Merge pull request #309 from Channingss/paddle_onnx

Paddle2ONNX support export ONNX opset9,10,11

README.md

@@ -61,6 +61,7 @@ x2paddle --framework=paddle2onnx --model=paddle_infer_model_dir --save_dir=onnx_
 |--without_data_format_optimization | **[可选]** For TensorFlow, 当指定该参数时，关闭NHWC->NCHW的优化，见[文档Q2](FAQ.md) |
 |--define_input_shape | **[可选]** For TensorFlow, 当指定该参数时，强制用户输入每个Placeholder的shape，见[文档Q2](FAQ.md) |
 |--params_merge | **[可选]** 当指定该参数时，转换完成后，inference_model中的所有模型参数将合并保存为一个文件__params__ |
+|--onnx_opset | **[可选]** 当framework为paddle2onnx时，该参数可设置转换为ONNX的OpSet版本，目前支持9、10、11，默认为10 |
 
 
 

x2paddle/convert.py

@@ -75,6 +75,12 @@ def arg_parser():
         action="store_true",
         default=False,
         help="define input shape for tf model")
+    parser.add_argument(
+        "--onnx_opset",
+        "-oo",
+        type=int,
+        default=10,
+        help="when paddle2onnx set onnx opset version to export")
     parser.add_argument(
         "--params_merge",
         "-pm",
@@ -172,7 +178,7 @@ def onnx2paddle(model_path, save_dir, params_merge=False):
         return
     print("Now translating model from onnx to paddle.")
 
-    from x2paddle.op_mapper.onnx_op_mapper import ONNXOpMapper
+    from x2paddle.op_mapper.onnx2paddle.onnx_op_mapper import ONNXOpMapper
     from x2paddle.decoder.onnx_decoder import ONNXDecoder
     from x2paddle.optimizer.onnx_optimizer import ONNXOptimizer
     model = ONNXDecoder(model_path)
@@ -186,12 +192,12 @@ def onnx2paddle(model_path, save_dir, params_merge=False):
     print("Paddle model and code generated.")
 
 
-def paddle2onnx(model_path, save_dir):
+def paddle2onnx(model_path, save_dir, opset_number):
     from x2paddle.decoder.paddle_decoder import PaddleDecoder
-    from x2paddle.op_mapper.paddle_op_mapper import PaddleOpMapper
+    from x2paddle.op_mapper.paddle2onnx.paddle_op_mapper import PaddleOpMapper
     model = PaddleDecoder(model_path, '__model__', '__params__')
     mapper = PaddleOpMapper()
-    mapper.convert(model.program, save_dir)
+    mapper.convert(model.program, save_dir, opset_number=opset_number)
 
 
 def main():
@@ -258,7 +264,7 @@ def main():
 
     elif args.framework == "paddle2onnx":
         assert args.model is not None, "--model should be defined while translating paddle model to onnx"
-        paddle2onnx(args.model, args.save_dir)
+        paddle2onnx(args.model, args.save_dir, args.onnx_opset)
 
     else:
         raise Exception(

x2paddle/op_mapper/onnx_op_mapper.py → x2paddle/op_mapper/onnx2paddle/onnx_op_mapper.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from x2paddle.op_mapper.onnx_opsets.opset9 import OpSet9
+from x2paddle.op_mapper.onnx2paddle.opsets.opset9 import OpSet9
 from x2paddle.core.op_mapper import OpMapper
 from x2paddle.op_mapper.onnx_opsets.custom_layer import *
 from x2paddle.decoder.onnx_decoder import ONNXGraph, ONNXGraphNode, ONNXGraphDataNode

x2paddle/op_mapper/onnx_opsets/custom_layer/__init__.py → x2paddle/op_mapper/onnx2paddle/opsets/custom_layer/__init__.py

@@ -13,10 +13,6 @@
 # limitations under the License.
 
 from .register import get_registered_layers
-#custom layer import begins
-
-from . import InstanceNormalization
-#custom layer import ends
 
 custom_layers = get_registered_layers()
 

x2paddle/op_mapper/paddle2onnx/opset10/opset.py

+#   Copyright (c) 2019  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import sys
+import x2paddle
+import os
+import numpy as np
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+import onnx
+from onnx import helper, onnx_pb
+from x2paddle.op_mapper.paddle2onnx.opset9.opset import OpSet9
+
+
+class OpSet10(OpSet9):
+    def __init__(self):
+        super(OpSet10, self).__init__()
+
+    def slice(self, op, block):
+        axes = op.attr('axes')
+        starts = op.attr('starts')
+        ends = op.attr('ends')
+        axes_name = self.get_name(op.type, 'axes')
+        starts_name = self.get_name(op.type, 'starts')
+        ends_name = self.get_name(op.type, 'ends')
+
+        axes_node = self.make_constant_node(axes_name,
+                                            onnx_pb.TensorProto.INT64, axes)
+        starts_node = self.make_constant_node(starts_name,
+                                              onnx_pb.TensorProto.INT64, starts)
+        ends_node = self.make_constant_node(ends_name,
+                                            onnx_pb.TensorProto.INT64, ends)
+        node = helper.make_node(
+            "Slice",
+            inputs=[op.input('Input')[0], starts_name, ends_name, axes_name],
+            outputs=op.output('Out'), )
+        return [starts_node, ends_node, axes_node, node]
+
+    def im2sequence(self, op, block):
+        from .paddle_custom_layer.im2sequence import im2sequence
+        return im2sequence(op, block)
+
+    def yolo_box(self, op, block):
+        from .paddle_custom_layer.yolo_box import yolo_box
+        return yolo_box(op, block)
+
+    def multiclass_nms(self, op, block):
+        from .paddle_custom_layer.multiclass_nms import multiclass_nms
+        return multiclass_nms(op, block)

x2paddle/op_mapper/paddle2onnx/opset10/paddle_custom_layer/im2sequence.py

+import onnx
+import numpy as np
+from onnx import onnx_pb, helper
+from x2paddle.op_mapper.paddle2onnx.opset9.paddle_custom_layer.im2sequence import im2sequence as im2sequence9
+
+
+def im2sequence(op, block):
+    return im2sequence9(op, block)

x2paddle/op_mapper/onnx_opsets/custom_layer/InstanceNormalization.py → x2paddle/op_mapper/paddle2onnx/opset10/paddle_custom_layer/multiclass_nms.py

@@ -12,45 +12,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .register import register
-
-
-def InstanceNormalization_shape(input_shape):
-    return input_shape
-
-
-def InstanceNormalization_layer(inputs, name=None):
-    # TODO(lvmengsi@baidu.com): Check the accuracy when using fluid.layers.layer_norm.
-    epsilon = 1e-5
-    input_ = inputs[0]
-    mean = fluid.layers.reduce_mean(input_, dim=[2, 3], keep_dim=True)
-    var = fluid.layers.reduce_mean(
-        fluid.layers.square(input_ - mean), dim=[2, 3], keep_dim=True)
-    if name is not None:
-        scale_name = name + "_scale"
-        offset_name = name + "_offset"
-
-    scale_param = inputs[1]
-    offset_param = inputs[2]
-    scale = fluid.layers.create_parameter(
-        name=scale_param.name, shape=input_.shape[1:2], dtype="float32")
-    offset = fluid.layers.create_parameter(
-        name=offset_param.name, shape=input_.shape[1:2], dtype="float32")
-
-    tmp = fluid.layers.elementwise_mul(x=(input_ - mean), y=scale, axis=1)
-    tmp = tmp / fluid.layers.sqrt(var + epsilon)
-    tmp = fluid.layers.elementwise_add(tmp, offset, axis=1)
-    return tmp
-
-
-def InstanceNormalization_weights(name, data=None):
-    weights_name = [name + '_scale']
-    return weights_name
-
-
-register(
-    kind='InstanceNormalization',
-    shape=InstanceNormalization_shape,
-    layer=InstanceNormalization_layer,
-    child_func=None,
-    weights=InstanceNormalization_weights)
+import math
+import sys
+import os
+import numpy as np
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+import onnx
+import warnings
+from onnx import helper, onnx_pb
+from x2paddle.op_mapper.paddle2onnx.opset9.paddle_custom_layer.multiclass_nms import multiclass_nms as multiclass_nms9
+
+
+def multiclass_nms(op, block):
+    """
+    Convert the paddle multiclass_nms to onnx op.
+    This op is get the select boxes from origin boxes.
+    """
+    return multiclass_nms9(op, block)

x2paddle/op_mapper/paddle2onnx/opset10/paddle_custom_layer/yolo_box.py

+import onnx
+import numpy as np
+from onnx import onnx_pb, helper
+from x2paddle.op_mapper.paddle2onnx.opset9.paddle_custom_layer.yolo_box import yolo_box as yolo_box9
+
+
+def yolo_box(op, block):
+    return yolo_box9(op, block)

x2paddle/op_mapper/paddle2onnx/opset11/opset.py

+#   Copyright (c) 2019  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import sys
+import x2paddle
+import os
+import numpy as np
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+import onnx
+from onnx import helper, onnx_pb
+from x2paddle.op_mapper.paddle2onnx.opset10.opset import OpSet10
+
+
+class OpSet11(OpSet10):
+    def __init__(self):
+        super(OpSet11, self).__init__()
+
+    def relu6(self, op, block):
+        min_name = self.get_name(op.type, 'min')
+        max_name = self.get_name(op.type, 'max')
+        min_node = self.make_constant_node(min_name, onnx_pb.TensorProto.FLOAT,
+                                           0)
+        max_node = self.make_constant_node(max_name, onnx_pb.TensorProto.FLOAT,
+                                           op.attr('threshold'))
+        node = helper.make_node(
+            'Clip',
+            inputs=[op.input('X')[0], min_name, max_name],
+            outputs=op.output('Out'), )
+        return [min_node, max_node, node]
+
+    def bilinear_interp(self, op, block):
+        input_names = op.input_names
+        coordinate_transformation_mode = ''
+        align_corners = op.attr('align_corners')
+        align_mode = op.attr('align_mode')
+        if align_corners:
+            coordinate_transformation_mode = 'align_corners'
+        elif align_mode == 1:
+            coordinate_transformation_mode = 'asymmetric'
+        else:
+            coordinate_transformation_mode = 'half_pixel'
+
+        if ('OutSize' in input_names and len(op.input('OutSize')) > 0) or (
+                'SizeTensor' in input_names and
+                len(op.input('SizeTensor')) > 0):
+            node_list = list()
+            roi_node = self.make_constant_node(
+                self.get_name(op.type, 'roi'), onnx_pb.TensorProto.FLOAT,
+                [1, 1, 1, 1, 1, 1, 1, 1])
+            roi_name = self.get_name(op.type, 'roi')
+            roi_node = self.make_constant_node(
+                roi_name, onnx_pb.TensorProto.FLOAT, [1, 1, 1, 1, 1, 1, 1, 1])
+            empty_name = self.get_name(op.type, 'empty')
+            empty_tensor = helper.make_tensor(
+                empty_name,
+                onnx_pb.TensorProto.FLOAT, (0, ),
+                np.array([]).astype('float32'),
+                raw=False)
+            empty_node = helper.make_node(
+                'Constant', [], outputs=[empty_name], value=empty_tensor)
+            shape_name0 = self.get_name(op.type, 'shape')
+            shape_node0 = helper.make_node(
+                'Shape', inputs=op.input('X'), outputs=[shape_name0])
+            starts_name = self.get_name(op.type, 'slice.starts')
+            starts_node = self.make_constant_node(
+                starts_name, onnx_pb.TensorProto.INT64, [0])
+            ends_name = self.get_name(op.type, 'slice.ends')
+            ends_node = self.make_constant_node(ends_name,
+                                                onnx_pb.TensorProto.INT64, [2])
+            shape_name1 = self.get_name(op.type, 'shape')
+            shape_node1 = helper.make_node(
+                'Slice',
+                inputs=[shape_name0, starts_name, ends_name],
+                outputs=[shape_name1])
+            node_list.extend([
+                roi_node, empty_node, shape_node0, starts_node, ends_node,
+                shape_node1
+            ])
+            if 'OutSize' in input_names and len(op.input('OutSize')) > 0:
+                cast_shape_name = self.get_name(op.type, "shape.cast")
+                cast_shape_node = helper.make_node(
+                    'Cast',
+                    inputs=op.input('OutSize'),
+                    outputs=[cast_shape_name],
+                    to=onnx_pb.TensorProto.INT64)
+                node_list.append(cast_shape_node)
+            else:
+                concat_shape_name = self.get_name(op.type, "shape.concat")
+                concat_shape_node = helper.make_node(
+                    "Concat",
+                    inputs=op.input('SizeTensor'),
+                    outputs=[concat_shape_name],
+                    axis=0)
+                cast_shape_name = self.get_name(op.type, "shape.cast")
+                cast_shape_node = helper.make_node(
+                    'Cast',
+                    inputs=[concat_shape_name],
+                    outputs=[cast_shape_name],
+                    to=onnx_pb.TensorProto.INT64)
+                node_list.extend([concat_shape_node, cast_shape_node])
+            shape_name3 = self.get_name(op.type, "shape.concat")
+            shape_node3 = helper.make_node(
+                'Concat',
+                inputs=[shape_name1, cast_shape_name],
+                outputs=[shape_name3],
+                axis=0)
+            result_node = helper.make_node(
+                'Resize',
+                inputs=[op.input('X')[0], roi_name, empty_name, shape_name3],
+                outputs=op.output('Out'),
+                mode='linear',
+                coordinate_transformation_mode=coordinate_transformation_mode)
+            node_list.extend([shape_node3, result_node])
+            return node_list
+        elif 'Scale' in input_names and len(op.input('Scale')) > 0:
+            node = helper.make_node(
+                'Resize',
+                inputs=[op.input('X')[0], op.input('Scale')[0]],
+                outputs=op.output('Out'),
+                mode='linear',
+                coordinate_transformation_mode=coordinate_transformation_mode)
+        else:
+            out_shape = [op.attr('out_h'), op.attr('out_w')]
+            scale = op.attr('scale')
+            if out_shape.count(-1) > 0:
+                scale_name = self.get_name(op.type, 'scale')
+                scale_node = self.make_constant_node(scale_name,
+                                                     onnx_pb.TensorProto.FLOAT,
+                                                     [1, 1, scale, scale])
+                roi_name = self.get_name(op.type, 'roi')
+                roi_node = self.make_constant_node(roi_name,
+                                                   onnx_pb.TensorProto.FLOAT,
+                                                   [1, 1, 1, 1, 1, 1, 1, 1])
+                node = helper.make_node(
+                    'Resize',
+                    inputs=[op.input('X')[0], roi_name, scale_name],
+                    outputs=op.output('Out'),
+                    mode='nearest',
+                    coordinate_transformation_mode=coordinate_transformation_mode
+                )
+                return [scale_node, roi_node, node]
+            else:
+                raise Exception("Unexpected situation happend")
+        return node
+
+    def nearest_interp(self, op, block):
+        input_names = op.input_names
+        coordinate_transformation_mode = ''
+        align_corners = op.attr('align_corners')
+        if align_corners:
+            coordinate_transformation_mode = 'align_corners'
+        else:
+            coordinate_transformation_mode = 'asymmetric'
+        if 'OutSize' in input_names and len(op.input('OutSize')) > 0:
+            node = helper.make_node(
+                'Resize',
+                inputs=[op.input('X')[0], '', op.input('OutSize')[0]],
+                outputs=op.output('Out'),
+                mode='nearest',
+                coordinate_transformation_mode=coordinate_transformation_mode)
+        elif 'Scale' in input_names and len(op.input('Scale')) > 0:
+            node = helper.make_node(
+                'Resize',
+                inputs=[op.input('X')[0], op.input('Scale')[0]],
+                outputs=op.output('Out'),
+                mode='nearest',
+                coordinate_transformation_mode=coordinate_transformation_mode)
+        else:
+            out_shape = [op.attr('out_h'), op.attr('out_w')]
+            scale = op.attr('scale')
+            if out_shape.count(-1) > 0:
+                scale_name = self.get_name(op.type, 'scale')
+                scale_node = self.make_constant_node(scale_name,
+                                                     onnx_pb.TensorProto.FLOAT,
+                                                     [1, 1, scale, scale])
+                roi_name = self.get_name(op.type, 'roi')
+                roi_node = self.make_constant_node(roi_name,
+                                                   onnx_pb.TensorProto.FLOAT,
+                                                   [1, 1, 1, 1, 1, 1, 1, 1])
+                node = helper.make_node(
+                    'Resize',
+                    inputs=[op.input('X')[0], roi_name, scale_name],
+                    outputs=op.output('Out'),
+                    mode='nearest',
+                    coordinate_transformation_mode=coordinate_transformation_mode
+                )
+                return [scale_node, roi_node, node]
+            else:
+                raise Exception("Unexpected situation happend")
+        return node
+
+    def hard_swish(self, op, block):
+        min_name = self.get_name(op.type, 'min')
+        max_name = self.get_name(op.type, 'max')
+        scale_name = self.get_name(op.type, 'scale')
+        offset_name = self.get_name(op.type, 'offset')
+        min_node = self.make_constant_node(min_name, onnx_pb.TensorProto.FLOAT,
+                                           0)
+        max_node = self.make_constant_node(max_name, onnx_pb.TensorProto.FLOAT,
+                                           op.attr('threshold'))
+        scale_node = self.make_constant_node(scale_name,
+                                             onnx_pb.TensorProto.FLOAT,
+                                             op.attr('scale'))
+        offset_node = self.make_constant_node(offset_name,
+                                              onnx_pb.TensorProto.FLOAT,
+                                              op.attr('offset'))
+
+        name0 = self.get_name(op.type, 'add')
+        node0 = helper.make_node(
+            'Add', inputs=[op.input('X')[0], offset_name], outputs=[name0])
+        name1 = self.get_name(op.type, 'relu')
+        node1 = helper.make_node(
+            'Clip',
+            inputs=[name0, min_name, max_name],
+            outputs=[name1], )
+        name2 = self.get_name(op.type, 'mul')
+        node2 = helper.make_node(
+            'Mul', inputs=[op.input('X')[0], name1], outputs=[name2])
+        node3 = helper.make_node(
+            'Div', inputs=[name2, scale_name], outputs=op.output('Out'))
+        return [
+            min_node, max_node, scale_node, offset_node, node0, node1, node2,
+            node3
+        ]
+
+    def im2sequence(self, op, block):
+        from .paddle_custom_layer.im2sequence import im2sequence
+        return im2sequence(op, block)
+
+    def yolo_box(self, op, block):
+        from .paddle_custom_layer.yolo_box import yolo_box
+        return yolo_box(op, block)
+
+    def multiclass_nms(self, op, block):
+        from .paddle_custom_layer.multiclass_nms import multiclass_nms
+        return multiclass_nms(op, block)

x2paddle/op_mapper/paddle2onnx/opset11/paddle_custom_layer/im2sequence.py

+import onnx
+import numpy as np
+from onnx import onnx_pb, helper
+from x2paddle.op_mapper.paddle2onnx.opset10.paddle_custom_layer.im2sequence import im2sequence as im2sequence10
+
+
+def im2sequence(op, block):
+    return im2sequence10(op, block)

x2paddle/op_mapper/paddle_custom_layer/multiclass_nms.py → x2paddle/op_mapper/paddle2onnx/opset11/paddle_custom_layer/multiclass_nms.py

@@ -125,7 +125,7 @@ def multiclass_nms(op, block):
                 vals=[value]))
         node_list.append(node)
 
-    # Ine this code block, we will deocde the raw score data, reshape N * C * M to 1 * N*C*M
+    # In this code block, we will deocde the raw score data, reshape N * C * M to 1 * N*C*M
     # and the same time, decode the select indices to 1 * D, gather the select_indices
     outputs_gather_1 = [result_name + "@gather_1"]
     node_gather_1 = onnx.helper.make_node(
@@ -405,12 +405,36 @@ def multiclass_nms(op, block):
 
     inputs_concat_final_results = outputs_cast_topk_class + outputs_unsqueeze_topk_scores +\
         outputs_gather_select_boxes
-    outputs_concat_final_results = outputs['Out']
-    node_concat_final_results = onnx.helper.make_node(
+    outputs_sort_by_socre_results = [result_name + "@concat_topk_scores"]
+    node_sort_by_socre_results = onnx.helper.make_node(
         'Concat',
         inputs=inputs_concat_final_results,
-        outputs=outputs_concat_final_results,
+        outputs=outputs_sort_by_socre_results,
         axis=2)
-    node_list.append(node_concat_final_results)
+    node_list.append(node_sort_by_socre_results)
 
+    # select topk classes indices
+    outputs_squeeze_cast_topk_class = [result_name + "@squeeze_cast_topk_class"]
+    node_squeeze_cast_topk_class = onnx.helper.make_node(
+        'Squeeze',
+        inputs=outputs_cast_topk_class,
+        outputs=outputs_squeeze_cast_topk_class,
+        axes=[0, 2])
+    node_list.append(node_squeeze_cast_topk_class)
+    outputs_topk_select_classes_indices = [result_name + "@topk_select_topk_classes_scores",\
+        result_name + "@topk_select_topk_classes_indices"]
+    node_topk_select_topk_indices = onnx.helper.make_node(
+        'TopK',
+        inputs=outputs_squeeze_cast_topk_class + outputs_cast_topk_indices,
+        outputs=outputs_topk_select_classes_indices,
+        largest=0)
+    node_list.append(node_topk_select_topk_indices)
+    outputs_concat_final_results = outputs['Out']
+    node_concat_final_results = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_sort_by_socre_results +
+        [outputs_topk_select_classes_indices[1]],
+        outputs=outputs_concat_final_results,
+        axis=1)
+    node_list.append(node_concat_final_results)
     return node_list

x2paddle/op_mapper/paddle_op_mapper.py → x2paddle/op_mapper/paddle2onnx/opset9/opset.py

@@ -23,7 +23,7 @@ import onnx
 from onnx import helper, onnx_pb
 
 
-class PaddleOpMapper(object):
+class OpSet9(object):
     def __init__(self):
         self.paddle_onnx_dtype_map = {
             core.VarDesc.VarType.FP32: onnx_pb.TensorProto.FLOAT,
@@ -34,62 +34,8 @@ class PaddleOpMapper(object):
             core.VarDesc.VarType.INT64: onnx_pb.TensorProto.INT64,
             core.VarDesc.VarType.BOOL: onnx_pb.TensorProto.BOOL
         }
-
         self.name_counter = dict()
 
-    def convert(self, program, save_dir):
-        weight_nodes = self.convert_weights(program)
-        op_nodes = list()
-        input_nodes = list()
-        output_nodes = list()
-        unsupported_ops = set()
-
-        print("Translating PaddlePaddle to ONNX...\n")
-        for block in program.blocks:
-            for i, op in enumerate(block.ops):
-                sys.stdout.write(
-                    "\rTotal:{}, Current:{} : {}                   ".format(
-                        len(block.ops), i + 1, op.type))
-                sys.stdout.flush()
-                if not hasattr(self, op.type):
-                    unsupported_ops.add(op.type)
-                    continue
-                if len(unsupported_ops) > 0:
-                    continue
-                node = getattr(self, op.type)(op, block)
-                if op.type == 'feed':
-                    input_nodes.append(node)
-                elif op.type == 'fetch':
-                    output_nodes.append(node)
-                else:
-                    if isinstance(node, list):
-                        op_nodes = op_nodes + node
-                    else:
-                        op_nodes.append(node)
-
-        if len(unsupported_ops) > 0:
-            print("\nThere's {} ops are not supported yet".format(
-                len(unsupported_ops)))
-            for op in unsupported_ops:
-                print("=========== {} ===========".format(op))
-            return
-
-        graph = helper.make_graph(
-            nodes=weight_nodes + op_nodes,
-            name='onnx_model_from_paddle',
-            initializer=[],
-            inputs=input_nodes,
-            outputs=output_nodes)
-        model = helper.make_model(graph, producer_name='X2Paddle')
-        onnx.checker.check_model(model)
-
-        if not os.path.isdir(save_dir):
-            os.makedirs(save_dir)
-        with open(os.path.join(save_dir, 'x2paddle_model.onnx'), 'wb') as f:
-            f.write(model.SerializeToString())
-        print("\nTranslated model saved in {}".format(
-            os.path.join(save_dir, 'x2paddle_model.onnx')))
-
     def get_name(self, op_name, var_name):
         name = 'p2o.{}.{}'.format(op_name, var_name)
         if name not in self.name_counter:
@@ -98,6 +44,21 @@ class PaddleOpMapper(object):
             self.name_counter[name] += 1
         return name + '.{}'.format(self.name_counter[name])
 
+    def make_constant_node(self, name, dtype, value=None):
+        if isinstance(value, list):
+            dims = (len(value), )
+        elif value is None:
+            dims = ()
+            value = []
+        else:
+            dims = ()
+            value = [value]
+        tensor = helper.make_tensor(
+            name=name, data_type=dtype, dims=dims, vals=value)
+        node = helper.make_node(
+            'Constant', inputs=[], outputs=[name], value=tensor)
+        return node
+
     def convert_weights(self, program):
         var_names = program.global_block().vars
         nodes = list()
@@ -118,21 +79,6 @@ class PaddleOpMapper(object):
             nodes.append(node)
         return nodes
 
-    def make_constant_node(self, name, dtype, value=None):
-        if isinstance(value, list):
-            dims = (len(value), )
-        elif value is None:
-            dims = ()
-            value = []
-        else:
-            dims = ()
-            value = [value]
-        tensor = helper.make_tensor(
-            name=name, data_type=dtype, dims=dims, vals=value)
-        node = helper.make_node(
-            'Constant', inputs=[], outputs=[name], value=tensor)
-        return node
-
     def conv2d(self, op, block):
         kernel_shape = block.var(op.input('Filter')[0]).shape
         node = helper.make_node(
@@ -182,41 +128,6 @@ class PaddleOpMapper(object):
             alpha=op.attr('alpha'))
         return node
 
-    def swish(self, op, block):
-        """
-        The activation swish,  y = x / (1 + exp(-beta * x))
-        """
-        beta = op.attr('beta')
-        beta_name = self.get_name(op.type, 'beta')
-        beta_node = onnx.helper.make_node(
-            'Constant',
-            name=beta_name,
-            inputs=[],
-            outputs=[beta_name],
-            value=onnx.helper.make_tensor(
-                name=beta_name,
-                data_type=onnx.TensorProto.FLOAT,
-                dims=(),
-                vals=[beta]))
-
-        beta_x_name = self.get_name(op.type, 'beta_x')
-        beta_x_node = onnx.helper.make_node(
-            'Mul',
-            name=beta_x_name,
-            inputs=[op.input('X')[0], beta_name],
-            outputs=[beta_x_name])
-        sigmoid_name = self.get_name(op.type, 'sigmoid')
-        sigmoid_node = onnx.helper.make_node(
-            'Sigmoid',
-            name=sigmoid_name,
-            inputs=[beta_x_name],
-            outputs=[sigmoid_name])
-        swish_node = onnx.helper.make_node(
-            'Mul',
-            inputs=[op.input('X')[0], sigmoid_name],
-            outputs=op.output('Out'))
-        return [beta_node, beta_x_node, sigmoid_node, swish_node]
-
     def elementwise_add(self, op, block):
         axis = op.attr('axis')
         x_shape = block.var(op.input('X')[0]).shape
@@ -285,6 +196,8 @@ class PaddleOpMapper(object):
                 pool_type[op.attr('pooling_type')][1],
                 inputs=op.input('X'),
                 outputs=op.output('Out'), )
+        elif op.attr('adaptive'):
+            raise Excpetion("ONNX cannot support adaptive pool")
         else:
             input_shape = block.var(op.input('X')[0]).shape
             k_size = op.attr('ksize')
@@ -431,17 +344,14 @@ class PaddleOpMapper(object):
         return self.conv2d(op, block)
 
     def relu6(self, op, block):
-        min_name = self.get_name(op.type, 'min')
-        max_name = self.get_name(op.type, 'max')
-        min_node = self.make_constant_node(min_name, onnx_pb.TensorProto.FLOAT,
-                                           0)
-        max_node = self.make_constant_node(max_name, onnx_pb.TensorProto.FLOAT,
-                                           op.attr('threshold'))
+        threshold = op.attr('threshold')
         node = helper.make_node(
             'Clip',
-            inputs=[op.input('X')[0], min_name, max_name],
-            outputs=op.output('Out'), )
-        return [min_node, max_node, node]
+            inputs=[op.input('X')[0]],
+            outputs=op.output('Out'),
+            max=threshold,
+            min=0.0)
+        return [node]
 
     def shape(self, op, block):
         node = helper.make_node(
@@ -469,21 +379,14 @@ class PaddleOpMapper(object):
         axes = op.attr('axes')
         starts = op.attr('starts')
         ends = op.attr('ends')
-        axes_name = self.get_name(op.type, 'axes')
-        starts_name = self.get_name(op.type, 'starts')
-        ends_name = self.get_name(op.type, 'ends')
-
-        axes_node = self.make_constant_node(axes_name,
-                                            onnx_pb.TensorProto.INT64, axes)
-        starts_node = self.make_constant_node(starts_name,
-                                              onnx_pb.TensorProto.INT64, starts)
-        ends_node = self.make_constant_node(ends_name,
-                                            onnx_pb.TensorProto.INT64, ends)
         node = helper.make_node(
             "Slice",
             inputs=[op.input('Input')[0], starts_name, ends_name, axes_name],
-            outputs=op.output('Out'), )
-        return [starts_node, ends_node, axes_node, node]
+            outputs=op.output('Out'),
+            axes=axes,
+            starts=starts,
+            ends=ends)
+        return [node]
 
     def fill_constant(self, op, block):
         value = op.attr('value')
@@ -578,27 +481,15 @@ class PaddleOpMapper(object):
 
     def bilinear_interp(self, op, block):
         input_names = op.input_names
-        coordinate_transformation_mode = 'half_pixel'
-        if op.attr('align_corners'):
-            coordinate_transformation_mode = 'align_corners'
+        input_shape = block.vars[op.input('X')[0]].shape
+        if op.attr('align_corners') or op.attr('align_mode') == 0:
+            raise Exception(
+                "Resize in onnx(opset<=10) only support coordinate_transformation_mode: 'asymmetric'."
+            )
         if ('OutSize' in input_names and len(op.input('OutSize')) > 0) or (
                 'SizeTensor' in input_names and
                 len(op.input('SizeTensor')) > 0):
             node_list = list()
-            roi_node = self.make_constant_node(
-                self.get_name(op.type, 'roi'), onnx_pb.TensorProto.FLOAT,
-                [1, 1, 1, 1, 1, 1, 1, 1])
-            roi_name = self.get_name(op.type, 'roi')
-            roi_node = self.make_constant_node(
-                roi_name, onnx_pb.TensorProto.FLOAT, [1, 1, 1, 1, 1, 1, 1, 1])
-            empty_name = self.get_name(op.type, 'empty')
-            empty_tensor = helper.make_tensor(
-                empty_name,
-                onnx_pb.TensorProto.FLOAT, (0, ),
-                np.array([]).astype('float32'),
-                raw=False)
-            empty_node = helper.make_node(
-                'Constant', [], outputs=[empty_name], value=empty_tensor)
             shape_name0 = self.get_name(op.type, 'shape')
             shape_node0 = helper.make_node(
                 'Shape', inputs=op.input('X'), outputs=[shape_name0])
@@ -613,16 +504,7 @@ class PaddleOpMapper(object):
                 'Slice',
                 inputs=[shape_name0, starts_name, ends_name],
                 outputs=[shape_name1])
-            node_list.extend([
-                roi_node, empty_node, shape_node0, starts_node, ends_node,
-                shape_node1
-            ])
-            #            shape_name2 = self.get_name(op.type, "shape.cast")
-            #            shape_node2 = helper.make_node(
-            #                'Cast',
-            #                inputs=op.input('OutSize'),
-            #                outputs=[shape_name2],
-            #                to=onnx_pb.TensorProto.INT64)
+            node_list.extend([shape_node0, starts_node, ends_node, shape_node1])
             if 'OutSize' in input_names and len(op.input('OutSize')) > 0:
                 cast_shape_name = self.get_name(op.type, "shape.cast")
                 cast_shape_node = helper.make_node(
@@ -632,7 +514,8 @@ class PaddleOpMapper(object):
                     to=onnx_pb.TensorProto.INT64)
                 node_list.append(cast_shape_node)
             else:
-                concat_shape_name = self.get_name(op.type, "shape.concat")
+                concat_shape_name = self.get_name(
+                    op.type, op.output('Out')[0] + "shape.concat")
                 concat_shape_node = helper.make_node(
                     "Concat",
                     inputs=op.input('SizeTensor'),
@@ -645,27 +528,46 @@ class PaddleOpMapper(object):
                     outputs=[cast_shape_name],
                     to=onnx_pb.TensorProto.INT64)
                 node_list.extend([concat_shape_node, cast_shape_node])
-            shape_name3 = self.get_name(op.type, "shape.concat")
-            shape_node3 = helper.make_node(
+            shape_name2 = self.get_name(op.type, "shape.concat")
+            shape_node2 = helper.make_node(
                 'Concat',
                 inputs=[shape_name1, cast_shape_name],
-                outputs=[shape_name3],
+                outputs=[shape_name2],
                 axis=0)
+            node_list.append(shape_node2)
+            cast_shape_name2 = self.get_name(op.type, "shape.cast")
+            cast_shape_node2 = helper.make_node(
+                'Cast',
+                inputs=[shape_name2],
+                outputs=[cast_shape_name2],
+                to=onnx_pb.TensorProto.FLOAT)
+            node_list.append(cast_shape_node2)
+            cast_shape_name0 = self.get_name(op.type, "shape.cast")
+            cast_shape_node0 = helper.make_node(
+                'Cast',
+                inputs=[shape_name0],
+                outputs=[cast_shape_name0],
+                to=onnx_pb.TensorProto.FLOAT)
+            node_list.append(cast_shape_node0)
+            outputs_h_w_scales = op.output('Out')[0] + "@out_hw_scales"
+            node_h_w_scales = helper.make_node(
+                'Div',
+                inputs=[cast_shape_name2, cast_shape_name0],
+                outputs=[outputs_h_w_scales])
+            node_list.append(node_h_w_scales)
             result_node = helper.make_node(
                 'Resize',
-                inputs=[op.input('X')[0], roi_name, empty_name, shape_name3],
+                inputs=[op.input('X')[0], outputs_h_w_scales],
                 outputs=op.output('Out'),
-                mode='linear',
-                coordinate_transformation_mode=coordinate_transformation_mode)
-            node_list.extend([shape_node3, result_node])
+                mode='linear')
+            node_list.extend([result_node])
             return node_list
         elif 'Scale' in input_names and len(op.input('Scale')) > 0:
             node = helper.make_node(
                 'Resize',
                 inputs=[op.input('X')[0], op.input('Scale')[0]],
                 outputs=op.output('Out'),
-                mode='linear',
-                coordinate_transformation_mode=coordinate_transformation_mode)
+                mode='linear')
         else:
             out_shape = [op.attr('out_h'), op.attr('out_w')]
             scale = op.attr('scale')
@@ -674,41 +576,34 @@ class PaddleOpMapper(object):
                 scale_node = self.make_constant_node(scale_name,
                                                      onnx_pb.TensorProto.FLOAT,
                                                      [1, 1, scale, scale])
-                roi_name = self.get_name(op.type, 'roi')
-                roi_node = self.make_constant_node(roi_name,
-                                                   onnx_pb.TensorProto.FLOAT,
-                                                   [1, 1, 1, 1, 1, 1, 1, 1])
                 node = helper.make_node(
                     'Resize',
-                    inputs=[op.input('X')[0], roi_name, scale_name],
+                    inputs=[op.input('X')[0], scale_name],
                     outputs=op.output('Out'),
-                    mode='nearest',
-                    coordinate_transformation_mode=coordinate_transformation_mode
-                )
-                return [scale_node, roi_node, node]
+                    mode='linear')
+                return [scale_node, node]
             else:
                 raise Exception("Unexpected situation happend")
         return node
 
     def nearest_interp(self, op, block):
         input_names = op.input_names
-        coordinate_transformation_mode = 'half_pixel'
         if op.attr('align_corners'):
-            coordinate_transformation_mode = 'align_corners'
+            raise Exception(
+                "Resize in onnx(opset<=10) only support coordinate_transformation_mode: 'asymmetric'."
+            )
         if 'OutSize' in input_names and len(op.input('OutSize')) > 0:
             node = helper.make_node(
                 'Resize',
-                inputs=[op.input('X')[0], '', op.input('OutSize')[0]],
+                inputs=[op.input('X')[0], op.input('OutSize')[0]],
                 outputs=op.output('Out'),
-                mode='nearest',
-                coordinate_transformation_mode=coordinate_transformation_mode)
+                mode='nearest')
         elif 'Scale' in input_names and len(op.input('Scale')) > 0:
             node = helper.make_node(
                 'Resize',
                 inputs=[op.input('X')[0], op.input('Scale')[0]],
                 outputs=op.output('Out'),
-                mode='nearest',
-                coordinate_transformation_mode=coordinate_transformation_mode)
+                mode='nearest')
         else:
             out_shape = [op.attr('out_h'), op.attr('out_w')]
             scale = op.attr('scale')
@@ -717,18 +612,12 @@ class PaddleOpMapper(object):
                 scale_node = self.make_constant_node(scale_name,
                                                      onnx_pb.TensorProto.FLOAT,
                                                      [1, 1, scale, scale])
-                roi_name = self.get_name(op.type, 'roi')
-                roi_node = self.make_constant_node(roi_name,
-                                                   onnx_pb.TensorProto.FLOAT,
-                                                   [1, 1, 1, 1, 1, 1, 1, 1])
                 node = helper.make_node(
                     'Resize',
-                    inputs=[op.input('X')[0], roi_name, scale_name],
+                    inputs=[op.input('X')[0], scale_name],
                     outputs=op.output('Out'),
-                    mode='nearest',
-                    coordinate_transformation_mode=coordinate_transformation_mode
-                )
-                return [scale_node, roi_node, node]
+                    mode='nearest')
+                return [scale_node, node]
             else:
                 raise Exception("Unexpected situation happend")
         return node
@@ -745,14 +634,8 @@ class PaddleOpMapper(object):
         return node
 
     def hard_swish(self, op, block):
-        min_name = self.get_name(op.type, 'min')
-        max_name = self.get_name(op.type, 'max')
         scale_name = self.get_name(op.type, 'scale')
         offset_name = self.get_name(op.type, 'offset')
-        min_node = self.make_constant_node(min_name, onnx_pb.TensorProto.FLOAT,
-                                           0)
-        max_node = self.make_constant_node(max_name, onnx_pb.TensorProto.FLOAT,
-                                           op.attr('threshold'))
         scale_node = self.make_constant_node(scale_name,
                                              onnx_pb.TensorProto.FLOAT,
                                              op.attr('scale'))
@@ -764,19 +647,20 @@ class PaddleOpMapper(object):
         node0 = helper.make_node(
             'Add', inputs=[op.input('X')[0], offset_name], outputs=[name0])
         name1 = self.get_name(op.type, 'relu')
+        min_value = op.attr('min')
+        max_value = op.attr('max')
         node1 = helper.make_node(
             'Clip',
-            inputs=[name0, min_name, max_name],
-            outputs=[name1], )
+            inputs=[name0],
+            outputs=[name1],
+            max=max_value,
+            min=min_value)
         name2 = self.get_name(op.type, 'mul')
         node2 = helper.make_node(
             'Mul', inputs=[op.input('X')[0], name1], outputs=[name2])
         node3 = helper.make_node(
             'Div', inputs=[name2, scale_name], outputs=op.output('Out'))
-        return [
-            min_node, max_node, scale_node, offset_node, node0, node1, node2,
-            node3
-        ]
+        return [scale_node, offset_node, node0, node1, node2, node3]
 
     def elementwise_mul(self, op, block):
         axis = op.attr('axis')
@@ -852,3 +736,11 @@ class PaddleOpMapper(object):
     def im2sequence(self, op, block):
         from .paddle_custom_layer.im2sequence import im2sequence
         return im2sequence(op, block)
+
+    def yolo_box(self, op, block):
+        from .paddle_custom_layer.yolo_box import yolo_box
+        return yolo_box(op, block)
+
+    def multiclass_nms(self, op, block):
+        from .paddle_custom_layer.multiclass_nms import multiclass_nms
+        return multiclass_nms(op, block)

x2paddle/op_mapper/paddle2onnx/opset9/paddle_custom_layer/multiclass_nms.py

+#   Copyright (c) 2019  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import sys
+import os
+import numpy as np
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+import onnx
+import warnings
+from onnx import helper, onnx_pb
+
+
+def multiclass_nms(op, block):
+    """
+    Convert the paddle multiclass_nms to onnx op.
+    This op is get the select boxes from origin boxes.
+    """
+    inputs = dict()
+    outputs = dict()
+    attrs = dict()
+    for name in op.input_names:
+        inputs[name] = op.input(name)
+    for name in op.output_names:
+        outputs[name] = op.output(name)
+    for name in op.attr_names:
+        attrs[name] = op.attr(name)
+
+    result_name = outputs['Out'][0]
+    background = attrs['background_label']
+    normalized = attrs['normalized']
+    if normalized == False:
+        warnings.warn(
+            'The parameter normalized of multiclass_nms OP of Paddle is False, which has diff with ONNX. \
+                         Please set normalized=True in multiclass_nms of Paddle')
+
+    #convert the paddle attribute to onnx tensor
+    name_score_threshold = [outputs['Out'][0] + "@score_threshold"]
+    name_iou_threshold = [outputs['Out'][0] + "@iou_threshold"]
+    name_keep_top_k = [outputs['Out'][0] + '@keep_top_k']
+    name_keep_top_k_2D = [outputs['Out'][0] + '@keep_top_k_1D']
+
+    node_score_threshold = onnx.helper.make_node(
+        'Constant',
+        inputs=[],
+        outputs=name_score_threshold,
+        value=onnx.helper.make_tensor(
+            name=name_score_threshold[0] + "@const",
+            data_type=onnx.TensorProto.FLOAT,
+            dims=(),
+            vals=[float(attrs['score_threshold'])]))
+
+    node_iou_threshold = onnx.helper.make_node(
+        'Constant',
+        inputs=[],
+        outputs=name_iou_threshold,
+        value=onnx.helper.make_tensor(
+            name=name_iou_threshold[0] + "@const",
+            data_type=onnx.TensorProto.FLOAT,
+            dims=(),
+            vals=[float(attrs['nms_threshold'])]))
+
+    node_keep_top_k = onnx.helper.make_node(
+        'Constant',
+        inputs=[],
+        outputs=name_keep_top_k,
+        value=onnx.helper.make_tensor(
+            name=name_keep_top_k[0] + "@const",
+            data_type=onnx.TensorProto.INT64,
+            dims=(),
+            vals=[np.int64(attrs['keep_top_k'])]))
+
+    node_keep_top_k_2D = onnx.helper.make_node(
+        'Constant',
+        inputs=[],
+        outputs=name_keep_top_k_2D,
+        value=onnx.helper.make_tensor(
+            name=name_keep_top_k_2D[0] + "@const",
+            data_type=onnx.TensorProto.INT64,
+            dims=[1, 1],
+            vals=[np.int64(attrs['keep_top_k'])]))
+
+    # the paddle data format is x1,y1,x2,y2
+    kwargs = {'center_point_box': 0}
+
+    name_select_nms = [outputs['Out'][0] + "@select_index"]
+    node_select_nms= onnx.helper.make_node(
+        'NonMaxSuppression',
+        inputs=inputs['BBoxes'] + inputs['Scores'] + name_keep_top_k +\
+            name_iou_threshold + name_score_threshold,
+        outputs=name_select_nms)
+    # step 1 nodes select the nms class
+    node_list = [
+        node_score_threshold, node_iou_threshold, node_keep_top_k,
+        node_keep_top_k_2D, node_select_nms
+    ]
+
+    # create some const value to use
+    name_const_value = [result_name+"@const_0",
+        result_name+"@const_1",\
+        result_name+"@const_2",\
+        result_name+"@const_-1"]
+    value_const_value = [0, 1, 2, -1]
+    for name, value in zip(name_const_value, value_const_value):
+        node = onnx.helper.make_node(
+            'Constant',
+            inputs=[],
+            outputs=[name],
+            value=onnx.helper.make_tensor(
+                name=name + "@const",
+                data_type=onnx.TensorProto.INT64,
+                dims=[1],
+                vals=[value]))
+        node_list.append(node)
+
+    # In this code block, we will deocde the raw score data, reshape N * C * M to 1 * N*C*M
+    # and the same time, decode the select indices to 1 * D, gather the select_indices
+    outputs_gather_1 = [result_name + "@gather_1"]
+    node_gather_1 = onnx.helper.make_node(
+        'Gather',
+        inputs=name_select_nms + [result_name + "@const_1"],
+        outputs=outputs_gather_1,
+        axis=1)
+    node_list.append(node_gather_1)
+
+    outputs_squeeze_gather_1 = [result_name + "@sequeeze_gather_1"]
+    node_squeeze_gather_1 = onnx.helper.make_node(
+        'Squeeze',
+        inputs=outputs_gather_1,
+        outputs=outputs_squeeze_gather_1,
+        axes=[1])
+    node_list.append(node_squeeze_gather_1)
+
+    outputs_gather_2 = [result_name + "@gather_2"]
+    node_gather_2 = onnx.helper.make_node(
+        'Gather',
+        inputs=name_select_nms + [result_name + "@const_2"],
+        outputs=outputs_gather_2,
+        axis=1)
+    node_list.append(node_gather_2)
+
+    #slice the class is not 0
+    if background == 0:
+        outputs_nonzero = [result_name + "@nonzero"]
+        node_nonzero = onnx.helper.make_node(
+            'NonZero', inputs=outputs_squeeze_gather_1, outputs=outputs_nonzero)
+        node_list.append(node_nonzero)
+    else:
+        name_thresh = [result_name + "@thresh"]
+        node_thresh = onnx.helper.make_node(
+            'Constant',
+            inputs=[],
+            outputs=name_thresh,
+            value=onnx.helper.make_tensor(
+                name=name_thresh[0] + "@const",
+                data_type=onnx.TensorProto.INT32,
+                dims=[1],
+                vals=[-1]))
+        node_list.append(node_thresh)
+
+        outputs_cast = [result_name + "@cast"]
+        node_cast = onnx.helper.make_node(
+            'Cast', inputs=outputs_squeeze_gather_1, outputs=outputs_cast, to=6)
+        node_list.append(node_cast)
+
+        outputs_greater = [result_name + "@greater"]
+        node_greater = onnx.helper.make_node(
+            'Greater',
+            inputs=outputs_cast + name_thresh,
+            outputs=outputs_greater)
+        node_list.append(node_greater)
+
+        outputs_nonzero = [result_name + "@nonzero"]
+        node_nonzero = onnx.helper.make_node(
+            'NonZero', inputs=outputs_greater, outputs=outputs_nonzero)
+        node_list.append(node_nonzero)
+
+    outputs_gather_1_nonzero = [result_name + "@gather_1_nonzero"]
+    node_gather_1_nonzero = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_gather_1 + outputs_nonzero,
+        outputs=outputs_gather_1_nonzero,
+        axis=0)
+    node_list.append(node_gather_1_nonzero)
+
+    outputs_gather_2_nonzero = [result_name + "@gather_2_nonzero"]
+    node_gather_2_nonzero = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_gather_2 + outputs_nonzero,
+        outputs=outputs_gather_2_nonzero,
+        axis=0)
+    node_list.append(node_gather_2_nonzero)
+
+    # reshape scores N * C * M to (N*C*M) * 1
+    outputs_reshape_scores_rank1 = [result_name + "@reshape_scores_rank1"]
+    node_reshape_scores_rank1 = onnx.helper.make_node(
+        "Reshape",
+        inputs=inputs['Scores'] + [result_name + "@const_-1"],
+        outputs=outputs_reshape_scores_rank1)
+    node_list.append(node_reshape_scores_rank1)
+
+    # get the shape of scores
+    outputs_shape_scores = [result_name + "@shape_scores"]
+    node_shape_scores = onnx.helper.make_node(
+        'Shape', inputs=inputs['Scores'], outputs=outputs_shape_scores)
+    node_list.append(node_shape_scores)
+
+    # gather the index: 2 shape of scores
+    outputs_gather_scores_dim1 = [result_name + "@gather_scores_dim1"]
+    node_gather_scores_dim1 = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_shape_scores + [result_name + "@const_2"],
+        outputs=outputs_gather_scores_dim1,
+        axis=0)
+    node_list.append(node_gather_scores_dim1)
+
+    # mul class * M
+    outputs_mul_classnum_boxnum = [result_name + "@mul_classnum_boxnum"]
+    node_mul_classnum_boxnum = onnx.helper.make_node(
+        'Mul',
+        inputs=outputs_gather_1_nonzero + outputs_gather_scores_dim1,
+        outputs=outputs_mul_classnum_boxnum)
+    node_list.append(node_mul_classnum_boxnum)
+
+    # add class * M * index
+    outputs_add_class_M_index = [result_name + "@add_class_M_index"]
+    node_add_class_M_index = onnx.helper.make_node(
+        'Add',
+        inputs=outputs_mul_classnum_boxnum + outputs_gather_2_nonzero,
+        outputs=outputs_add_class_M_index)
+    node_list.append(node_add_class_M_index)
+
+    # Squeeze the indices to 1 dim
+    outputs_squeeze_select_index = [result_name + "@squeeze_select_index"]
+    node_squeeze_select_index = onnx.helper.make_node(
+        'Squeeze',
+        inputs=outputs_add_class_M_index,
+        outputs=outputs_squeeze_select_index,
+        axes=[0, 2])
+    node_list.append(node_squeeze_select_index)
+
+    # gather the data from flatten scores
+    outputs_gather_select_scores = [result_name + "@gather_select_scores"]
+    node_gather_select_scores = onnx.helper.make_node('Gather',
+        inputs=outputs_reshape_scores_rank1 + \
+            outputs_squeeze_select_index,
+        outputs=outputs_gather_select_scores,
+        axis=0)
+    node_list.append(node_gather_select_scores)
+
+    # get nums to input TopK
+    outputs_shape_select_num = [result_name + "@shape_select_num"]
+    node_shape_select_num = onnx.helper.make_node(
+        'Shape',
+        inputs=outputs_gather_select_scores,
+        outputs=outputs_shape_select_num)
+    node_list.append(node_shape_select_num)
+
+    outputs_gather_select_num = [result_name + "@gather_select_num"]
+    node_gather_select_num = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_shape_select_num + [result_name + "@const_0"],
+        outputs=outputs_gather_select_num,
+        axis=0)
+    node_list.append(node_gather_select_num)
+
+    outputs_unsqueeze_select_num = [result_name + "@unsqueeze_select_num"]
+    node_unsqueeze_select_num = onnx.helper.make_node(
+        'Unsqueeze',
+        inputs=outputs_gather_select_num,
+        outputs=outputs_unsqueeze_select_num,
+        axes=[0])
+    node_list.append(node_unsqueeze_select_num)
+
+    outputs_concat_topK_select_num = [result_name + "@conat_topK_select_num"]
+    node_conat_topK_select_num = onnx.helper.make_node(
+        'Concat',
+        inputs=outputs_unsqueeze_select_num + name_keep_top_k_2D,
+        outputs=outputs_concat_topK_select_num,
+        axis=0)
+    node_list.append(node_conat_topK_select_num)
+
+    outputs_cast_concat_topK_select_num = [
+        result_name + "@concat_topK_select_num"
+    ]
+    node_outputs_cast_concat_topK_select_num = onnx.helper.make_node(
+        'Cast',
+        inputs=outputs_concat_topK_select_num,
+        outputs=outputs_cast_concat_topK_select_num,
+        to=6)
+    node_list.append(node_outputs_cast_concat_topK_select_num)
+    # get min(topK, num_select)
+    outputs_compare_topk_num_select = [result_name + "@compare_topk_num_select"]
+    node_compare_topk_num_select = onnx.helper.make_node(
+        'ReduceMin',
+        inputs=outputs_cast_concat_topK_select_num,
+        outputs=outputs_compare_topk_num_select,
+        keepdims=0)
+    node_list.append(node_compare_topk_num_select)
+
+    # unsqueeze the indices to 1D tensor
+    outputs_unsqueeze_topk_select_indices = [
+        result_name + "@unsqueeze_topk_select_indices"
+    ]
+    node_unsqueeze_topk_select_indices = onnx.helper.make_node(
+        'Unsqueeze',
+        inputs=outputs_compare_topk_num_select,
+        outputs=outputs_unsqueeze_topk_select_indices,
+        axes=[0])
+    node_list.append(node_unsqueeze_topk_select_indices)
+
+    # cast the indices to INT64
+    outputs_cast_topk_indices = [result_name + "@cast_topk_indices"]
+    node_cast_topk_indices = onnx.helper.make_node(
+        'Cast',
+        inputs=outputs_unsqueeze_topk_select_indices,
+        outputs=outputs_cast_topk_indices,
+        to=7)
+    node_list.append(node_cast_topk_indices)
+
+    # select topk scores  indices
+    outputs_topk_select_topk_indices = [result_name + "@topk_select_topk_values",\
+        result_name + "@topk_select_topk_indices"]
+    node_topk_select_topk_indices = onnx.helper.make_node(
+        'TopK',
+        inputs=outputs_gather_select_scores + outputs_cast_topk_indices,
+        outputs=outputs_topk_select_topk_indices)
+    node_list.append(node_topk_select_topk_indices)
+
+    # gather topk label, scores, boxes
+    outputs_gather_topk_scores = [result_name + "@gather_topk_scores"]
+    node_gather_topk_scores = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_gather_select_scores +
+        [outputs_topk_select_topk_indices[1]],
+        outputs=outputs_gather_topk_scores,
+        axis=0)
+    node_list.append(node_gather_topk_scores)
+
+    outputs_gather_topk_class = [result_name + "@gather_topk_class"]
+    node_gather_topk_class = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_gather_1_nonzero +
+        [outputs_topk_select_topk_indices[1]],
+        outputs=outputs_gather_topk_class,
+        axis=1)
+    node_list.append(node_gather_topk_class)
+
+    # gather the boxes need to gather the boxes id, then get boxes
+    outputs_gather_topk_boxes_id = [result_name + "@gather_topk_boxes_id"]
+    node_gather_topk_boxes_id = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_gather_2_nonzero +
+        [outputs_topk_select_topk_indices[1]],
+        outputs=outputs_gather_topk_boxes_id,
+        axis=1)
+    node_list.append(node_gather_topk_boxes_id)
+
+    # squeeze the gather_topk_boxes_id to 1 dim
+    outputs_squeeze_topk_boxes_id = [result_name + "@squeeze_topk_boxes_id"]
+    node_squeeze_topk_boxes_id = onnx.helper.make_node(
+        'Squeeze',
+        inputs=outputs_gather_topk_boxes_id,
+        outputs=outputs_squeeze_topk_boxes_id,
+        axes=[0, 2])
+    node_list.append(node_squeeze_topk_boxes_id)
+
+    outputs_gather_select_boxes = [result_name + "@gather_select_boxes"]
+    node_gather_select_boxes = onnx.helper.make_node(
+        'Gather',
+        inputs=inputs['BBoxes'] + outputs_squeeze_topk_boxes_id,
+        outputs=outputs_gather_select_boxes,
+        axis=1)
+    node_list.append(node_gather_select_boxes)
+
+    # concat the final result
+    # before concat need to cast the class to float
+    outputs_cast_topk_class = [result_name + "@cast_topk_class"]
+    node_cast_topk_class = onnx.helper.make_node(
+        'Cast',
+        inputs=outputs_gather_topk_class,
+        outputs=outputs_cast_topk_class,
+        to=1)
+    node_list.append(node_cast_topk_class)
+
+    outputs_unsqueeze_topk_scores = [result_name + "@unsqueeze_topk_scores"]
+    node_unsqueeze_topk_scores = onnx.helper.make_node(
+        'Unsqueeze',
+        inputs=outputs_gather_topk_scores,
+        outputs=outputs_unsqueeze_topk_scores,
+        axes=[0, 2])
+    node_list.append(node_unsqueeze_topk_scores)
+
+    inputs_concat_final_results = outputs_cast_topk_class + outputs_unsqueeze_topk_scores +\
+        outputs_gather_select_boxes
+    outputs_sort_by_socre_results = [result_name + "@concat_topk_scores"]
+    node_sort_by_socre_results = onnx.helper.make_node(
+        'Concat',
+        inputs=inputs_concat_final_results,
+        outputs=outputs_sort_by_socre_results,
+        axis=2)
+    node_list.append(node_sort_by_socre_results)
+
+    # select topk classes indices
+    outputs_squeeze_cast_topk_class = [result_name + "@squeeze_cast_topk_class"]
+    node_squeeze_cast_topk_class = onnx.helper.make_node(
+        'Squeeze',
+        inputs=outputs_cast_topk_class,
+        outputs=outputs_squeeze_cast_topk_class,
+        axes=[0, 2])
+    node_list.append(node_squeeze_cast_topk_class)
+    outputs_neg_squeeze_cast_topk_class = [
+        result_name + "@neg_squeeze_cast_topk_class"
+    ]
+    node_neg_squeeze_cast_topk_class = onnx.helper.make_node(
+        'Neg',
+        inputs=outputs_squeeze_cast_topk_class,
+        outputs=outputs_neg_squeeze_cast_topk_class)
+    node_list.append(node_neg_squeeze_cast_topk_class)
+    outputs_topk_select_classes_indices = [result_name + "@topk_select_topk_classes_scores",\
+        result_name + "@topk_select_topk_classes_indices"]
+    node_topk_select_topk_indices = onnx.helper.make_node(
+        'TopK',
+        inputs=outputs_neg_squeeze_cast_topk_class + outputs_cast_topk_indices,
+        outputs=outputs_topk_select_classes_indices)
+    node_list.append(node_topk_select_topk_indices)
+    outputs_concat_final_results = outputs['Out']
+    node_concat_final_results = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_sort_by_socre_results +
+        [outputs_topk_select_classes_indices[1]],
+        outputs=outputs_concat_final_results,
+        axis=1)
+    node_list.append(node_concat_final_results)
+    return node_list

x2paddle/op_mapper/paddle2onnx/paddle_op_mapper.py

+#   Copyright (c) 2019  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import sys
+import x2paddle
+import os
+import numpy as np
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+import onnx
+from onnx import helper, onnx_pb
+from x2paddle.op_mapper.paddle2onnx.opset9.opset import OpSet9
+from x2paddle.op_mapper.paddle2onnx.opset10.opset import OpSet10
+from x2paddle.op_mapper.paddle2onnx.opset11.opset import OpSet11
+
+
+class PaddleOpMapper(object):
+    def __init__(self):
+        self.support_opsets = [9, 10, 11]
+        self.default_opset = 10
+        self.name_counter = dict()
+        self.op_set = None
+
+    def convert(self, program, save_dir, opset_number=10):
+        self.op_set = self.create_opset(opset_number)
+        weight_nodes = self.op_set.convert_weights(program)
+        op_nodes = list()
+        input_nodes = list()
+        output_nodes = list()
+        unsupported_ops = set()
+
+        print("Translating PaddlePaddle to ONNX...\n")
+        for block in program.blocks:
+            for i, op in enumerate(block.ops):
+                sys.stdout.write("\rTotal:{}, Current:{} : {} ".format(
+                    len(block.ops), i + 1, op.type))
+                sys.stdout.flush()
+                if not hasattr(self.op_set, op.type):
+                    unsupported_ops.add(op.type)
+                    continue
+                if len(unsupported_ops) > 0:
+                    continue
+                node = getattr(self.op_set, op.type)(op, block)
+                if op.type == 'feed':
+                    print(node.name)
+                    input_nodes.append(node)
+                elif op.type == 'fetch':
+                    output_nodes.append(node)
+                else:
+                    if isinstance(node, list):
+                        op_nodes = op_nodes + node
+                    else:
+                        op_nodes.append(node)
+
+        if len(unsupported_ops) > 0:
+            print("\nThere's {} ops are not supported yet".format(
+                len(unsupported_ops)))
+            for op in unsupported_ops:
+                print("=========== {} ===========".format(op))
+            return
+
+        graph = helper.make_graph(
+            nodes=weight_nodes + op_nodes,
+            name='onnx_model_from_paddle',
+            initializer=[],
+            inputs=input_nodes,
+            outputs=output_nodes)
+        opset_imports = [helper.make_opsetid("", opset_number)]
+        model = helper.make_model(
+            graph, producer_name='X2Paddle', opset_imports=opset_imports)
+        onnx.checker.check_model(model)
+
+        if not os.path.isdir(save_dir):
+            os.makedirs(save_dir)
+        with open(os.path.join(save_dir, 'x2paddle_model.onnx'), 'wb') as f:
+            f.write(model.SerializeToString())
+        print("\nTranslated model saved in {}".format(
+            os.path.join(save_dir, 'x2paddle_model.onnx')))
+
+    def create_opset(self, opset_number):
+        run_opset = self.default_opset
+        opset = ''
+        if opset_number in self.support_opsets:
+            run_opset = opset_number
+        else:
+            for support_opset_number in self.support_opsets:
+                if support_opset_number < opset_number:
+                    run_opset = support_opset_number
+                else:
+                    break
+        print(
+            'Now, onnx2paddle support convert onnx model opset_verison {},'
+            'opset_verison of your onnx model is {}, automatically treated as op_set: {}.'
+            .format(self.support_opsets, opset_number, run_opset))
+        opset = 'OpSet' + str(run_opset)
+        return eval(opset)()