Merge pull request #2096 from heavengate/yolo_syncbn

sync bn for yolov3.

PaddleCV/yolov3/README.md

@@ -134,14 +134,22 @@ Evaluation is to evaluate the performance of a trained model. This sample provid
 
 - Set ```export CUDA_VISIBLE_DEVICES=0``` to specifiy one GPU to eval.
 
-Evalutaion result is shown as below:
+If train with '--syncbn=False', Evalutaion result is shown as below:
 
 |   input size  | mAP(IoU=0.50:0.95) | mAP(IoU=0.50) | mAP(IoU=0.75) |
 | :------: | :------: | :------: | :------: |
-| 608x608| 37.7 | 59.8 | 40.8 |
+| 608x608 | 37.7 | 59.8 | 40.8 |
 | 416x416 | 36.5 | 58.2 | 39.1 |
 | 320x320 | 34.1 | 55.4 | 36.3 |
 
+If train with '--syncbn=True', Evalutaion result is shown as below:
+
+|   input size  | mAP(IoU=0.50:0.95) | mAP(IoU=0.50) | mAP(IoU=0.75) |
+| :------: | :------: | :------: | :------: |
+| 608x608 | 38.9 | 61.1 | 42.0 |
+| 416x416 | 37.5 | 59.6 | 40.2 |
+| 320x320 | 34.8 | 56.4 | 36.9 |
+
 - **NOTE:** evaluations based on `pycocotools` evaluator, predict bounding boxes with `score < 0.05` were not filtered out. Some frameworks which filtered out predict bounding boxes with `score < 0.05` will cause a drop in accuracy.
 
 ## Inference and Visualization

PaddleCV/yolov3/README_cn.md

@@ -134,15 +134,23 @@ Train Loss
 
 - 通过设置export CUDA\_VISIBLE\_DEVICES=0指定单卡GPU评估。
 
-模型评估结果：
+若训练时指定'--syncbn=False', 模型评估精度如下:
 
 |   input size  | mAP(IoU=0.50:0.95) | mAP(IoU=0.50) | mAP(IoU=0.75) |
 | :------: | :------: | :------: | :------: |
-| 608x608| 37.7 | 59.8 | 40.8 |
+| 608x608 | 37.7 | 59.8 | 40.8 |
 | 416x416 | 36.5 | 58.2 | 39.1 |
 | 320x320 | 34.1 | 55.4 | 36.3 |
 
-- **注意: **评估结果基于`pycocotools`评估器，没有滤除`score < 0.05`的预测框，其他框架有此滤除操作会导致精度下降。
+若训练时指定'--syncbn=True', 模型评估精度如下:
+
+|   input size  | mAP(IoU=0.50:0.95) | mAP(IoU=0.50) | mAP(IoU=0.75) |
+| :------: | :------: | :------: | :------: |
+| 608x608 | 38.9 | 61.1 | 42.0 |
+| 416x416 | 37.5 | 59.6 | 40.2 |
+| 320x320 | 34.8 | 56.4 | 36.9 |
+
+- **注意：** 评估结果基于`pycocotools`评估器，没有滤除`score < 0.05`的预测框，其他框架有此滤除操作会导致精度下降。
 
 
 

PaddleCV/yolov3/models/yolov3.py

@@ -46,19 +46,9 @@ def yolo_detection_block(input, channel, is_test=True, name=None):
     return route, tip
 
 def upsample(input, scale=2,name=None):
-    # get dynamic upsample output shape
-    shape_nchw = fluid.layers.shape(input)
-    shape_hw = fluid.layers.slice(shape_nchw, axes=[0], starts=[2], ends=[4])
-    shape_hw.stop_gradient = True
-    in_shape = fluid.layers.cast(shape_hw, dtype='int32')
-    out_shape = in_shape * scale
-    out_shape.stop_gradient = True
-
-    # reisze by actual_shape
     out = fluid.layers.resize_nearest(
         input=input,
-        scale=scale,
-        actual_shape=out_shape,
+        scale=float(scale),
         name=name)
     return out
 

PaddleCV/yolov3/train.py

@@ -82,6 +82,7 @@ def train():
 
     build_strategy= fluid.BuildStrategy()
     build_strategy.memory_optimize = True
+    build_strategy.sync_batch_norm = cfg.syncbn 
     compile_program = fluid.compiler.CompiledProgram(
             fluid.default_main_program()).with_data_parallel(
             loss_name=loss.name, build_strategy=build_strategy)

PaddleCV/yolov3/utility.py

@@ -112,6 +112,7 @@ def parse_args():
     add_arg('no_mixup_iter',    int,    40000,  "Disable mixup in last N iter.")
     # TRAIN TEST INFER
     add_arg('input_size',       int,    608,    "Image input size of YOLOv3.")
+    add_arg('syncbn',           bool,   True,   "Whether to use synchronized batch normalization.")
     add_arg('random_shape',     bool,   True,   "Resize to random shape for train reader.")
     add_arg('valid_thresh',     float,  0.005,  "Valid confidence score for NMS.")
     add_arg('nms_thresh',       float,  0.45,   "NMS threshold.")