[Bug]: Exported end2end ONNX model produces poor COCO validation results #146
Description
Before Reporting
-
I have pulled the latest code of main branch to run again and the bug still existed. 我已经拉取了主分支上最新的代码，重新运行之后，问题仍不能解决。
-
I have read the README carefully and no error occured during the installation process. (Otherwise, we recommand that you can ask a question using the Question template) 我已经仔细阅读了README上的操作指引，并且在安装过程中没有错误发生。（否则，我们建议您使用Question模板向我们进行提问）
Search before reporting
OS
Ubuntu 24.04
Device
RTX 4090
CUDA version
12.5
TensorRT version
No response
Python version
3.10
PyTorch version
1.13.1
torchvision version
0.14.1
Describe the bug
I have been attempting to train and export a tinynasL25_s model on the COCO dataset, but getting terrible results from the exported model. I have exported the model end2end, which if I am interpreting the code correctly should have given me at most 100 detections after NMS, but in many cases I am still getting 1000+ detections. Detections do however appear to be filtered by a minimum confidence score of 0.05.
After 60 epochs of training on the COCO dataset I get an evaluation score of:
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.365
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.514
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.394
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.208
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.403
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.488
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.320
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.549
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.613
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.424
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.673
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.771
I export this model using:
python tools/converter.py -f configs/damoyolo_tinynasL25_S.py -c workdirs/damoyolo_tinynasL25_S/epoch_60_ckpt.pth --batch_size 1 --img_size 640 --end2end --ort
But when evaluating the exported model on COCO I get the following results:
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.008
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.017
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.003
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.047
I have debugged the pre-processing and post-processing steps I have added for running the ONNX model and it looks consistent with the demo script. I am using the same COCO validation scripts for YOLOX and RT-DETR models and have no issues with those. Looks to me like something must be wrong with the export script.
To Reproduce
- Train model using configs/damoyolo_tinynasL25_S.py
- Export model to ONNX using: python tools/converter.py -f configs/damoyolo_tinynasL25_S.py -c workdirs/damoyolo_tinynasL25_S/epoch_60_ckpt.pth --batch_size 1 --img_size 640 --end2end --ort
- Evaluate ONNX model on COCO validation set
Hyper-parameters/Configs
No response
Logs
No response
Screenshots
No response
Additional
No response