Add cam support name of PPYOLOEWithAuxHead (#7719)

* GradCAM Refinement

* ROI CAM bugfix

* Add support for blazeface, ssd, retinanet

* Change nms return_idx default value to False

* Add use_extra_data attr

* pre-commit code style

* fix yolo extra_data process

* Add cam support name of PPYOLOEWithAuxHead

docs/tutorials/GradCAM_cn.md

@@ -2,7 +2,7 @@
 
 ## 1.简介
 
-基于backbone/roi特征图计算物体预测框的cam(类激活图)
+基于backbone/roi特征图计算物体预测框的cam(类激活图), 目前支持基于FasterRCNN/MaskRCNN系列, PPYOLOE系列, 以及BlazeFace, SSD, Retinanet网络。
 
 ## 2.使用方法
 * 以PP-YOLOE为例，准备好数据之后，指定网络配置文件、模型权重地址和图片路径以及输出文件夹路径，使用脚本调用tools/cam_ppdet.py计算图片中物体预测框的grad_cam热力图。下面为运行脚本示例。
@@ -27,7 +27,7 @@ python tools/cam_ppdet.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml --infer
 </center>
 <br><center>cam_ppyoloe/225.jpg</center></br>
 
-## 3. 目前支持基于FasterRCNN/MaskRCNN, PPYOLOE系列以及BlazeFace, SSD, Retinanet网络。
+## 3. 目前支持基于FasterRCNN/MaskRCNN系列, PPYOLOE系列以及BlazeFace, SSD, Retinanet网络。
 * PPYOLOE网络热图可视化脚本
 ```bash
 python tools/cam_ppdet.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml --infer_img demo/000000014439.jpg --cam_out cam_ppyoloe --target_feature_layer_name model.backbone -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams

docs/tutorials/GradCAM_en.md

 # Object detection grad_cam heatmap
 
 ## 1.Introduction
-Calculate the cam (class activation map) of the object predict bbox based on the backbone/roi feature map
+Calculate the cam (class activation map) of the object predict bbox based on the backbone/roi feature map, currently supports networks based on FasterRCNN/MaskRCNN series, PPYOLOE series and BlazeFace, SSD, Retinanet.
 
 ## 2.Usage
 * Taking PP-YOLOE as an example, after preparing the data, specify the network configuration file, model weight address, image path and output folder path, and then use the script to call tools/cam_ppdet.py to calculate the grad_cam heat map of the prediction box. Below is an example run script.
@@ -27,7 +27,7 @@ python tools/cam_ppdet.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml --infer
 <br><center>cam_ppyoloe/225.jpg</center></br>
 
 
-## 3.Currently supports networks based on FasterRCNN/MaskRCNN, PPYOLOE series and BlazeFace, SSD, Retinanet.
+## 3.Currently supports networks based on FasterRCNN/MaskRCNN series, PPYOLOE series and BlazeFace, SSD, Retinanet.
 * PPYOLOE bbox heat map visualization script (with backbone featuremap)
 ```bash
 python tools/cam_ppdet.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml --infer_img demo/000000014439.jpg --cam_out cam_ppyoloe -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams

ppdet/utils/cam_utils.py

@@ -143,9 +143,9 @@ class BBoxCAM:
         # set for get extra_data before nms
         trainer.model.use_extra_data=True
         # set for record the bbox index before nms
-        if cfg.architecture == 'FasterRCNN' or cfg.architecture == 'MaskRCNN':
+        if cfg.architecture in ['FasterRCNN', 'MaskRCNN']:
             trainer.model.bbox_post_process.nms.return_index = True
-        elif cfg.architecture == 'YOLOv3':
+        elif cfg.architecture in ['YOLOv3', 'PPYOLOE', 'PPYOLOEWithAuxHead']:
             if trainer.model.post_process is not None:
                 # anchor based YOLOs: YOLOv3,PP-YOLO
                 trainer.model.post_process.nms.return_index = True

tools/cam_ppdet.py

@@ -67,7 +67,10 @@ def parse_args():
     return args
 
 def run(FLAGS, cfg):
-    assert cfg.architecture in ['FasterRCNN', 'MaskRCNN', 'YOLOv3', 'BlazeFace', 'SSD', 'RetinaNet'],  'Only supported cam for faster_rcnn based and yolov3 based architecture for now,  the others are not supported temporarily!'
+    assert cfg.architecture in ['FasterRCNN', 'MaskRCNN', 'YOLOv3', 'PPYOLOE',
+                                'PPYOLOEWithAuxHead', 'BlazeFace', 'SSD', 'RetinaNet'],  \
+        'Only supported cam for faster_rcnn based and yolov3 based architecture for now,  ' \
+        'the others are not supported temporarily!'
 
     bbox_cam = BBoxCAM(FLAGS, cfg)
     bbox_cam.get_bboxes_cams()