=str, help='efficientdet-[d0, d1, ..]') am i right? btw, i trained my own datasets from scratch.
Same issue here; I tried downgrading torchvision to 0.4.0 and using efficientdet-d1, but it still shows "No boxes to NMS"
Edit: After further inspection, it seems that during validation, the model actually can predict the boxes, but only a portion of the time. So, at least on my part, it just has to do with training the model more and cleaning up my dataset so that the model can predict better.
Edit2: The model didn't improve even after 150 epochs, so I switched to a different implentation (see below)
Same issue here; I tried downgrading torchvision to 0.4.0 and using efficientdet-d1, but it still shows "No boxes to NMS"
Edit: After further inspection, it seems that during validation, the model actually can predict the boxes, but only a portion of the time. So, at least on my part, it just has to do with training the model more and cleaning up my dataset so that the model can predict better.
Have you any updated for your result? because I'm also training on d1 model and getting low results for map.
@quangtn266 Yes so I tried training more, but even after 150 epochs there was still no improvement. So, I moved over to this different implementation: https://github.com/signatrix/efficientdet
No offense intended for toandaominh, but the one I switched to works insanely well and is much more effective; I didn't have any issues with the other implementation.
Hope this helps
pytroch 1.1 and torchvision 0.3.0 . only d0 cant train normaly ,
you mean that model from d1 to d7 can train normally and only d0 meets the problem, right?
i can train normal now , see the code of 'class EfficientDet' . maybe do test after 10 or more epochs can work well , but at the same time , i found that this Implementation have lots of difference from the paper.
The official implementation is out! https://github.com/google/automl/tree/master/efficientdet