FAQs

[zylo117]

Q1. Why the mAP is so low?

A1. Check loss graph for overfitting or underfitting.

Q2. How low is low for losses?

A2. Losses are just losses. It doesn't mean anything directly. You should run coco_eval to get a useful result. But for comparison, here are my losses on training coco with d0, cls loss: 0.25, reg loss 0.25.

Q3. Why are my losses are so low but the mAP is also low?

A3. First refer to Q2. If anchors don't fit your gt boxes, the losses will be falsely low. You should try more suitable anchors. Try this, mnslarcher/kmeans-anchors-ratios

Q4. Why is reg_loss so low but cls_loss doesn't drop a bit for a while?

A4. Category ids must start from 1 as mentioned on readme.

Q5. pycocotools crashes?

A5. Make sure numpy version is not greater than 1.17. I think it's a bug of pycocotools.

Q6. KeyError: '???' when training custom dataset?

A6. Check image_ids. image_ids should be int, not str. Some annotation convertion tools may use str but it's not standard.

Q7. Losses are low, mAP is good enough, but nothing comes out of inference?

A7. Make sure the inference parameters and settings are the same as training and evaluation.

Q8. Shape mismatch warning on training?

[Warning] Ignoring Error(s) in loading state_dict for EfficientDetBackbone:
size mismatch for classifier.header.pointwise_conv.conv.weight: copying a param with shape torch.Size([810, ??, 1, 1]) from checkpoint, the shape in current model is torch.Size([???, ??, 1, 1]).

A8. It's caused by training custom datasets with coco pretrained weights. Don't panic if you see this, this might be because you load a pretrained weights with different number of classes. The rest of the weights should be loaded already.

Q9. Shape mismatch warning on inference?

A9. Be aware, there is no background class in efficientdet.