The MAP of auther's pre-trained model EfficientDet-d0 on VOC is about 0.56?

[majian-stu]

I have run the script eval.py with the model EfficientDet-d0 offered by author, but the MAP is only about 0.56. Do you know why? I hope you can help me. Thank you!

[linerfei]

My results is 0.56 either, not 0.62. By the way, I trained the network without pre trained network, and the results is lower than 0.36. SGD or adamW reaches the same results , why?

[majian-stu]

My results is 0.56 either, not 0.62. By the way, I trained the network without pre trained network, and the results is lower than 0.36. SGD or adamW reaches the same results , why?

Thank you for your reply. I also encountered this problem when training for both VOC and COCO dataset. Although the loss has been reduced to a very low level, mAP is poor.

[quangtn266]

My results is 0.56 either, not 0.62. By the way, I trained the network without pre trained network, and the results is lower than 0.36. SGD or adamW reaches the same results , why?

Thank you for your reply. I also encountered this problem when training for both VOC and COCO dataset. Although the loss has been reduced to a very low level, mAP is poor.

I got the same issue. After trying to fix this issue for 1 month, I decide to change another model :))

[majian-stu]

My results is 0.56 either, not 0.62. By the way, I trained the network without pre trained network, and the results is lower than 0.36. SGD or adamW reaches the same results , why?

Thank you for your reply. I also encountered this problem when training for both VOC and COCO dataset. Although the loss has been reduced to a very low level, mAP is poor.

I got the same issue. After trying to fix this issue for 1 month, I decide to change another model :))

That‘s a sad story. I really hope the author can reply us...

[linerfei]

@majian-stu ,have you used pre trained efficientNet-B0 ? I changed the code 'from_pretrained('efficientnet-b0') ' and get from_name('efficientnet-b0') instead. I think this may lead to lower ap.

[majian-stu]

@majian-stu ,have you used pre trained efficientNet-B0 ? I changed the code 'from_pretrained('efficientnet-b0') ' and get from_name('efficientnet-b0') instead. I think this may lead to lower ap.

Yes, I used the params from pre-trained model efficientnet-b0, Do you means I should use the params from efficientdet-d0 feeding into backnone?

[majian-stu]

@majian-stu ,have you used pre trained efficientNet-B0 ? I changed the code 'from_pretrained('efficientnet-b0') ' and get from_name('efficientnet-b0') instead. I think this may lead to lower ap.

Yes, I used the params from pre-trained model efficientnet-b0, Do you means I should use the params from efficientdet-d0 feeding into backnone?

I have change the code "self.backbone = EfficientNet.from_pretrained(MODEL_MAP[network])" to "self.backbone = EfficientNet.from_name(MODEL_MAP[network])", and run eval.py again, but the mAP is still 0.56.

[kentaroy47]

the accuracy is quite strange. our code provides mAP of 77%, but its still low. (should get ~85% with the paper's reported accuracy). Maybe adding image augumentations will get some accuracy boost? I think lots of lr scheduling configurations will be required to get good performance on coco. https://github.com/kentaroy47/efficientdet.pytorch

[y78h11b09]

My model of efficient-d0 trained 255 epoches provides mAP of 74.8% on VOC 07, and 70.7% on VOC12

[xpngzhng]

@kentaroy47 hi, I have read your repo. Have you try coco dataset? I made some modifications to the bifpn implementation in this repo and add it to mmdetection, using efficientdet-d2 I obtain mAP 0.347

[kentaroy47]

@xpngzhng I actually haven't finished coco training. That's a good coco accuracy even compared to faster-rcnns!

[djaym7]

My model of efficient-d0 trained 255 epoches provides mAP of 74.8% on VOC 07, and 70.7% on VOC12

For testing on 2007 did you use 2007+2012 trainval for training the model ?