re-implementation of Retinanet detection : https://arxiv.org/abs/1708.02002
- Python 3.7
- Numpy
- pytorch >= 1.5.0
- batch size : 16
- optimizer : SGD
- epoch : 30
- initial learning rate 0.01
- weight decay : 5e-4
- momentum : 0.9
- scheduler : cosineannealing LR (min : 5e-5)
- Dataset
- COCO
- VOC
- Model
- Loss (Focal loss and smooth l1 loss)
- Coder
- Distributed training (distributed data parallel)
- VOC
start..evaluation
90.25% = aeroplane AP
84.79% = bicycle AP
87.98% = bird AP
76.75% = boat AP
63.18% = bottle AP
90.86% = bus AP
92.45% = car AP
95.38% = cat AP
66.37% = chair AP
91.59% = cow AP
78.50% = diningtable AP
92.67% = dog AP
90.68% = horse AP
86.94% = motorbike AP
84.95% = person AP
59.69% = pottedplant AP
89.97% = sheep AP
81.69% = sofa AP
89.28% = train AP
83.85% = tvmonitor AP
mAP = 83.89%
- COCO
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.345
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.533
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.369
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.181
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.383
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.493
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.307
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.464
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.488
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.291
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.545
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.669
mAP : 0.3446613792728923
mean Loss : 0.3393706703444406
Eval Time : 253.8811
| methods | Traning Dataset | Testing Dataset | Resolution | AP | AP50 | AP75 | Time | Fps |
|---|---|---|---|---|---|---|---|---|
| papers | COCOtrain2017 | COCO test-dev | 600 x 600 | 34.0 | 52.5 | 36.5 | 98 | 10.20 |
| papers | COCOtrain2017 | COCOval2017(minival) | 600 x 600 | 34.3 | 53.2 | 36.9 | 98 | 10.20 |
| ours | COCOtrain2017 | COCO test-dev | 600 x 600 | - | - | - | - | - |
| ours | COCOtrain2017 | COCOval2017(minival) | 600 x 600 | 34.5 | 53.3 | 36.9 | - | 23.03 |
(FPS for RIX3090 1AE)
- Distributed Data Parallel for fully using the gpu memory and decreasing training time
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 465.19.01 Driver Version: 465.19.01 CUDA Version: 11.3 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA GeForce ... Off | 00000000:01:00.0 Off | N/A |
| 72% 69C P2 328W / 350W | 14395MiB / 24268MiB | 98% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 1 NVIDIA GeForce ... Off | 00000000:21:00.0 Off | N/A |
| 65% 65C P2 325W / 350W | 13394MiB / 24268MiB | 99% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
2800~2900 s/epoch
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 465.19.01 Driver Version: 465.19.01 CUDA Version: 11.3 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA GeForce ... Off | 00000000:01:00.0 Off | N/A |
| 70% 69C P2 298W / 350W | 11154MiB / 24268MiB | 98% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 1 NVIDIA GeForce ... Off | 00000000:21:00.0 Off | N/A |
| 62% 63C P2 291W / 350W | 8764MiB / 24268MiB | 83% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
240x s/epoch training time : 285x s/epoch -> 240x s/epoch (improvement about 15 %)
- training
main.py --config ./configs/retinanet_coco_train.txt
main.py --config ./configs/retinanet_voc_train.txt
- testing & demo
1- download "VOC, COCO" best weights from here and here
2- and set saves file like bellows:
logs
|-- retinanet_coco
|-- saves
|-- retinanet_coco.best.pth.tar
|-- retinanet_voc
|-- saves
|-- retinanet_voc.best.pth.tar
3- and run script like bellows:
(test)
test.py --config ./configs/retinanet_coco_test.txt
test.py --config ./configs/retinanet_voc_test.txt
(demo)
demo.py --config ./configs/retinanet_coco_demo.txt
demo.py --config ./configs/retinanet_voc_demo.txt
https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Object-Detection
https://github.com/NVIDIA/retinanet-examples
https://github.com/yhenon/pytorch-retinanet
https://github.com/liangheming/retinanetv1
If you found this implementation and pretrained model helpful, please consider citation
@misc{csm-kr_retinanet_pytorch,
author={Sungmin, Cho},
publisher = {GitHub},
title={retinanet_pytorch},
url={https://github.com/csm-kr/retinanet_pytorch},
year={2022},
}