Merge branch 'develop' into fix_adative_pool

Author: channings

docs/en/model_zoo.md

@@ -42,7 +42,8 @@ Due to ONNX's limit, Paddle2ONNX is not able to support all the detection models
 |YoloV3_ResNet34|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO.md#yolo-v3) |
 |YoloV3_MobileNet|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO.md#yolo-v3) |
 |FasterRCNN|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO.md#faster--mask-r-cnn) |
-|FasterRCNN_with_DCN|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO.md#deformable-convnets-v2) |
+|FasterRCNN_FPN|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO.md#faster--mask-r-cnn) |
+|FasterRCNN_FPN_DCN|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO.md#deformable-convnets-v2) |
 
 ## Image segmentation
 Support UNet,HRNet and DeepLab. Test models are from [PaddleSeg](https://github.com/PaddlePaddle/PaddleSeg).

docs/zh/model_zoo.md

@@ -38,7 +38,8 @@
 |YoloV3_ResNet34|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO_cn.md#yolo-v3-%E5%9F%BA%E4%BA%8Epasacl-voc%E6%95%B0%E6%8D%AE%E9%9B%86) |
 |YoloV3_MobileNet|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO_cn.md#yolo-v3-%E5%9F%BA%E4%BA%8Epasacl-voc%E6%95%B0%E6%8D%AE%E9%9B%86) |
 |FasterRCNN|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO_cn.md#faster--mask-r-cnn) |
-|FasterRCNN_with_DCN|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO_cn.md#deformable-%E5%8D%B7%E7%A7%AF%E7%BD%91%E7%BB%9Cv2) |
+|FasterRCNN_FPN|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO_cn.md#faster--mask-r-cnn) |
+|FasterRCNN_FPN_DCN|[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.4/docs/MODEL_ZOO_cn.md#deformable-%E5%8D%B7%E7%A7%AF%E7%BD%91%E7%BB%9Cv2) |
 
 ## 图像分割
 支持的模型有UNet, HRNet, DeepLab模型，测试模型来自于[PaddleSeg](https://github.com/PaddlePaddle/PaddleSeg)。

paddle2onnx/op_mapper/math.py

@@ -152,6 +152,21 @@ def opset_1(cls, graph, node, **kw):
             'MatMul', inputs=[flatten_x, flatten_y], outputs=node.output('Out'))
 
 
+@op_mapper('affine_channel')
+class AffineChannel():
+    support_opset_version_range = (1, 12)
+
+    @classmethod
+    def opset_1(cls, graph, node, **kw):
+        x = node.input('X', 0)
+        bias = node.input('Bias', 0)
+        scale = node.input('Scale', 0)
+        scale = graph.make_node('Unsqueeze', inputs=scale, axes=[0, 2, 3])
+        bias = graph.make_node('Unsqueeze', inputs=bias, axes=[0, 2, 3])
+        x = graph.make_node('Mul', inputs=[x, scale])
+        x = graph.make_node('Add', inputs=[x, bias], outputs=node.output('Out'))
+
+
 @op_mapper('bmm')
 class BMM():
     support_opset_version_range = (1, 12)