[onnx]support paddle1.8, remove onnxruntime

.pre-commit-config.yaml

--   repo: local
+-   repo: https://github.com/PaddlePaddle/mirrors-yapf.git
+    sha: 0d79c0c469bab64f7229c9aca2b1186ef47f0e37
     hooks:
     -   id: yapf
-        name: yapf
-        entry: yapf
-        language: system
-        args: [-i, --style .style.yapf]
         files: \.py$
-
 -   repo: https://github.com/pre-commit/pre-commit-hooks
     sha: a11d9314b22d8f8c7556443875b731ef05965464
     hooks:
@@ -18,6 +14,7 @@
     -   id: check-symlinks
     -   id: check-added-large-files
 -   repo: local
+
     hooks:
     -   id: copyright_checker
         name: copyright_checker

.travis.yml

 language: python
 
 python:
-  - '2.7'
   - '3.5'
-  - '3.6'
 
 script:
   - if [[ $TRAVIS_PYTHON_VERSION != 2.7 ]]; then /bin/bash ./tools/check_code_style.sh; fi

FAQ.md

@@ -11,3 +11,6 @@ x2paddle -f tensorflow -m tf.pb -s pd-model --without_data_format_optimization -
 ```
 
 > 1. 目前Tensorflow的CV模型大部分均为`NHWC`的输入格式，而Paddle的默认输入格式为`NCHW`，因此X2Paddle在转换过程中，会对如`axis`， `shape`等参数进行转换，适应Paddle的NCHW格式。但在这种情况下，可能会由于TensorFlow模型太复杂，导致出错。  指定`--without_data_format_optimization`后，会停止对`axis`，`shape`等参数的优化（这可能会带来一定数量的transpose操作）
+
+**Q3. ONNX模型转换过程中，提示『Unknown shape for input tensor[tensor name: "input"] -> shape: ['batch', 'sequence']， Please define shape of input here』**  
+A：该提示信息表示从ONNX的模型中获取到输入tensor(tensor名为"input:)的shape是语义象征性的['batch', 'sequence']，而不是dim为int类型的shape，从而可能会因为部分node的shape无法推理，导致转换失败。所以用户可以尝试手动在提示后输入详细的shape信息，如:-1,3,224,224  其中-1表示Batch

README.md

@@ -10,12 +10,12 @@ X2Paddle在多个主流的CV模型上，测试过TensorFlow/Caffe/ONNX模型的
 ## 环境依赖
 
 python == 2.7 | python >= 3.5  
-paddlepaddle >= 1.6.0  
+paddlepaddle >= 1.8.0  
 
 **按需安装以下依赖**  
 tensorflow ： tensorflow == 1.14.0  
 caffe ： 无  
-onnx ： onnx == 1.6.0  onnxruntime == 1.0.0
+onnx ： onnx == 1.6.0
 
 ## 安装
 ### 安装方式一（推荐）
@@ -44,10 +44,15 @@ x2paddle --framework=caffe --prototxt=deploy.prototxt --weight=deploy.caffemodel
 ```
 x2paddle --framework=onnx --model=onnx_model.onnx --save_dir=pd_model
 ```
+### Paddle2ONNX
+```
+# 注意：paddle_infer_model_dir下需包含__model__和__params__两个文件
+x2paddle --framework=paddle2onnx --model=paddle_infer_model_dir --save_dir=onnx_model
+```
 ### 参数选项
 | 参数 | |
 |----------|--------------|
-|--framework | 源模型类型 (tensorflow、caffe、onnx) |
+|--framework | 源模型类型 (tensorflow、caffe、onnx、paddle2onnx) |
 |--prototxt | 当framework为caffe时，该参数指定caffe模型的proto文件路径 |
 |--weight | 当framework为caffe时，该参数指定caffe模型的参数文件路径 |
 |--save_dir | 指定转换后的模型保存目录路径 |
@@ -58,6 +63,7 @@ x2paddle --framework=onnx --model=onnx_model.onnx --save_dir=pd_model
 |--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__ |
 
 
+
 ## 使用转换后的模型
 转换后的模型包括`model_with_code`和`inference_model`两个目录。  
 `model_with_code`中保存了模型参数，和转换后的python模型代码  

setup.py

@@ -11,8 +11,7 @@ setuptools.setup(
     version=x2paddle.__version__,
     author="dltp-sz",
     author_email="dltp-sz@baidu.com",
-    description=
-    "a toolkit for converting trained model to PaddlePaddle from other deep learning frameworks.",
+    description="a toolkit for converting trained model to PaddlePaddle from other deep learning frameworks.",
     long_description=long_description,
     long_description_content_type="text/plain",
     url="https://github.com/PaddlePaddle/x2paddle",
@@ -23,6 +22,4 @@ setuptools.setup(
         "Operating System :: OS Independent",
     ],
     license='Apache 2.0',
-    entry_points={'console_scripts': [
-        'x2paddle=x2paddle.convert:main',
-    ]})
+    entry_points={'console_scripts': ['x2paddle=x2paddle.convert:main', ]})

tools/merge_params.py

@@ -5,12 +5,14 @@ model_dir = sys.argv[1]
 new_model_dir = sys.argv[2]
 exe = fluid.Executor(fluid.CPUPlace())
 [inference_program, feed_target_names,
- fetch_targets] = fluid.io.load_inference_model(dirname=model_dir, executor=exe)
+ fetch_targets] = fluid.io.load_inference_model(
+     dirname=model_dir, executor=exe)
 
 print(feed_target_names)
-fluid.io.save_inference_model(dirname=new_model_dir,
-                              feeded_var_names=feed_target_names,
-                              target_vars=fetch_targets,
-                              executor=exe,
-                              main_program=inference_program,
-                              params_filename="__params__")
+fluid.io.save_inference_model(
+    dirname=new_model_dir,
+    feeded_var_names=feed_target_names,
+    target_vars=fetch_targets,
+    executor=exe,
+    main_program=inference_program,
+    params_filename="__params__")

x2paddle/__init__.py

-__version__ = "0.7.1"
+__version__ = "0.7.4"

x2paddle/convert.py

-#   Copyright (c) 2019  PaddlePaddle Authors. All Rights Reserved.
+# Copyright (c) 2020  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.
@@ -19,32 +19,37 @@ import sys
 
 def arg_parser():
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model",
-                        "-m",
-                        type=_text_type,
-                        default=None,
-                        help="define model file path for tensorflow or onnx")
-    parser.add_argument("--prototxt",
-                        "-p",
-                        type=_text_type,
-                        default=None,
-                        help="prototxt file of caffe model")
-    parser.add_argument("--weight",
-                        "-w",
-                        type=_text_type,
-                        default=None,
-                        help="weight file of caffe model")
-    parser.add_argument("--save_dir",
-                        "-s",
-                        type=_text_type,
-                        default=None,
-                        help="path to save translated model")
+    parser.add_argument(
+        "--model",
+        "-m",
+        type=_text_type,
+        default=None,
+        help="define model file path for tensorflow or onnx")
+    parser.add_argument(
+        "--prototxt",
+        "-p",
+        type=_text_type,
+        default=None,
+        help="prototxt file of caffe model")
+    parser.add_argument(
+        "--weight",
+        "-w",
+        type=_text_type,
+        default=None,
+        help="weight file of caffe model")
+    parser.add_argument(
+        "--save_dir",
+        "-s",
+        type=_text_type,
+        default=None,
+        help="path to save translated model")
     parser.add_argument(
         "--framework",
         "-f",
         type=_text_type,
         default=None,
-        help="define which deeplearning framework(tensorflow/caffe/onnx)")
+        help="define which deeplearning framework(tensorflow/caffe/onnx/paddle2onnx)"
+    )
     parser.add_argument(
         "--caffe_proto",
         "-c",
@@ -52,27 +57,30 @@ def arg_parser():
         default=None,
         help="optional: the .py file compiled by caffe proto file of caffe model"
     )
-    parser.add_argument("--version",
-                        "-v",
-                        action="store_true",
-                        default=False,
-                        help="get version of x2paddle")
+    parser.add_argument(
+        "--version",
+        "-v",
+        action="store_true",
+        default=False,
+        help="get version of x2paddle")
     parser.add_argument(
         "--without_data_format_optimization",
         "-wo",
         action="store_true",
         default=False,
         help="tf model conversion without data format optimization")
-    parser.add_argument("--define_input_shape",
-                        "-d",
-                        action="store_true",
-                        default=False,
-                        help="define input shape for tf model")
-    parser.add_argument("--params_merge",
-                        "-pm",
-                        action="store_true",
-                        default=False,
-                        help="define whether merge the params")
+    parser.add_argument(
+        "--define_input_shape",
+        "-d",
+        action="store_true",
+        default=False,
+        help="define input shape for tf model")
+    parser.add_argument(
+        "--params_merge",
+        "-pm",
+        action="store_true",
+        default=False,
+        help="define whether merge the params")
 
     return parser
 
@@ -117,7 +125,6 @@ def tf2paddle(model_path,
         optimizer.merge_bias()
         optimizer.optimize_sub_graph()
 
-
 #        optimizer.merge_batch_norm()
 #        optimizer.merge_prelu()
     else:
@@ -165,16 +172,35 @@ 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.onnx.onnx_helper import ONNXOpMapperFactory
     from x2paddle.decoder.onnx_decoder import ONNXDecoder
     from x2paddle.optimizer.onnx_optimizer import ONNXOptimizer
-    import onnxruntime
     model = ONNXDecoder(model_path)
-    mapper = ONNXOpMapper(model, save_dir)
+    factory = ONNXOpMapperFactory()
+    mapper = factory.create_onnx_op_mapper(model)
+    print("Model optimizing ...")
     optimizer = ONNXOptimizer(mapper)
+    print("Model optimized.")
 
-    optimizer.delete_redundance_code()
+    print("Paddle model and code generating ...")
     mapper.save_inference_model(save_dir, params_merge)
+    print("Paddle model and code generated.")
+
+
+def paddle2onnx(model_path, save_dir):
+    from x2paddle.decoder.paddle_decoder import PaddleDecoder
+    from x2paddle.op_mapper.paddle_op_mapper import PaddleOpMapper
+    model = PaddleDecoder(model_path, '__model__', '__params__')
+    mapper = PaddleOpMapper()
+    mapper.convert(model.program, save_dir)
+
+
+def paddle2onnx(model_path, save_dir):
+    from x2paddle.decoder.paddle_decoder import PaddleDecoder
+    from x2paddle.op_mapper.paddle_op_mapper import PaddleOpMapper
+    model = PaddleDecoder(model_path, '__model__', '__params__')
+    mapper = PaddleOpMapper()
+    mapper.convert(model.program, save_dir)
 
 
 def main():
@@ -196,22 +222,13 @@ def main():
     assert args.framework is not None, "--framework is not defined(support tensorflow/caffe/onnx)"
     assert args.save_dir is not None, "--save_dir is not defined"
 
-    if args.framework == "onnx":
-        try:
-            import onnxruntime as rt
-            version = rt.__version__
-            if version != '1.0.0':
-                print("[ERROR] onnxruntime==1.0.0 is required")
-                return
-        except:
-            print(
-                "[ERROR] onnxruntime is not installed, use \"pip install onnxruntime==1.0.0\"."
-            )
-
     try:
         import paddle
         v0, v1, v2 = paddle.__version__.split('.')
-        if int(v0) != 1 or int(v1) < 6:
+        print("paddle.__version__ = {}".format(paddle.__version__))
+        if v0 == '0' and v1 == '0' and v2 == '0':
+            print("[WARNING] You are use develop version of paddlepaddle")
+        elif int(v0) != 1 or int(v1) < 6:
             print("[ERROR] paddlepaddle>=1.6.0 is required")
             return
     except:
@@ -243,11 +260,18 @@ def main():
     elif args.framework == "onnx":
         assert args.model is not None, "--model should be defined while translating onnx model"
         params_merge = False
+
         if args.params_merge:
             params_merge = True
         onnx2paddle(args.model, args.save_dir, params_merge)
+
+    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)
+
     else:
-        raise Exception("--framework only support tensorflow/caffe/onnx now")
+        raise Exception(
+            "--framework only support tensorflow/caffe/onnx/paddle2onnx now")
 
 
 if __name__ == "__main__":

x2paddle/core/fluid_code.py

@@ -46,8 +46,9 @@ class Layer(object):
             for input in self.inputs:
                 if isinstance(input, GraphNode):
                     if hasattr(input, "index"):
-                        in_list += (input.layer_name +
-                                    "[{}]".format(input.index) + ", ")
+                        in_list += (
+                            input.layer_name + "[{}]".format(input.index) + ", "
+                        )
                     else:
                         in_list += (input.layer_name + ", ")
                 elif isinstance(input, six.string_types):
@@ -71,8 +72,8 @@ class Layer(object):
                     layer_code = layer_code + key + "={}, ".format(input)
         elif isinstance(self.inputs, GraphNode):
             if hasattr(self.inputs, "index"):
-                layer_code += (self.inputs.layer_name +
-                               "[{}]".format(self.inputs.index))
+                layer_code += (
+                    self.inputs.layer_name + "[{}]".format(self.inputs.index))
             else:
                 layer_code += (self.inputs.layer_name)
             if self.op != "=":
@@ -88,6 +89,8 @@ class Layer(object):
         for key, value in param_attr.items():
             if '\n' in str(value):
                 value = string(str(value).replace('\n', ','))
+            if str(key) == 'attr':
+                value = 'ParamAttr(' + str(value) + ')'
             layer_code = layer_code + key + "={}, ".format(value)
         layer_code = layer_code.strip(", ")
 

x2paddle/core/op_mapper.py

@@ -29,11 +29,14 @@ def export_paddle_param(param, param_name, dir):
         "bool": [framework_pb2.VarType.BOOL, None]
     }
     shape = param.shape
+    if str(param.dtype) in ['uint8', 'uint_8', 'bool']:
+        param = param.astype('int64')
     if len(shape) == 0:
         assert param.size == 1, "Unexpected situation happend!"
         shape = [1]
-    assert str(param.dtype) in dtype_map, "Unknown dtype of params."
-
+    assert str(
+        param.dtype) in dtype_map, "Unknown dtype {} of params: {}.".format(
+            str(param.dtype), param_name)
     fp = open(os.path.join(dir, param_name), 'wb')
     numpy.array([0], dtype='int32').tofile(fp)
     numpy.array([0], dtype='int64').tofile(fp)
@@ -64,10 +67,8 @@ def run_net(param_dir="./"):
         b = os.path.exists(os.path.join(param_dir, var.name))
         return b
 
-    fluid.io.load_vars(exe,
-                       param_dir,
-                       fluid.default_main_program(),
-                       predicate=if_exist)
+    fluid.io.load_vars(
+        exe, param_dir, fluid.default_main_program(), predicate=if_exist)
 
 
 class OpMapper(object):
@@ -98,8 +99,8 @@ class OpMapper(object):
     def add_codes(self, codes, indent=0):
         if isinstance(codes, list):
             for code in codes:
-                self.paddle_codes += (self.tab * indent + code.strip('\n') +
-                                      '\n')
+                self.paddle_codes += (
+                    self.tab * indent + code.strip('\n') + '\n')
         elif isinstance(codes, str):
             self.paddle_codes += (self.tab * indent + codes.strip('\n') + '\n')
         else:
@@ -135,24 +136,25 @@ class OpMapper(object):
                     os.path.join(os.path.join(py_code_dir, var.name)))
                 return b
 
-            fluid.io.load_vars(exe,
-                               py_code_dir,
-                               fluid.default_main_program(),
-                               predicate=if_exist)
+            fluid.io.load_vars(
+                exe,
+                py_code_dir,
+                fluid.default_main_program(),
+                predicate=if_exist)
             if params_merge:
-                fluid.io.save_inference_model(dirname=os.path.join(
-                    save_dir, "inference_model"),
-                                              feeded_var_names=input_names,
-                                              target_vars=outputs,
-                                              executor=exe,
-                                              params_filename="__params__")
+                fluid.io.save_inference_model(
+                    dirname=os.path.join(save_dir, "inference_model"),
+                    feeded_var_names=input_names,
+                    target_vars=outputs,
+                    executor=exe,
+                    params_filename="__params__")
             else:
-                fluid.io.save_inference_model(dirname=os.path.join(
-                    save_dir, "inference_model"),
-                                              feeded_var_names=input_names,
-                                              target_vars=outputs,
-                                              executor=exe,
-                                              params_filename=None)
+                fluid.io.save_inference_model(
+                    dirname=os.path.join(save_dir, "inference_model"),
+                    feeded_var_names=input_names,
+                    target_vars=outputs,
+                    executor=exe,
+                    params_filename=None)
         except:
             raise Exception(
                 "Paddle code was saved in {}/model.py, but seems there's wrong exist, please check model.py manually."

x2paddle/decoder/caffe_decoder.py

@@ -49,13 +49,11 @@ class CaffeResolver(object):
 class CaffeGraphNode(GraphNode):
     def __init__(self, layer, type_str, layer_name=None):
         if layer_name is None:
-            super(CaffeGraphNode,
-                  self).__init__(layer,
-                                 layer.name.replace('/', '_').replace('-', '_'))
+            super(CaffeGraphNode, self).__init__(
+                layer, layer.name.replace('/', '_').replace('-', '_'))
         else:
-            super(CaffeGraphNode,
-                  self).__init__(layer,
-                                 layer_name.replace('/', '_').replace('-', '_'))
+            super(CaffeGraphNode, self).__init__(
+                layer, layer_name.replace('/', '_').replace('-', '_'))
         self.layer_type = type_str
         self.fluid_code = FluidCode()
         self.data = None
@@ -171,6 +169,14 @@ class CaffeGraph(Graph):
         self.input2layers(input_layers)
         self.transform_input_layers(layers, input_layers)
         layers = input_layers + layers
+        for layer in layers:
+            if hasattr(layer, 'name'):
+                name = getattr(layer, 'name')
+                setattr(layer, 'name', name.replace('/', '_').replace('-', '_'))
+            for i, name in enumerate(layer.bottom):
+                layer.bottom[i] = name.replace('/', '_').replace('-', '_')
+            for i, name in enumerate(layer.top):
+                layer.top[i] = name.replace('/', '_').replace('-', '_')
 
         top_layer = {}
         for layer in layers:
@@ -232,10 +238,12 @@ class CaffeDecoder(object):
 
     def load_using_pb(self):
         data = self.resolver.NetParameter()
-
         data.MergeFromString(open(self.model_path, 'rb').read())
-        pair = lambda layer: (layer.name, self.normalize_pb_data(layer))
         layers = data.layers or data.layer
+        for layer in layers:
+            setattr(layer, 'name',
+                    layer.name.replace('/', '_').replace('-', '_'))
+        pair = lambda layer: (layer.name, self.normalize_pb_data(layer))
         self.params = [pair(layer) for layer in layers if layer.blobs]
 
     def normalize_pb_data(self, layer):
@@ -246,21 +254,20 @@ class CaffeDecoder(object):
                 if layer.type == 'PReLU':
                     c_o, c_i, h, w = map(int, [1] + \
                         list(dims) + [1]* (3 - len(dims)))
-                elif layer.type == 'Normalize':
+                elif layer.type == 'Normalize' and len(dims) == 4:
                     data = np.asarray(list(blob.data), dtype=np.float32)
                     transformed.append(data)
                     continue
                 else:
-                    c_o, c_i, h, w = map(int, [1] * (4 - len(dims)) \
-                        + list(dims))
-
+                    c_o, c_i, h, w = map(int,
+                                         [1] * (4 - len(dims)) + list(dims))
             else:
                 c_o = blob.num
                 c_i = blob.channels
                 h = blob.height
                 w = blob.width
-            data = np.asarray(list(blob.data),
-                              dtype=np.float32).reshape(c_o, c_i, h, w)
+            data = np.asarray(
+                list(blob.data), dtype=np.float32).reshape(c_o, c_i, h, w)
 
             transformed.append(data)
         return transformed

x2paddle/decoder/onnx_decoder.py

@@ -14,6 +14,7 @@
 
 from x2paddle.core.graph import GraphNode, Graph
 from x2paddle.core.fluid_code import FluidCode
+from x2paddle.decoder.onnx_shape_inference import SymbolicShapeInference
 from onnx.checker import ValidationError
 from onnx.checker import check_model
 from onnx.utils import polish_model
@@ -53,7 +54,7 @@ class ONNXGraphNode(GraphNode):
         convert ONNX node attributes to dict
         """
         return {
-            attr.name: self.get_attribute_value2(attr)
+            attr.name: self.get_attribute_value(attr)
             for attr in self.layer.attribute
         }
 
@@ -64,16 +65,15 @@ class ONNXGraphNode(GraphNode):
             return None
         return self.attr_map['value']
 
-    def get_attribute_value2(self, attr):
+    def get_attribute_value(self, attr):
         """
         get_attribute_value enhanced
         """
         if attr.type == onnx.AttributeProto.TENSOR:
             dtype = np.dtype(TENSOR_TYPE_TO_NP_TYPE[attr.t.data_type])
             data = attr.t.raw_data
-            value = np.frombuffer(data,
-                                  dtype=dtype,
-                                  count=(len(data) // dtype.itemsize))
+            value = np.frombuffer(
+                data, dtype=dtype, count=(len(data) // dtype.itemsize))
         elif attr.type == onnx.AttributeProto.STRING:
             value = attr.s
             value = value.decode() if isinstance(value, bytes) else value
@@ -131,43 +131,90 @@ class ONNXGraphDataNode(GraphNode):
 
 class ONNXGraph(Graph):
     def __init__(self, onnx_model):
-        super(ONNXGraph, self).__init__(onnx_model.graph)
-        self.onnx_model = onnx_model
+        super(ONNXGraph, self).__init__(onnx_model)
+        self.fixed_input_shape = {}
         self.initializer = {}
         self.place_holder_nodes = list()
+        self.value_infos = {}
+        self.graph = onnx_model.graph
         self.get_place_holder_nodes()
-        self.value_infos = self.inferred_model_value_info(self.model)
-        self.results_of_inference = dict()
+        print("shape inferencing ...")
+        self.graph = SymbolicShapeInference.infer_shapes(
+            onnx_model, fixed_input_shape=self.fixed_input_shape)
+        print("shape inferenced.")
+        self.build()
+        self.collect_value_infos()
+        self.allocate_shapes()
 
     def get_inner_nodes(self):
         """
         generate inner node of ONNX model
         """
         inner_nodes = []
-        if not isinstance(self.model, onnx.GraphProto):
+        if not isinstance(self.graph, onnx.GraphProto):
             logger.error('graph is not a GraphProto instance')
             return
-        for initializer in self.model.initializer:
+        for initializer in self.graph.initializer:
             name = initializer.name
             inner_nodes.append(name)
         return inner_nodes
 
+    def get_symbolic_shape(self, dims):
+        shape = []
+        for dim in dims:
+            if dim.HasField('dim_param'):
+                shape.append(dim.dim_param)
+            else:
+                shape.append(dim.dim_value)
+        return shape
+
+    def check_input_shape(self, vi):
+        if vi.type.HasField('tensor_type'):
+            for dim in vi.type.tensor_type.shape.dim:
+                if dim.HasField(
+                        'dim_param') and vi.name not in self.fixed_input_shape:
+                    shape = self.get_symbolic_shape(
+                        vi.type.tensor_type.shape.dim)
+                    print(
+                        "Unknown shape for input tensor[tensor name: '{}'] -> shape: {}, Please define shape of input here,\nNote:you can use visualization tools like Netron to check input shape."
+                        .format(vi.name, shape))
+                    right_shape_been_input = False
+                    while not right_shape_been_input:
+                        try:
+                            shape = raw_input(
+                                "Shape of Input(e.g. -1,3,224,224), enter 'N' to skip: "
+                            )
+                        except:
+                            shape = input(
+                                "Shape of Input(e.g. -1,3,224,224), enter 'N' to skip: "
+                            )
+                        if shape.count("-1") > 1:
+                            print("Only 1 dimension can be -1, type again:)")
+                        else:
+                            right_shape_been_input = True
+                    if shape == 'N':
+                        break
+                    shape = [int(dim) for dim in shape.strip().split(',')]
+                    assert shape.count(-1) <= 1, "Only one dimension can be -1"
+                    self.fixed_input_shape[vi.name] = shape
+                    break
+
     def get_place_holder_nodes(self):
         """
         generate place_holder node of ONNX model
         """
         inner_nodes = self.get_inner_nodes()
-        input_nodes = [value.name for value in self.model.input]
-        for ipt_data in input_nodes:
-            if ipt_data not in inner_nodes:
-                self.place_holder_nodes.append(ipt_data)
+        for ipt_vi in self.graph.input:
+            if ipt_vi.name not in inner_nodes:
+                self.check_input_shape(ipt_vi)
+                self.place_holder_nodes.append(ipt_vi.name)
 
     def get_output_nodes(self):
         """
         generate output_nodes node of ONNX model
         """
         inner_nodes = self.get_inner_nodes()
-        output_nodes = [value.name for value in self.model.output]
+        output_nodes = [value.name for value in self.graph.output]
         for opt_data in output_nodes:
             if opt_data not in inner_nodes:
                 self.output_nodes.append(opt_data)
@@ -184,11 +231,11 @@ class ONNXGraph(Graph):
         """
         build topo_sort of ONNX model
         """
-        for layer in self.model.node:
+        for layer in self.graph.node:
             node = ONNXGraphNode(layer)
             self.node_map[layer.name] = node
 
-        for layer in self.model.input:
+        for layer in self.graph.input:
             if layer.name not in self.node_map:
                 is_place_holder = self.is_place_holder_nodes(layer.name)
                 self.node_map[layer.name] = ONNXGraphDataNode(
@@ -197,7 +244,7 @@ class ONNXGraph(Graph):
                     is_global_input=is_place_holder)
 
         #set data node's weight
-        for initializer in self.model.initializer:
+        for initializer in self.graph.initializer:
             name = initializer.name
             weight = to_array(initializer)
             if name in self.node_map:
@@ -205,9 +252,8 @@ class ONNXGraph(Graph):
                     self.node_map[name].weight = weight
                     self.node_map[name].embeded_as = []
             else:
-                self.node_map[name] = ONNXGraphDataNode(initializer,
-                                                        layer_name=name,
-                                                        is_global_input=False)
+                self.node_map[name] = ONNXGraphDataNode(
+                    initializer, layer_name=name, is_global_input=False)
                 self.node_map[name].weight = weight
                 self.node_map[name].embeded_as = []
 
@@ -230,7 +276,7 @@ class ONNXGraph(Graph):
                 continue
             if in_node not in self.node_map:
                 flag = 0
-                for nd in self.model.node:
+                for nd in self.graph.node:
                     for idx, opt in enumerate(nd.output):
                         if opt == in_node:
                             self.connect(nd.name, layer_name)
@@ -258,81 +304,86 @@ class ONNXGraph(Graph):
                 ipt_node.index = node.which_child[ipt_node.layer_name]
             return ipt_node
 
-    def graph_weights(self, graph):
+    def graph_weights(self):
         """
         generator for weights
         """
 
-        if not isinstance(graph, onnx.GraphProto):
+        if not isinstance(self.graph, onnx.GraphProto):
             logger.error('graph is not a GraphProto instance')
             return
 
-        for initializer in graph.initializer:
+        for initializer in self.graph.initializer:
             name = initializer.name
             weight = to_array(initializer)
             yield name, weight
 
-    def inferred_model_value_info(self, graph):
+    def collect_value_infos(self):
         """
         collect value/type info for an ONNX model
         """
-        assert isinstance(graph,
+        assert isinstance(self.graph,
                           onnx.GraphProto), 'model is not a ModelProto instance'
 
-        value_info = Dict()
-        for item in graph.value_info:
-            value_info[item.name] = {
+        for item in self.graph.value_info:
+            self.value_infos[item.name] = {
                 'dtype':
                 TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
                 'shape':
                 [dim.dim_value for dim in item.type.tensor_type.shape.dim],
                 'external': False
             }
-        for item in graph.input:
-            assert item.name not in value_info
-            value_info[item.name] = {
-                'dtype':
-                TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
-                'shape':
-                [dim.dim_value for dim in item.type.tensor_type.shape.dim],
-                'external': True
-            }
-        for item in graph.output:
-            assert item.name not in value_info
-            value_info[item.name] = {
-                'dtype':
-                TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
-                'shape':
-                [dim.dim_value for dim in item.type.tensor_type.shape.dim],
-                'external': True
-            }
-        return value_info
+
+    # for item in self.graph.input:
+    #     self.value_infos[item.name] = {
+    #         'dtype':
+    #         TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
+    #         'shape':
+    #         [dim.dim_value for dim in item.type.tensor_type.shape.dim],
+    #         'external': True
+    #     }
+    # for item in self.graph.output:
+    #     #assert item.name not in value_info
+    #     self.value_infos[item.name] = {
+    #         'dtype':
+    #         TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
+    #         'shape':
+    #         [dim.dim_value for dim in item.type.tensor_type.shape.dim],
+    #         'external': True
+    #     }
+
+    def allocate_shapes(self):
+        """
+        run shape inference
+        """
+        for layer in self.graph.node:
+            node = self.node_map[layer.name]
+            for opt in layer.output:
+                if opt in self.value_infos:
+                    value_info = self.value_infos[opt]
+                    #if len(value_info['shape']) == 0 or value_info[
+                    #        'dtype'] is None or 0 in value_info['shape']:
+                    #    #TODO add node shape inference
+                    node.dtype = value_info['dtype']
+                    node.out_shapes.append(value_info['shape'])
+                else:
+                    node.out_shapes.append([])
 
 
 class ONNXDecoder(object):
     def __init__(self, onnx_model):
-        model = onnx.load(onnx_model)
+        onnx_model = onnx.load(onnx_model)
         print('model ir_version: {}, op version: {}'.format(
-            model.ir_version, model.opset_import[0].version))
-        if model.opset_import[0].version < 9:
-            _logger.warning(
-                'Now, onnx2paddle support convert onnx model opset_verison == 9,'
-                'opset_verison of your onnx model is %d < 9,'
-                'some operator maybe unsuccessful in convertion.',
-                model.opset_import[0].version)
-
-        check_model(model)
-        self.check_model_running_state(onnx_model)
-
-        model = onnx.shape_inference.infer_shapes(model)
-        model = self.optimize_model_skip_op_for_inference(model)
-        model = self.optimize_model_strip_initializer(model)
-        self.standardize_variable_name(model.graph)
-
-        self.model = model
-        graph = model.graph
-        self.onnx_graph = ONNXGraph(model)
-        self.onnx_graph.build()
+            onnx_model.ir_version, onnx_model.opset_import[0].version))
+        self.op_set = onnx_model.opset_import[0].version
+
+        check_model(onnx_model)
+
+        onnx_model = self.optimize_model_skip_op(onnx_model)
+        onnx_model = self.optimize_model_strip_initializer(onnx_model)
+        onnx_model = self.optimize_node_name(onnx_model)
+        self.graph = ONNXGraph(onnx_model)
+        #self.onnx_model = onnx_model
 
     def build_value_refs(self, nodes):
         """
@@ -375,14 +426,13 @@ class ONNXDecoder(object):
                     processed += 1
         return processed
 
-    def optimize_model_skip_op_for_inference(self, model, op_list=None):
+    def optimize_model_skip_op(self, model, op_list=None):
         """
         skip ops can be bypassed for inference
         """
+        nodes = model.graph.node
         if op_list is None:
             op_list = ['Dropout']
-
-        nodes = model.graph.node
         input_refs, output_refs = self.build_value_refs(nodes)
         ret = type(model)()
         ret.CopyFrom(model)
@@ -475,38 +525,11 @@ class ONNXDecoder(object):
             name = name.replace(s, '_')
         return 'x2paddle_' + name
 
-    def check_model_running_state(self, model_path):
-        import onnxruntime as rt
-        model = onnx.load(model_path)
-        model = onnx.shape_inference.infer_shapes(model)
-        if len(model.graph.value_info) < len(model.graph.node) - 1:
-            _logger.warning(
-                'During conversion of your  model, some operators will be assignd node.out_shape==None, '
-                'refer to https://github.com/onnx/onnx/blob/master/docs/ShapeInference.md'
-            )
-        try:
-            datatype_map = {
-                'tensor(int64)': 'int',
-                'tensor(float)': 'float32',
-                'tensor(int32)': 'int32'
-            }
-            input_dict = {}
-            sess = rt.InferenceSession(model_path)
-            for ipt in sess.get_inputs():
-                datatype = datatype_map[ipt.type]
-                input_dict[ipt.name] = np.random.random(
-                    ipt.shape).astype(datatype)
-
-            res = sess.run(None, input_feed=input_dict)
-        except:
-            raise Exception(
-                "onnxruntime inference onnx model failed, Please confirm the correctness of onnx model by onnxruntime, if onnx model is correct, please submit issue in github."
-            )
-
-    def standardize_variable_name(self, graph):
+    def optimize_node_name(self, model):
         """
         standardize variable name for paddle's code
         """
+        graph = model.graph
         for initializer in graph.initializer:
             initializer.name = self.make_variable_name(initializer.name)
         for ipt in graph.input:
@@ -525,3 +548,4 @@ class ONNXDecoder(object):
                     node.input[i] = self.make_variable_name(node.input[i])
             for i in range(len(node.output)):
                 node.output[i] = self.make_variable_name(node.output[i])
+        return model

x2paddle/decoder/paddle_decoder.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 paddle.fluid as fluid
+
+
+class PaddleDecoder(object):
+    def __init__(self,
+                 model_dir,
+                 model_filename='__model__',
+                 params_filename=None):
+        exe = fluid.Executor(fluid.CPUPlace())
+        [self.program, feed, fetchs] = fluid.io.load_inference_model(
+            model_dir,
+            exe,
+            model_filename=model_filename,
+            params_filename=params_filename)

x2paddle/decoder/tf_decoder.py

@@ -48,7 +48,10 @@ class TFGraphNode(GraphNode):
 
     @property
     def out_shapes(self):
-        values = self.layer.attr["_output_shapes"].list.shape
+        if self.layer_type == "OneShotIterator":
+            values = self.layer.attr["output_shapes"].list.shape
+        else:
+            values = self.layer.attr["_output_shapes"].list.shape
         out_shapes = list()
         for value in values:
             shape = [dim.size for dim in value.dim]
@@ -62,6 +65,8 @@ class TFGraphNode(GraphNode):
             dtype = self.layer.attr[k].type
             if dtype > 0:
                 break
+        if dtype == 0:
+            dtype = self.layer.attr['output_types'].list.type[0]
         if dtype not in self.dtype_map:
             raise Exception("Dtype[{}] not in dtype_map".format(dtype))
         return self.dtype_map[dtype]
@@ -115,13 +120,13 @@ class TFGraph(Graph):
     def build(self):
         for layer in self.model.node:
             self.node_map[layer.name.replace('/', '_').replace(
-                '-', '_')] = TFGraphNode(layer, data_format=self.tf_data_format)
+                '-', '_')] = TFGraphNode(
+                    layer, data_format=self.tf_data_format)
 
         for layer_name, node in self.node_map.items():
             for in_node in node.layer.input:
-                in_node = in_node.replace('/',
-                                          '_').replace('-',
-                                                       '_').replace('^', '')
+                in_node = in_node.replace('/', '_').replace('-', '_').replace(
+                    '^', '')
                 if in_node not in self.node_map:
                     if in_node.strip().split(':')[0] in self.node_map:
                         self.connect(in_node.strip().split(':')[0], layer_name)
@@ -226,7 +231,7 @@ class TFGraph(Graph):
     def _remove_identity_node(self):
         identity_ops = [
             'Identity', 'StopGradient', 'Switch', 'Merge',
-            'PlaceholderWithDefault'
+            'PlaceholderWithDefault', 'IteratorGetNext'
         ]
         identity_node = list()
         for node_name, node in self.node_map.items():
@@ -317,7 +322,7 @@ class TFDecoder(object):
         graph_def = cp.deepcopy(graph_def)
         input_map = dict()
         for layer in graph_def.node:
-            if layer.op != "Placeholder":
+            if layer.op != "Placeholder" and layer.op != "OneShotIterator":
                 continue
             graph_node = TFGraphNode(layer)
             dtype = graph_node.layer.attr['dtype'].type
@@ -335,6 +340,14 @@ class TFDecoder(object):
                 if shape.count(-1) > 1:
                     need_define_shape = 2
 
+            if need_define_shape == 1:
+                try:
+                    shape = graph_node.out_shapes[0]
+                    if len(shape) > 0 and shape.count(-1) < 2:
+                        need_define_shape = 0
+                except:
+                    pass
+
             if need_define_shape > 0:
                 shape = None
                 if graph_node.get_attr("shape"):
@@ -377,10 +390,10 @@ class TFDecoder(object):
                         shape=shape,
                         name="x2paddle_{}".format(layer.name))
                 except:
-                    x2paddle_input = tf.placeholder(dtype=dtype,
-                                                    shape=shape,
-                                                    name="x2paddle_{}".format(
-                                                        layer.name))
+                    x2paddle_input = tf.placeholder(
+                        dtype=dtype,
+                        shape=shape,
+                        name="x2paddle_{}".format(layer.name))
 
                 input_map["{}:0".format(layer.name)] = x2paddle_input
                 if shape.count(None) > 0:

x2paddle/op_mapper/caffe_custom_layer/convolutiondepthwise.py

@@ -122,16 +122,17 @@ def convolutiondepthwise_layer(inputs,
     c_out = num_output if num_output is not None else input_shape[0][1]
     group = int(c_in / (c_in / c_out)) if c_in > c_out else int(c_in /
                                                                 (c_out / c_in))
-    out = fluid.layers.conv2d(input,
-                              dilation=[dila_h, dila_w],
-                              filter_size=[k_h, k_w],
-                              stride=[s_h, s_w],
-                              padding=[p_h, p_w],
-                              groups=group,
-                              num_filters=c_out,
-                              param_attr=name + '_weights',
-                              bias_attr=name + '_bias',
-                              name=name)
+    out = fluid.layers.conv2d(
+        input,
+        dilation=[dila_h, dila_w],
+        filter_size=[k_h, k_w],
+        stride=[s_h, s_w],
+        padding=[p_h, p_w],
+        groups=group,
+        num_filters=c_out,
+        param_attr=name + '_weights',
+        bias_attr=name + '_bias',
+        name=name)
     return out
 
 
@@ -142,7 +143,8 @@ def convolutiondepthwise_weights(name, data=None):
     return weights_name
 
 
-register(kind='ConvolutionDepthwise',
-         shape=convolutiondepthwise_shape,
-         layer=convolutiondepthwise_layer,
-         weights=convolutiondepthwise_weights)
+register(
+    kind='ConvolutionDepthwise',
+    shape=convolutiondepthwise_shape,
+    layer=convolutiondepthwise_layer,
+    weights=convolutiondepthwise_weights)

x2paddle/op_mapper/caffe_custom_layer/detectionoutput.py

@@ -12,7 +12,6 @@ def detectionoutput_layer(inputs,
                           share_location=True,
                           keep_top_k=100,
                           confidence_threshold=0.1,
-                          num_classes=2,
                           input_shape=None,
                           name=None):
     nms_param_str = nms_param
@@ -37,9 +36,9 @@ def detectionoutput_layer(inputs,
     pb = fluid.layers.reshape(x=pb, shape=[-1, 4])
     pbv = fluid.layers.reshape(x=pbv, shape=[-1, 4])
     mbox_loc = inputs[0]
-    mbox_loc = fluid.layers.reshape(x=mbox_loc, shape=[0, -1, 4])
-    mbox_conf_flatten = fluid.layers.reshape(x=mbox_conf_flatten,
-                                             shape=[0, -1, num_classes])
+    mbox_loc = fluid.layers.reshape(x=mbox_loc, shape=[-1, pb.shape[0], 4])
+    mbox_conf_flatten = fluid.layers.reshape(
+        x=mbox_conf_flatten, shape=[0, pb.shape[0], -1])
 
     default = {"nms_threshold": 0.3, "top_k": 10, "eta": 1.0}
     fields = ['eta', 'top_k', 'nms_threshold']
@@ -65,7 +64,8 @@ def detectionoutput_weights(name, data=None):
     return weights_name
 
 
-register(kind='DetectionOutput',
-         shape=detectionoutput_shape,
-         layer=detectionoutput_layer,
-         weights=detectionoutput_weights)
+register(
+    kind='DetectionOutput',
+    shape=detectionoutput_shape,
+    layer=detectionoutput_layer,
+    weights=detectionoutput_weights)

x2paddle/op_mapper/caffe_custom_layer/normalize.py

@@ -20,9 +20,8 @@ def normalize_layer(inputs,
         attr=name + '_scale')
     scale_param = fluid.layers.reshape(x=scale_param, \
                   shape=[1] if channel_shared else [input_shape[0][1]])
-    out = fluid.layers.elementwise_mul(x=l2_norm,
-                                       y=scale_param,
-                                       axis=-1 if channel_shared else 1)
+    out = fluid.layers.elementwise_mul(
+        x=l2_norm, y=scale_param, axis=-1 if channel_shared else 1)
     return out
 
 
@@ -31,7 +30,8 @@ def normalize_weights(name, data=None):
     return weights_name
 
 
-register(kind='Normalize',
-         shape=normalize_shape,
-         layer=normalize_layer,
-         weights=normalize_weights)
+register(
+    kind='Normalize',
+    shape=normalize_shape,
+    layer=normalize_layer,
+    weights=normalize_weights)

x2paddle/op_mapper/caffe_custom_layer/permute.py

@@ -23,7 +23,8 @@ def permute_weights(name, data=None):
     return weights_name
 
 
-register(kind='Permute',
-         shape=permute_shape,
-         layer=permute_layer,
-         weights=permute_weights)
+register(
+    kind='Permute',
+    shape=permute_shape,
+    layer=permute_layer,
+    weights=permute_weights)

x2paddle/op_mapper/caffe_custom_layer/priorbox.py

@@ -30,18 +30,19 @@ def priorbox_layer(inputs,
     steps = tuple(step) if type(step) is list or type(step) is tuple else (step,
                                                                            step)
 
-    box, variance_ = fluid.layers.prior_box(input,
-                                            image,
-                                            min_sizes=min_size,
-                                            max_sizes=max_size,
-                                            aspect_ratios=aspect_ratio,
-                                            variance=variance,
-                                            flip=flip,
-                                            clip=clip,
-                                            steps=steps,
-                                            offset=offset,
-                                            name=name,
-                                            min_max_aspect_ratios_order=True)
+    box, variance_ = fluid.layers.prior_box(
+        input,
+        image,
+        min_sizes=min_size,
+        max_sizes=max_size,
+        aspect_ratios=aspect_ratio,
+        variance=variance,
+        flip=flip,
+        clip=clip,
+        steps=steps,
+        offset=offset,
+        name=name,
+        min_max_aspect_ratios_order=True)
     box = fluid.layers.reshape(box, [1, 1, -1])
     variance_ = fluid.layers.reshape(variance_, [1, 1, -1])
     out = fluid.layers.concat([box, variance_], axis=1)
@@ -53,7 +54,8 @@ def priorbox_weights(name, data=None):
     return weights_name
 
 
-register(kind='PriorBox',
-         shape=priorbox_shape,
-         layer=priorbox_layer,
-         weights=priorbox_weights)
+register(
+    kind='PriorBox',
+    shape=priorbox_shape,
+    layer=priorbox_layer,
+    weights=priorbox_weights)

x2paddle/op_mapper/caffe_custom_layer/register.py

@@ -23,8 +23,7 @@ def register(kind, shape, layer, weights):
         kind = [kind]
     else:
         assert type(
-            kind
-        ) is list, 'invalid param "kind" for register, not a list or str'
+            kind) is list, 'invalid param "kind" for register, not a list or str'
 
     for k in kind:
         assert type(

x2paddle/op_mapper/caffe_custom_layer/roipooling.py

@@ -21,11 +21,12 @@ def roipooling_layer(inputs,
     input = inputs[0]
     roi = inputs[1]
     roi = fluid.layers.slice(roi, axes=[1], starts=[1], ends=[5])
-    out = fluid.layers.roi_pool(input,
-                                roi,
-                                pooled_height=pooled_h,
-                                pooled_width=pooled_w,
-                                spatial_scale=spatial_scale)
+    out = fluid.layers.roi_pool(
+        input,
+        roi,
+        pooled_height=pooled_h,
+        pooled_width=pooled_w,
+        spatial_scale=spatial_scale)
     return out
 
 
@@ -34,7 +35,8 @@ def roipooling_weights(name, data=None):
     return weights_name
 
 
-register(kind='ROIPooling',
-         shape=roipooling_shape,
-         layer=roipooling_layer,
-         weights=roipooling_weights)
+register(
+    kind='ROIPooling',
+    shape=roipooling_shape,
+    layer=roipooling_layer,
+    weights=roipooling_weights)

x2paddle/op_mapper/caffe_custom_layer/select.py

@@ -30,11 +30,12 @@ def select_layer(inputs,
     out = []
     for i in range(len(slice_point)):
         out.append(
-            fluid.layers.slice(input,
-                               axes=[axis],
-                               starts=[slice_point[i]],
-                               ends=[slice_point[i + 1]],
-                               name=name + '_' + str(i)))
+            fluid.layers.slice(
+                input,
+                axes=[axis],
+                starts=[slice_point[i]],
+                ends=[slice_point[i + 1]],
+                name=name + '_' + str(i)))
         if i == len(slice_point) - 2:
             break
     return out
@@ -45,7 +46,8 @@ def select_weights(name, data=None):
     return weights_name
 
 
-register(kind='Select',
-         shape=select_shape,
-         layer=select_layer,
-         weights=select_weights)
+register(
+    kind='Select',
+    shape=select_shape,
+    layer=select_layer,
+    weights=select_weights)

x2paddle/op_mapper/caffe_custom_layer/shufflechannel.py

@@ -17,7 +17,8 @@ def shufflechannel_weights(name, data=None):
     return weights_name
 
 
-register(kind='ShuffleChannel',
-         shape=shufflechannel_shape,
-         layer=shufflechannel_layer,
-         weights=shufflechannel_weights)
+register(
+    kind='ShuffleChannel',
+    shape=shufflechannel_shape,
+    layer=shufflechannel_layer,
+    weights=shufflechannel_weights)

x2paddle/op_mapper/caffe_shape.py

@@ -33,8 +33,8 @@ def get_kernel_parameters(params):
         [s_h, s_w] = [params.stride] * 2
     elif len(params.stride) > 0:
         s_h = params.stride_h if params.stride_h > 0 else params.stride[0]
-        s_w = params.stride_w if params.stride_w > 0 else params.stride[
-            len(params.stride) - 1]
+        s_w = params.stride_w if params.stride_w > 0 else params.stride[len(
+            params.stride) - 1]
     elif params.stride_h > 0 or params.stride_w > 0:
         s_h = params.stride_h
         s_w = params.stride_w
@@ -293,12 +293,15 @@ def shape_reshape(layer, input_shape):
             explicit_count *= count(l)
         for i in range(len(copy_axes)):
             explicit_count *= outshape[start_axis + copy_axes[i]]
-        outshape[start_axis + inferred_axis] = -1
-        outshape[0] = 0
-    else:
-        outshape[0] = -1
+        assert input_count % explicit_count == 0, "[Reshape]botom count[%d] "\
+                "must be divisible by product of the specified dimensions[%d] "\
+                % (input_count, explicit_count)
+        outshape[start_axis + inferred_axis] = int(input_count / explicit_count)
 
     output_count = count(outshape)
+    assert output_count == input_count, "[Reshape]output count[%d] must match input count[%d]" % (
+        output_count, input_count)
+    outshape[0] = -1
     return [outshape]
 
 
@@ -342,10 +345,9 @@ def shape_flatten(layer, input_shape):
     output_shape = inshape[0:start_axis]
     if len(inshape[start_axis:end_axis]) != 0:
         flat_sz = reduce(lambda a, b: a * b, inshape[start_axis:end_axis])
-        flat_sz = -1
-        output_shape[0] = 0
         output_shape += [flat_sz]
     output_shape += inshape[end_axis:len(inshape)]
+    output_shape[0] = -1
     return [output_shape]
 
 

x2paddle/op_mapper/onnx_custom_layer/InstanceNormalization.py → x2paddle/op_mapper/onnx/custom_layer/InstanceNormalization.py

@@ -24,21 +24,18 @@ def InstanceNormalization_layer(inputs, name=None):
     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)
+    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")
+    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)
@@ -51,8 +48,9 @@ def InstanceNormalization_weights(name, data=None):
     return weights_name
 
 
-register(kind='InstanceNormalization',
-         shape=InstanceNormalization_shape,
-         layer=InstanceNormalization_layer,
-         child_func=None,
-         weights=InstanceNormalization_weights)
+register(
+    kind='InstanceNormalization',
+    shape=InstanceNormalization_shape,
+    layer=InstanceNormalization_layer,
+    child_func=None,
+    weights=InstanceNormalization_weights)

x2paddle/op_mapper/onnx_custom_layer/register.py → x2paddle/op_mapper/onnx/custom_layer/register.py

@@ -36,8 +36,7 @@ def register(kind, shape, layer, child_func, weights):
         kind = [kind]
     else:
         assert type(
-            kind
-        ) is list, 'invalid param "kind" for register, not a list or str'
+            kind) is list, 'invalid param "kind" for register, not a list or str'
 
     for k in kind:
         assert type(

x2paddle/op_mapper/onnx/onnx_helper.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.
+
+from x2paddle.op_mapper.onnx.opset9 import ONNXOpMapperOpSet9
+
+
+class ONNXOpMapperFactory:
+    def __init__(self):
+        self.support_op_sets = [9, ]
+        self.default_op_set = 9
+
+    def create_onnx_op_mapper(self, decoder):
+        run_op_set = self.default_op_set
+        OpMapper = ''
+        if decoder.op_set in self.support_op_sets:
+            OpMapper = 'ONNXOpMapperOpSet' + str(decoder.op_set)
+        elif decoder.op_set < self.default_op_set:
+            OpMapper = 'ONNXOpMapperOpSet' + str(self.default_op_set)
+        else:
+            for op_set in self.support_op_sets:
+                if decoder.op_set > op_set:
+                    run_op_set = op_set
+                else:
+                    break
+            OpMapper = 'ONNXOpMapperOpSet' + str(run_op_set)
+        print(
+            'Now, onnx2paddle support convert onnx model opset_verison {},'
+            'opset_verison of your onnx model is {}, automatically treated as op_set: {}.'
+            .format(self.support_op_sets, decoder.op_set, run_op_set))
+        return eval(OpMapper)(decoder)

x2paddle/op_mapper/onnx_directly_map.py

@@ -28,61 +28,55 @@ default_op_mapping_field_values['FILL_NAME_FIELD'] = True
 default_op_mapping = {
     'Shape': ['shape', ['X'], ['Out']],
     'Clip': [
-        'clip', ['X'], ['Out'],
-        dict(),
-        dict(
-            min=(_np.asarray([255, 255, 127, 255],
-                             dtype=_np.uint8).view(_np.float32)[0]),
-            max=(_np.asarray([255, 255, 127, 127],
-                             dtype=_np.uint8).view(_np.float32)[0]),
-        )
+        'clip', ['X'], ['Out'], dict(), dict(
+            min=(_np.asarray(
+                [255, 255, 127, 255], dtype=_np.uint8).view(_np.float32)[0]),
+            max=(_np.asarray(
+                [255, 255, 127, 127], dtype=_np.uint8).view(_np.float32)[0]), )
     ],
     'Erf': ['erf', ['X'], ['Out']],
     'Ceil': ['ceil', ['X'], ['Out']],
     'ReduceMean': [
-        'reduce_mean', ['X'], ['Out'],
-        dict(axes='dim', keepdims='keep_dim'),
-        dict(keep_dim=1)
+        'reduce_mean', ['X'], ['Out'], dict(
+            axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
     ],
     'ReduceSum': [
-        'reduce_sum', ['X'], ['Out'],
-        dict(axes='dim', keepdims='keep_dim'),
-        dict(keep_dim=1)
+        'reduce_sum', ['X'], ['Out'], dict(
+            axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
     ],
     'ReduceMin': [
-        'reduce_min', ['X'], ['Out'],
-        dict(axes='dim', keepdims='keep_dim'),
-        dict(keep_dim=1)
+        'reduce_min', ['X'], ['Out'], dict(
+            axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
+    ],
+    'ReduceMax': [
+        'reduce_max', ['X'], ['Out'], dict(
+            axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
     ],
     #active function
     'Relu': ['relu', ['X'], ['Out']],
-    'LeakyRelu': ['leaky_relu', ['X'], ['Out'],
-                  dict(), dict(alpha=.01)],
-    'Elu': ['elu', ['X'], ['Out'],
-            dict(), dict(alpha=1.)],
+    'LeakyRelu': ['leaky_relu', ['X'], ['Out'], dict(), dict(alpha=.01)],
+    'Elu': ['elu', ['X'], ['Out'], dict(), dict(alpha=1.)],
     'ThresholdedRelu': [
-        'thresholded_relu', ['X'], ['Out'],
-        dict(alpha='threshold'),
+        'thresholded_relu', ['X'], ['Out'], dict(alpha='threshold'),
         dict(alpha=1.)
     ],
     'Tanh': ['tanh', ['X'], ['Out']],
     'Sigmoid': ['sigmoid', ['X'], ['Out']],
     'HardSigmoid': [
-        'hard_sigmoid', ['X'], ['Out'],
-        dict(alpha='slope', beta='offset'),
-        dict(slope=.2, offset=.5)
+        'hard_sigmoid', ['X'], ['Out'], dict(
+            alpha='slope', beta='offset'), dict(
+                slope=.2, offset=.5)
     ],
     'Softsign': ['softsign', ['X'], ['Out']],
     'Softplus': ['softplus', ['X'], ['Out']],
     'Exp': ['exp', ['X'], ['Out']],
-    'Softmax': ['softmax', ['X'], ['Out'],
-                dict(), dict(axis=1)],
+    'Softmax': ['softmax', ['X'], ['Out'], dict(), dict(axis=1)],
     'Sqrt': ['sqrt', ['X'], ['Out']],
     'Floor': ['floor', ['X'], ['Out']],
     'Abs': ['abs', ['X'], ['Out']],
 }
 
 default_ioa_constraint = {
-    'Gather':
-    [(lambda i, o, a: a.get('axis', 0) == 0, 'only axis = 0 is supported')],
+    'Gather': [(lambda i, o, a: a.get('axis', 0) == 0,
+                'only axis = 0 is supported')],
 }

x2paddle/op_mapper/paddle_custom_layer/im2sequence.py

+import onnx
+import numpy as np
+from onnx import onnx_pb, helper
+
+im2seq_counter = 0
+
+
+def im2sequence(op, block):
+    global im2sequence_counter
+    n, c, h, w = block.var(op.input('X')[0]).shape
+    assert h > 0 and w > 0, "Only supported fixed input shape for im2sequence operator."
+    stride_h, stride_w = op.attr('strides')
+    paddings = op.attr('paddings')
+    assert op.attr(
+        'out_stride'
+    ) != 1, "Only out_stride==1 is supported for im2sequence operator."
+    h = h + paddings[0] + paddings[1]
+    w = w + paddings[1] + paddings[2]
+    kernel_h, kernel_w = op.attr('kernels')
+    out_h = 1 + (h - kernel_h + stride_h - 1) // stride_h
+    out_w = 1 + (w - kernel_w + stride_w - 1) // stride_w
+    h_steps = list()
+    for i in range(out_h):
+        h_steps.append([i * stride_h, i * stride_h + kernel_h])
+    w_steps = list()
+    for i in range(out_w):
+        w_steps.append([i * stride_w, i * stride_w + kernel_w])
+
+    nodes = list()
+    slice_blocks = list()
+    for i in range(out_h):
+        for j in range(out_w):
+            starts_name = "im2sequence.starts.{}.{}.{}".format(im2seq_counter,
+                                                               i, j)
+            starts_tensor = helper.make_tensor(
+                name=starts_name,
+                data_type=onnx_pb.TensorProto.INT64,
+                dims=[4],
+                vals=[0, 0, h_steps[i][0], w_steps[j][0]])
+            ends_name = "im2sequence.ends.{}.{}.{}".format(im2seq_counter, i, j)
+            ends_tensor = helper.make_tensor(
+                name=ends_name,
+                data_type=onnx_pb.TensorProto.INT64,
+                dims=[4],
+                vals=[999999, 999999, h_steps[i][1], w_steps[j][1]])
+            starts_node = helper.make_node(
+                'Constant',
+                inputs=[],
+                outputs=[starts_name],
+                value=starts_tensor)
+            ends_node = helper.make_node(
+                'Constant', inputs=[], outputs=[ends_name], value=ends_tensor)
+            nodes.extend([starts_node, ends_node])
+
+            slice_block_name = "im2sequence.slice.{}.{}.{}".format(
+                im2seq_counter, i, j)
+            slice_block_node = helper.make_node(
+                'Slice',
+                inputs=[op.input('X')[0], starts_name, ends_name],
+                outputs=[slice_block_name])
+            flatten_block_name = "im2sequence.flatten.{}.{}.{}".format(
+                im2seq_counter, i, j)
+            flatten_block_node = helper.make_node(
+                "Flatten",
+                inputs=[slice_block_name],
+                outputs=[flatten_block_name],
+                axis=0)
+            nodes.extend([slice_block_node, flatten_block_node])
+            slice_blocks.append(flatten_block_name)
+    concat_block_name = "im2sequence.concat_block.{}".format(im2seq_counter)
+    #    concat_block_node = helper.make_node("Concat", inputs=slice_blocks, outputs=[concat_block_name], axis=0)
+    concat_block_node = helper.make_node(
+        "Concat", inputs=slice_blocks, outputs=op.output('Out'), axis=0)
+    nodes.append(concat_block_node)
+    print("\n\n==========Importance Notice===========")
+    print(
+        "Since im2sequence operator is used in your paddlepaddle model, the translated onnx model only support input data with batch_size=1."
+    )
+    print("======================================\n")
+    return nodes

x2paddle/op_mapper/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)
+
+    # Ine 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_concat_final_results = outputs['Out']
+    node_concat_final_results = onnx.helper.make_node(
+        'Concat',
+        inputs=inputs_concat_final_results,
+        outputs=outputs_concat_final_results,
+        axis=2)
+    node_list.append(node_concat_final_results)
+
+    return node_list

x2paddle/optimizer/caffe_optimizer.py

@@ -41,10 +41,11 @@ class CaffeOptimizer(object):
                     if is_delete_node:
                         parent_node.fluid_code.clear()
                     node.fluid_code.clear()
-                    node.fluid_code.add_layer("batch_norm",
-                                              inputs=input,
-                                              output=node,
-                                              param_attr=parent_param_attr)
+                    node.fluid_code.add_layer(
+                        "batch_norm",
+                        inputs=input,
+                        output=node,
+                        param_attr=parent_param_attr)
 
     def merge_op_activation(self):
         for node_name in self.graph.topo_sort:
@@ -62,7 +63,8 @@ class CaffeOptimizer(object):
                     if is_delete_node:
                         parent_node.fluid_code.clear()
                     node.fluid_code.clear()
-                    node.fluid_code.add_layer(op,
-                                              inputs=input,
-                                              output=node,
-                                              param_attr=parent_param_attr)
+                    node.fluid_code.add_layer(
+                        op,
+                        inputs=input,
+                        output=node,
+                        param_attr=parent_param_attr)

x2paddle/optimizer/onnx_optimizer.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 # TODO useless node remove
-from x2paddle.op_mapper.onnx_op_mapper import ONNXOpMapper
 
 
 class ONNXOptimizer(object):

x2paddle_model_zoo.md

@@ -13,6 +13,7 @@
 | ShuffleNet | [code](https://github.com/TropComplique/shufflenet-v2-tensorflow) |-|
 | mNASNet | [code](https://github.com/tensorflow/tpu/tree/master/models/official/mnasnet) |-|
 | EfficientNet | [code](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet) |-|
+| Inception_V3 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v3.py) |-|
 | Inception_V4 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v4.py) |-|
 | Inception_ResNet_V2 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_resnet_v2.py) |-|
 | VGG16 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-|
@@ -47,8 +48,8 @@
 ## ONNX
 **注：** 部分模型来源于PyTorch，PyTorch的转换可参考[pytorch_to_onnx.md](pytorch_to_onnx.md)
 
-| 模型 | 来源 | operator version|
-|-------|--------|---------|
+| 模型 | 来源 | operator version|备注|
+|-------|--------|---------|---------|
 | ResNet18 | [torchvison.model.resnet18](https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py) |9|
 | ResNet34 | [torchvison.model.resnet34](https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py) |9|
 | ResNet50 | [torchvison.model.resnet50](https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py) |9|
@@ -64,4 +65,6 @@
 | mNASNet | [pytorch(personal practice)](https://github.com/rwightman/gen-efficientnet-pytorch) |9|
 | EfficientNet | [pytorch(personal practice)](https://github.com/rwightman/gen-efficientnet-pytorch) |9|
 | SqueezeNet | [onnx official](https://s3.amazonaws.com/download.onnx/models/opset_9/squeezenet.tar.gz) |9|
-|Ultra-Light-Fast-Generic-Face-Detector-1MB| [onnx_model](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master/models/onnx)| |
+|Ultra-Light-Fast-Generic-Face-Detector-1MB| [onnx_model](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master/models/onnx)|9 |
+|BERT| [pytorch(huggingface)](https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb)|11|转换时需指定input shape，见[文档Q3](FAQ.md)|
+|GPT2| [pytorch(huggingface)](https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb)|11|转换时需指定input shape，见[文档Q3](FAQ.md)|