未验证 提交 86d11239 编写于 作者: F Feng Ni 提交者: GitHub

add p2 and largesize visdrone model (#6213)

* add p2 and largesize visdrone model, test=document_fix

* add size doc, test=document_fix
上级 43e3b436
......@@ -7,11 +7,16 @@ PaddleDetection团队提供了针对VisDrone-DET小目标数航拍场景的基
| 模型 | COCOAPI mAP<sup>val<br>0.5:0.95 | COCOAPI mAP<sup>val<br>0.5 | COCOAPI mAP<sup>test_dev<br>0.5:0.95 | COCOAPI mAP<sup>test_dev<br>0.5 | MatlabAPI mAP<sup>test_dev<br>0.5:0.95 | MatlabAPI mAP<sup>test_dev<br>0.5 | 下载 | 配置文件 |
|:---------|:------:|:------:| :----: | :------:| :------: | :------:| :----: | :------:|
|PP-YOLOE-s| 23.5 | 39.9 | 19.4 | 33.6 | 23.68 | 40.66 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_s_80e_visdrone.pdparams) | [配置文件](./ppyoloe_crn_s_80e_visdrone.yml) |
|PP-YOLOE-l| 29.8 | 48.3 | 23.0 | 38.6 | 27.29 | 45.52 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_80e_visdrone.pdparams) | [配置文件](./ppyoloe_crn_l_80e_visdrone.yml) |
|PP-YOLOE-l| 29.2 | 47.3 | 23.5 | 39.1 | 28.00 | 46.20 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_80e_visdrone.pdparams) | [配置文件](./ppyoloe_crn_l_80e_visdrone.yml) |
|PP-YOLOE-P2-l| 30.0 | 49.2 | 24.1 | 40.9 | 28.47 | 48.16 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyoloe_p2_crn_l_80e_visdrone.pdparams) | [配置文件](./ppyoloe_p2_crn_l_80e_visdrone.yml) |
|PP-YOLOE-l largesize| 40.3 | 63.5 | 31.3 | 51.8 | 36.13 | 59.96 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_80e_visdrone_largesize.pdparams) | [配置文件](./ppyoloe_crn_l_80e_visdrone_largesize.yml) |
**注意:**
- PP-YOLOE模型训练过程中使用8 GPUs进行混合精度训练,如果**GPU卡数**或者**batch size**发生了改变,你需要按照公式 **lr<sub>new</sub> = lr<sub>default</sub> * (batch_size<sub>new</sub> * GPU_number<sub>new</sub>) / (batch_size<sub>default</sub> * GPU_number<sub>default</sub>)** 调整学习率。
- 具体使用教程请参考[ppyoloe](../configs/ppyoloe#getting-start)
- 具体使用教程请参考[ppyoloe](../ppyoloe#getting-start)
- PP-YOLOE-P2是指增加P2层(1/4下采样层)的特征,共输出4个PPYOLOEHead。
- largesize是指使用以1600尺度为基础的多尺度训练和1920尺度预测,相应的训练batch_size也减小,以速度来换取高精度。
- MatlabAPI测试是使用官网评测工具[VisDrone2018-DET-toolkit](https://github.com/VisDrone/VisDrone2018-DET-toolkit)
## 引用
......
......@@ -18,7 +18,7 @@ TrainReader:
epoch: 80
LearningRate:
base_lr: 0.001
base_lr: 0.01
schedulers:
- !CosineDecay
max_epochs: 96
......
_BASE_: [
'ppyoloe_crn_l_80e_visdrone.yml',
]
weights: output/ppyoloe_crn_l_80e_visdrone_largesize/model_final
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams
TrainReader:
batch_size: 2
LearningRate:
base_lr: 0.0025
worker_num: 2
eval_height: &eval_height 1920
eval_width: &eval_width 1920
eval_size: &eval_size [*eval_height, *eval_width]
TrainReader:
sample_transforms:
- Decode: {}
- RandomDistort: {}
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
- RandomCrop: {}
- RandomFlip: {}
batch_transforms:
- BatchRandomResize: {target_size: [1024, 1088, 1152, 1216, 1280, 1344, 1408, 1472, 1536, 1600, 1664, 1728, 1792, 1856, 1920], random_size: True, random_interp: True, keep_ratio: False}
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
- Permute: {}
- PadGT: {}
batch_size: 2
shuffle: true
drop_last: true
use_shared_memory: true
collate_batch: true
EvalReader:
sample_transforms:
- Decode: {}
- Resize: {target_size: *eval_size, keep_ratio: False, interp: 2}
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
- Permute: {}
batch_size: 2
TestReader:
inputs_def:
image_shape: [3, *eval_height, *eval_width]
sample_transforms:
- Decode: {}
- Resize: {target_size: *eval_size, keep_ratio: False, interp: 2}
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
- Permute: {}
batch_size: 1
......@@ -18,7 +18,7 @@ TrainReader:
epoch: 80
LearningRate:
base_lr: 0.001
base_lr: 0.01
schedulers:
- !CosineDecay
max_epochs: 96
......
_BASE_: [
'ppyoloe_crn_l_80e_visdrone.yml',
]
weights: output/ppyoloe_p2_crn_l_80e_visdrone/model_final
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams
TrainReader:
batch_size: 4
LearningRate:
base_lr: 0.005
CSPResNet:
return_idx: [0, 1, 2, 3]
CustomCSPPAN:
out_channels: [768, 384, 192, 64]
PPYOLOEHead:
fpn_strides: [32, 16, 8, 4]
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册