Add yolov3_mobilenet_v3 (#443)

* add yolov3_mobilenet_v3

configs/yolov3_darknet.yml

 architecture: YOLOv3
 use_gpu: true
-max_iters: 500200
+max_iters: 500000
 log_smooth_window: 20
 save_dir: output
-snapshot_iter: 2000
+snapshot_iter: 10000
 metric: COCO
 pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/DarkNet53_pretrained.tar
 weights: output/yolov3_darknet/model_final

configs/yolov3_mobilenet_v1.yml

 architecture: YOLOv3
 use_gpu: true
-max_iters: 500200
+max_iters: 500000
 log_smooth_window: 20
 save_dir: output
-snapshot_iter: 2000
+snapshot_iter: 10000
 metric: COCO
 pretrain_weights: http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar
 weights: output/yolov3_mobilenet_v1/model_final

configs/yolov3_mobilenet_v3.yml

+architecture: YOLOv3
+use_gpu: true
+max_iters: 500000
+log_smooth_window: 20
+save_dir: output
+snapshot_iter: 10000
+metric: COCO
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV3_large_x1_0_pretrained.tar
+weights: output/yolov3_mobilenet_v3/model_final
+num_classes: 80
+use_fine_grained_loss: false
+
+YOLOv3:
+  backbone: MobileNetV3
+  yolo_head: YOLOv3Head
+
+MobileNetV3:
+  norm_type: sync_bn
+  norm_decay: 0.
+  model_name: large
+  scale: 1.
+  with_extra_blocks: false
+
+YOLOv3Head:
+  anchor_masks: [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
+  anchors: [[10, 13], [16, 30], [33, 23],
+            [30, 61], [62, 45], [59, 119],
+            [116, 90], [156, 198], [373, 326]]
+  norm_decay: 0.
+  yolo_loss: YOLOv3Loss
+  nms:
+    background_label: -1
+    keep_top_k: 100
+    nms_threshold: 0.45
+    nms_top_k: 1000
+    normalized: false
+    score_threshold: 0.01
+
+YOLOv3Loss:
+  # batch_size here is only used for fine grained loss, not used
+  # for training batch_size setting, training batch_size setting
+  # is in configs/yolov3_reader.yml TrainReader.batch_size, batch
+  # size here should be set as same value as TrainReader.batch_size
+  batch_size: 8
+  ignore_thresh: 0.7
+  label_smooth: false
+
+LearningRate:
+  base_lr: 0.001
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones:
+    - 400000
+    - 450000
+  - !LinearWarmup
+    start_factor: 0.
+    steps: 4000
+
+OptimizerBuilder:
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0005
+    type: L2
+
+_READER_: 'yolov3_reader.yml'

configs/yolov3_r34.yml

 architecture: YOLOv3
 use_gpu: true
-max_iters: 500200
+max_iters: 500000
 log_smooth_window: 20
 save_dir: output
-snapshot_iter: 2000
+snapshot_iter: 10000
 metric: COCO
 pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet34_pretrained.tar
 weights: output/yolov3_r34/model_final

docs/MODEL_ZOO.md

@@ -133,6 +133,9 @@ The backbone models pretrained on ImageNet are available. All backbone models ar
 | MobileNet-V1      | ImageNet | 608  |  False    |    8    |   270e  |    78.302     |  29.3  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
 | MobileNet-V1      | ImageNet | 416  |  False    |    8    |   270e  |      -        |  29.3  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
 | MobileNet-V1      | ImageNet | 320  |  False    |    8    |   270e  |      -        |  27.1  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
+| MobileNet-V3      | ImageNet | 608  |  False    |    8    |   270e  |      -        |  31.6  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v3.pdparams) |
+| MobileNet-V3      | ImageNet | 416  |  False    |    8    |   270e  |      -        |  29.9  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v3.pdparams) |
+| MobileNet-V3      | ImageNet | 320  |  False    |    8    |   270e  |      -        |  27.1  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v3.pdparams) |
 | ResNet34          | ImageNet | 608  |  False    |    8    |   270e  |    63.356     |  36.2  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) |
 | ResNet34          | ImageNet | 416  |  False    |    8    |   270e  |      -        |  34.3  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) |
 | ResNet34          | ImageNet | 320  |  False    |    8    |   270e  |      -        |  31.4  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) |

docs/MODEL_ZOO_cn.md

@@ -129,6 +129,9 @@ Paddle提供基于ImageNet的骨架网络预训练模型。所有预训练模型
 | MobileNet-V1        | ImageNet | 608  |  否    |    8    |   270e  |    78.302     |  29.3  | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
 | MobileNet-V1        | ImageNet | 416  |  否    |    8    |   270e  |      -        |  29.3  | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
 | MobileNet-V1        | ImageNet | 320  |  否    |    8    |   270e  |      -        |  27.1  | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
+| MobileNet-V1        | ImageNet | 608  |  否    |    8    |   270e  |      -        |  31.6  | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v3.pdparams) |
+| MobileNet-V1        | ImageNet | 416  |  否    |    8    |   270e  |      -        |  29.9  | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v3.pdparams) |
+| MobileNet-V1        | ImageNet | 320  |  否    |    8    |   270e  |      -        |  27.1  | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v3.pdparams) |
 | ResNet34            | ImageNet | 608  |  否    |    8    |   270e  |    63.356     |  36.2  | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) |
 | ResNet34            | ImageNet | 416  |  否    |    8    |   270e  |      -        |  34.3  | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) |
 | ResNet34            | ImageNet | 320  |  否    |    8    |   270e  |      -        |  31.4  | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) |

ppdet/modeling/backbones/mobilenet_v3.py

@@ -15,7 +15,8 @@ class MobileNetV3():
                  model_name='small',
                  with_extra_blocks=False,
                  conv_decay=0.0,
-                 bn_decay=0.0,
+                 norm_type='bn',
+                 norm_decay=0.0,
                  extra_block_filters=[[256, 512], [128, 256], [128, 256],
                                       [64, 128]]):
         self.scale = scale
@@ -23,7 +24,7 @@ class MobileNetV3():
         self.with_extra_blocks = with_extra_blocks
         self.extra_block_filters = extra_block_filters
         self.conv_decay = conv_decay
-        self.bn_decay = bn_decay
+        self.norm_decay = norm_decay
         self.inplanes = 16
         self.end_points = []
         self.block_stride = 1
@@ -90,9 +91,9 @@ class MobileNetV3():
             bias_attr=False)
         bn_name = name + '_bn'
         bn_param_attr = ParamAttr(
-            name=bn_name + "_scale", regularizer=L2Decay(self.bn_decay))
+            name=bn_name + "_scale", regularizer=L2Decay(self.norm_decay))
         bn_bias_attr = ParamAttr(
-            name=bn_name + "_offset", regularizer=L2Decay(self.bn_decay))
+            name=bn_name + "_offset", regularizer=L2Decay(self.norm_decay))
         bn = fluid.layers.batch_norm(
             input=conv,
             param_attr=bn_param_attr,
@@ -238,6 +239,8 @@ class MobileNetV3():
         i = 0
         for layer_cfg in cfg:
             self.block_stride *= layer_cfg[5]
+            if layer_cfg[5] == 2:
+                blocks.append(conv)
             conv = self._residual_unit(
                 input=conv,
                 num_in_filter=inplanes,
@@ -250,9 +253,10 @@ class MobileNetV3():
                 name='conv' + str(i + 2))
             inplanes = int(scale * layer_cfg[2])
             i += 1
+        blocks.append(conv)
 
         if not self.with_extra_blocks:
-            return conv
+            return blocks
 
         # extra block
         conv_extra = self._conv_bn_layer(

ppdet/utils/checkpoint.py

@@ -92,6 +92,13 @@ def _load_state(path):
     return state
 
 
+def _strip_postfix(path):
+    path, ext = os.path.splitext(path)
+    assert ext in ['', '.pdparams', '.pdopt', '.pdmodel'], \
+            "Unknown postfix {} from weights".format(ext)
+    return path
+
+
 def load_params(exe, prog, path, ignore_params=[]):
     """
     Load model from the given path.
@@ -106,6 +113,8 @@ def load_params(exe, prog, path, ignore_params=[]):
 
     if is_url(path):
         path = _get_weight_path(path)
+
+    path = _strip_postfix(path)
     if not (os.path.isdir(path) or os.path.exists(path + '.pdparams')):
         raise ValueError("Model pretrain path {} does not "
                          "exists.".format(path))
@@ -151,6 +160,8 @@ def load_checkpoint(exe, prog, path):
     """
     if is_url(path):
         path = _get_weight_path(path)
+
+    path = _strip_postfix(path)
     if not (os.path.isdir(path) or os.path.exists(path + '.pdparams')):
         raise ValueError("Model pretrain path {} does not "
                          "exists.".format(path))

ppdet/utils/download.py

@@ -201,6 +201,10 @@ def get_path(url, root_dir, md5sum=None, check_exist=True):
     else:
         exist_flag = False
         fullname = _download(url, root_dir, md5sum)
+
+        # new weights format which postfix is 'pdparams' not
+        # need to decompress
+        if osp.splitext(fullname)[-1] != '.pdparams':
             _decompress(fullname)
 
     return fullpath, exist_flag