fix yolo backbone

examples/yolov3/.gitignore

 dataset/voc*
+pretrain_weights/darknet53_pretrained.pdparams

examples/yolov3/README.md

@@ -99,9 +99,17 @@ YOLOv3 的网络结构由基础特征提取网络、multi-scale特征融合层
   |   ...
   ```
 
+### 预训练权重下载
+
+YOLOv3模型训练需下载骨干网络DarkNet53的预训练权重，可通过如下方式下载。
+
+```bash
+sh pretrain_weights/download.sh
+```
+
 ### 模型训练
 
-数据准备完毕后，可使用`main.py`脚本启动训练和评估，如下脚本会自动每epoch交替进行训练和模型评估，并将checkpoint默认保存在`yolo_checkpoint`目录下。
+数据和预训练权重下载完成后，可使用`main.py`脚本启动训练和评估，如下脚本会自动每epoch交替进行训练和模型评估，并将checkpoint默认保存在`yolo_checkpoint`目录下。
 
 YOLOv3模型训练总batch_size为64训练，以下以使用4卡Tesla P40每卡batch_size为16训练介绍训练方式。对于静态图和动态图，多卡训练中`--batch_size`为每卡上的batch_size，即总batch_size为`--batch_size`乘以卡数。
 
@@ -117,7 +125,7 @@ python main.py --help
 使用如下方式进行多卡训练:
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --data=<path/to/dataset> --batch_size=16
+CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --data=<path/to/dataset> --batch_size=16 --pretrain_weights=./pretrain_weights/darknet53_pretrained
 ```
 
 #### 动态图训练
@@ -127,7 +135,7 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --data=
 使用如下方式进行多卡训练:
 
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py -m paddle.distributed.launch --data=<path/to/dataset> --batch_size=16 -d
+CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py -m paddle.distributed.launch --data=<path/to/dataset> --batch_size=16 -d --pretrain_weights=./pretrain_weights/darknet53_pretrained
 ```
 
 

examples/yolov3/main.py

@@ -28,6 +28,7 @@ from paddle.io import DataLoader
 from hapi.model import Model, Input, set_device
 from hapi.distributed import DistributedBatchSampler
 from hapi.datasets import COCODataset
+from hapi.download import get_weights_path
 from hapi.vision.transforms import *
 from hapi.vision.models import yolov3_darknet53, YoloLoss
 
@@ -124,7 +125,7 @@ def main():
                    model_mode='eval' if FLAGS.eval_only else 'train',
                    pretrained=pretrained)
 
-    if FLAGS.pretrain_weights is not None:
+    if FLAGS.pretrain_weights and not FLAGS.eval_only:
         model.load(FLAGS.pretrain_weights, skip_mismatch=True, reset_optimizer=True)
 
     optim = make_optimizer(len(batch_sampler), parameter_list=model.parameters())
@@ -196,7 +197,8 @@ if __name__ == '__main__':
     parser.add_argument(
         "-j", "--num_workers", default=4, type=int, help="reader worker number")
     parser.add_argument(
-        "-p", "--pretrain_weights", default=None, type=str,
+        "-p", "--pretrain_weights",
+        default="./pretrain_weights/darknet53_pretrained", type=str,
         help="path to pretrained weights")
     parser.add_argument(
         "-r", "--resume", default=None, type=str,

examples/yolov3/pretrain_weights/download.sh

+DIR="$( cd "$(dirname "$0")" ; pwd -P   )"
+cd "$DIR"
+
+echo "Downloading https://paddlemodels.bj.bcebos.com/hapi/darknet53_pretrained.pdparams"
+wget https://paddlemodels.bj.bcebos.com/hapi/darknet53_pretrained.pdparams

hapi/vision/models/darknet.py

@@ -136,7 +136,7 @@ class LayerWarp(fluid.dygraph.Layer):
 DarkNet_cfg = {53: ([1, 2, 8, 8, 4])}
 
 
-class DarkNet(Model):
+class DarkNet(fluid.dygraph.Layer):
     """DarkNet model from
     `"YOLOv3: An Incremental Improvement" <https://arxiv.org/abs/1804.02767>`_
 

hapi/vision/models/yolov3.py

@@ -118,7 +118,7 @@ class YOLOv3(Model):
         self.nms_posk = 100
         self.draw_thresh = 0.5
 
-        self.backbone = darknet53(pretrained=(model_mode=='train'))
+        self.backbone = darknet53(pretrained=False)
         self.block_outputs = []
         self.yolo_blocks = []
         self.route_blocks = []