Update ConvNeXt models (#6363)

* update convnext yolox ppyoloe * fix cfg pretrain weights * fix pdparams

Update ConvNeXt models (#6363)
* update convnext yolox ppyoloe * fix cfg pretrain weights * fix pdparams
b9307fff · Feng Ni · GitHub · 546cada6 · b9307fff · b9307fff
7 changed file
--- a/configs/convnext/README.md
+++ b/configs/convnext/README.md
+# ConvNeXt (A ConvNet for the 2020s)
+
+## 模型库
+### ConvNeXt on COCO
+
+| 网络网络                  | 输入尺寸 | 图片数/GPU | 学习率策略 | mAP<sup>val<br>0.5:0.95 | mAP<sup>val<br>0.5 | Params(M) | FLOPs(G) |    下载链接       | 配置文件 |
+| :------------- | :------- | :-------: | :------: | :------------: | :---------------------: | :----------------: |:---------: | :------: |:---------------: |
+| PP-YOLOE-ConvNeXt-tiny | 640 |    16      |   36e    |  44.6  |  63.3 |  33.04  |  13.87 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyoloe_convnext_tiny_36e_coco.pdparams) | [配置文件](./ppyoloe_convnext_tiny_36e_coco.yml) |
+| YOLOX-ConvNeXt-s       | 640 |    8       |   36e    |  44.6  |  65.3 |  36.20  |  27.52 | [下载链接](https://paddledet.bj.bcebos.com/models/yolox_convnext_s_36e_coco.pdparams) | [配置文件](./yolox_convnext_s_36e_coco.yml) |
+
+
+## Citations
+```
+@Article{liu2022convnet,
+  author  = {Zhuang Liu and Hanzi Mao and Chao-Yuan Wu and Christoph Feichtenhofer and Trevor Darrell and Saining Xie},
+  title   = {A ConvNet for the 2020s},
+  journal = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
+  year    = {2022},
+}
+```
--- a/configs/convnext/ppyoloe_convnext_tiny_36e_coco.yml
+++ b/configs/convnext/ppyoloe_convnext_tiny_36e_coco.yml
+_BASE_: [
+  '../datasets/coco_detection.yml',
+  '../runtime.yml',
+  '../ppyoloe/_base_/ppyoloe_crn.yml',
+  '../ppyoloe/_base_/ppyoloe_reader.yml',
+]
+depth_mult: 0.25
+width_mult: 0.50
+
+log_iter: 100
+snapshot_epoch: 5
+weights: output/ppyoloe_convnext_tiny_36e_coco/model_final
+pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/convnext_tiny_22k_224.pdparams
+
+
+YOLOv3:
+  backbone: ConvNeXt
+  neck: CustomCSPPAN
+  yolo_head: PPYOLOEHead
+  post_process: ~
+
+ConvNeXt:
+  arch: 'tiny'
+  drop_path_rate: 0.4
+  layer_scale_init_value: 1.0
+  return_idx: [1, 2, 3]
+
+
+PPYOLOEHead:
+  static_assigner_epoch: 12
+  nms:
+    nms_top_k: 10000
+    keep_top_k: 300
+    score_threshold: 0.01
+    nms_threshold: 0.7
+
+
+TrainReader:
+  batch_size: 16
+
+
+epoch: 36
+LearningRate:
+  base_lr: 0.0002
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [36]
+    use_warmup: false
+
+OptimizerBuilder:
+  regularizer: false
+  optimizer:
+    type: AdamW
+    weight_decay: 0.0005
--- a/configs/yolox/yolox_convnext_s_36e_coco.yml
+++ b/configs/yolox/yolox_convnext_s_36e_coco.yml
 _BASE_: [
  '../datasets/coco_detection.yml',
  '../runtime.yml',
-  './_base_/yolox_convnext.yml',
-  './_base_/yolox_reader.yml'
+  '../yolox/_base_/yolox_cspdarknet.yml',
+  '../yolox/_base_/yolox_reader.yml'
 ]
 depth_mult: 0.33
 width_mult: 0.50
@@ -10,7 +10,21 @@ width_mult: 0.50
 log_iter: 100
 snapshot_epoch: 5
 weights: output/yolox_convnext_s_36e_coco/model_final
-pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/convnext_tiny_1k_224_ema.pdparams
+pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/convnext_tiny_22k_224.pdparams
+
+
+YOLOX:
+  backbone: ConvNeXt
+  neck: YOLOCSPPAN
+  head: YOLOXHead
+  size_stride: 32
+  size_range: [15, 25] # multi-scale range [480*480 ~ 800*800]
+
+ConvNeXt:
+  arch: 'tiny'
+  drop_path_rate: 0.4
+  layer_scale_init_value: 1.0
+  return_idx: [1, 2, 3]


 TrainReader:
@@ -19,7 +33,7 @@ TrainReader:


 YOLOXHead:
-  l1_epoch: 20
+  l1_epoch: 30
  nms:
    name: MultiClassNMS
    nms_top_k: 10000

--- a/configs/yolox/README.md
+++ b/configs/yolox/README.md
@@ -22,13 +22,13 @@

 | 网络网络        | 输入尺寸   | 图片数/GPU | 学习率策略 | 模型推理耗时(ms) | mAP<sup>val<br>0.5:0.95 | mAP<sup>val<br>0.5 | Params(M) | FLOPs(G) |    下载链接       | 配置文件 |
 | :------------- | :------- | :-------: | :------: | :------------: | :---------------------: | :----------------: |:---------: | :------: |:---------------: |:-----: |
-| YOLOXv2-tiny    |  416     |    8      |   300e    |     1.9    |  32.4  |  50.2 |  5.03 |  6.33  | [下载链接](https://paddledet.bj.bcebos.com/models/yoloxv2_tiny_300e_coco.pdparams) | [配置文件](./yoloxv2_tiny_300e_coco.yml) |
+| YOLOX-cdn-tiny    |  416     |    8      |   300e    |     1.9    |  32.4  |  50.2 |  5.03 |  6.33  | [下载链接](https://paddledet.bj.bcebos.com/models/yolox_cdn_tiny_300e_coco.pdparams) | [配置文件](./yolox_cdn_tiny_300e_coco.yml) |
 | YOLOX-crn-s     |  640     |    8      |   300e    |     3.0    |  40.4  |  59.6 |  7.7  |  24.69 | [下载链接](https://paddledet.bj.bcebos.com/models/yolox_crn_s_300e_coco.pdparams) | [配置文件](./yolox_crn_s_300e_coco.yml) |
-| YOLOX-ConvNeXt-s|  640     |    8      |   36e     |     -      |  **44.6**  |  **65.3** |  36.2 |  27.52 | [下载链接](https://paddledet.bj.bcebos.com/models/yolox_convnext_s_36e_coco.pdparams) | [配置文件](./yolox_convnext_s_36e_coco.yml) |
+| YOLOX-ConvNeXt-s|  640     |    8      |   36e     |     -      |  **44.6**  |  **65.3** |  36.2 |  27.52 | [下载链接](https://paddledet.bj.bcebos.com/models/yolox_convnext_s_36e_coco.pdparams) | [配置文件](../convnext/yolox_convnext_s_36e_coco.yml) |


 **注意:**
-  - YOLOX模型训练使用COCO train2017作为训练集，YOLOXv2表示使用与YOLOv5 releases v6.0之后版本相同的主干网络，YOLOX-crn表示使用与PPYOLOE相同的主干网络CSPResNet，YOLOX-ConvNeXt表示使用ConvNeXt作为主干网络；
+  - YOLOX模型训练使用COCO train2017作为训练集，YOLOX-cdn表示使用与YOLOv5 releases v6.0之后版本相同的主干网络，YOLOX-crn表示使用与PPYOLOE相同的主干网络CSPResNet，YOLOX-ConvNeXt表示使用ConvNeXt作为主干网络；
  - YOLOX模型训练过程中默认使用8 GPUs进行混合精度训练，默认每卡batch_size为8，默认lr为0.01为8卡总batch_size=64的设置，如果**GPU卡数**或者每卡**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；
  - 为保持高mAP的同时提高推理速度，可以将[yolox_cspdarknet.yml](_base_/yolox_cspdarknet.yml)中的`nms_top_k`修改为`1000`，将`keep_top_k`修改为`100`，将`score_threshold`修改为`0.01`，mAP会下降约0.1~0.2%；
  - 为快速的demo演示效果，可以将[yolox_cspdarknet.yml](_base_/yolox_cspdarknet.yml)中的`score_threshold`修改为`0.25`，将`nms_threshold`修改为`0.45`，但mAP会下降较多；

--- a/configs/yolox/_base_/yolox_convnext.yml
+++ b/configs/yolox/_base_/yolox_convnext.yml
-_BASE_: [
-  './yolox_cspdarknet.yml',
-]
-
-YOLOX:
-  backbone: ConvNeXt
-  neck: YOLOCSPPAN
-  head: YOLOXHead
-  size_stride: 32
-  size_range: [15, 25] # multi-scale range [480*480 ~ 800*800]
-
-ConvNeXt:
-  arch: 'tiny'
-  drop_path_rate: 0.4
-  layer_scale_init_value: 1.0
-  return_idx: [1, 2, 3]
--- a/configs/yolox/yoloxv2_tiny_300e_coco.yml
+++ b/configs/yolox/yoloxv2_tiny_300e_coco.yml
 _BASE_: [
  'yolox_tiny_300e_coco.yml'
 ]
-weights: output/yoloxv2_tiny_300e_coco/model_final
+depth_mult: 0.33
+width_mult: 0.375
+
+log_iter: 100
+snapshot_epoch: 10
+weights: output/yolox_cdn_tiny_300e_coco/model_final

 CSPDarkNet:
  arch: "P5" # using the same backbone of YOLOv5 releases v6.0 and later version

--- a/configs/yolox/yolox_crn_s_300e_coco.yml
+++ b/configs/yolox/yolox_crn_s_300e_coco.yml
@@ -11,6 +11,8 @@ width_mult: 0.50
 log_iter: 100
 snapshot_epoch: 10
 weights: output/yolox_crn_s_300e_coco/model_final
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/CSPResNetb_s_pretrained.pdparams
+

 YOLOX:
  backbone: CSPResNet