new ppyolo class

paddlex/cv/__init__.py

@@ -26,6 +26,7 @@ ResNet50 = models.ResNet50
 DarkNet53 = models.DarkNet53
 # detection
 YOLOv3 = models.YOLOv3
+PPYOLO = models.PPYOLO
 #EAST = models.EAST
 FasterRCNN = models.FasterRCNN
 MaskRCNN = models.MaskRCNN

paddlex/cv/models/__init__.py

@@ -38,6 +38,7 @@ from .classifier import HRNet_W18
 from .classifier import AlexNet
 from .base import BaseAPI
 from .yolo_v3 import YOLOv3
+from .ppyolo import PPYOLO
 from .faster_rcnn import FasterRCNN
 from .mask_rcnn import MaskRCNN
 from .unet import UNet

paddlex/cv/nets/detection/yolo_v3.py

@@ -496,15 +496,14 @@ class YOLOv3:
             gt_label = inputs['gt_label']
             gt_score = inputs['gt_score']
             im_size = inputs['im_size']
-            #num_boxes = fluid.layers.shape(gt_box)[1]
-            #im_size_wh = fluid.layers.reverse(im_size, axis=1)
-            #whwh = fluid.layers.concat([im_size_wh, im_size_wh], axis=1)
-            #whwh = fluid.layers.unsqueeze(whwh, axes=[1])
-            #whwh = fluid.layers.expand(whwh, expand_times=[1, num_boxes, 1])
-            #whwh = fluid.layers.cast(whwh, dtype='float32')
-            #whwh.stop_gradient = True
-            #normalized_box = fluid.layers.elementwise_div(gt_box, whwh)
-            normalized_box = gt_box
+            num_boxes = fluid.layers.shape(gt_box)[1]
+            im_size_wh = fluid.layers.reverse(im_size, axis=1)
+            whwh = fluid.layers.concat([im_size_wh, im_size_wh], axis=1)
+            whwh = fluid.layers.unsqueeze(whwh, axes=[1])
+            whwh = fluid.layers.expand(whwh, expand_times=[1, num_boxes, 1])
+            whwh = fluid.layers.cast(whwh, dtype='float32')
+            whwh.stop_gradient = True
+            normalized_box = fluid.layers.elementwise_div(gt_box, whwh)
 
             targets = []
             if self.use_fine_grained_loss:

paddlex/cv/transforms/__init__.py

@@ -91,7 +91,10 @@ def arrange_transforms(model_type, class_name, transforms, mode='train'):
     elif model_type == 'segmenter':
         arrange_transform = seg_transforms.ArrangeSegmenter
     elif model_type == 'detector':
-        arrange_name = 'Arrange{}'.format(class_name)
+        if class_name == "PPYOLO":
+            arrange_name = 'ArrangeYOLOv3'
+        else:
+            arrange_name = 'Arrange{}'.format(class_name)
         arrange_transform = getattr(det_transforms, arrange_name)
     else:
         raise Exception("Unrecognized model type: {}".format(self.model_type))

paddlex/det.py

@@ -17,6 +17,7 @@ from . import cv
 
 FasterRCNN = cv.models.FasterRCNN
 YOLOv3 = cv.models.YOLOv3
+PPYOLO = cv.models.PPYOLO
 MaskRCNN = cv.models.MaskRCNN
 transforms = cv.transforms.det_transforms
 visualize = cv.models.utils.visualize.visualize_detection

tutorials/train/object_detection/ppyolo.py

@@ -43,17 +43,7 @@ eval_dataset = pdx.datasets.VOCDetection(
 num_classes = len(train_dataset.labels)
 
 # API说明: https://paddlex.readthedocs.io/zh_CN/develop/apis/models/detection.html#paddlex-det-yolov3
-model = pdx.det.YOLOv3(
-    num_classes=num_classes,
-    backbone='ResNet50_vd',
-    with_dcn_v2=True,
-    use_coord_conv=True,
-    use_iou_aware=True,
-    use_spp=True,
-    use_drop_block=True,
-    scale_x_y=1.05,
-    use_iou_loss=True,
-    use_matrix_nms=True)
+model = pdx.det.PPYOLO(num_classes=num_classes)
 
 # API说明: https://paddlex.readthedocs.io/zh_CN/develop/apis/models/detection.html#train
 # 各参数介绍与调整说明：https://paddlex.readthedocs.io/zh_CN/develop/appendix/parameters.html
@@ -64,6 +54,5 @@ model.train(
     eval_dataset=eval_dataset,
     learning_rate=0.000125,
     lr_decay_epochs=[210, 240],
-    use_ema=True,
     save_dir='output/ppyolo',
     use_vdl=True)