Merge pull request #2 from PaddlePaddle/develop

dev

FAQ.md

@@ -5,29 +5,10 @@ A：该提示信息表示无法从TensorFlow的pb模型中获取到输入tensor(
 
 
 **Q2. TensorFlow模型转换失败怎么解决?**  
-A: 目前TensorFlow模型转换失败存在几个问题。1) 存在暂未支持的OP，此信息会在转换时输出; 2) NHWC优化导致部分参数出错；3）Batch维度带来的出错 4）其它
-
-对于（1）问题，建议自行添加或发起Issue；
-
-其中（2）、(3)、（4）问题目前没有明确的报错信息，当您遇到模型转换失败时，请尝试如下的步骤后，再进行转换测试
-
+A: 如果并非是由缺少OP导致，那可能是由于TensorFlow模型转换时（NHWC->NCHW格式转换导致），在这种情况下，可以采用关闭格式优化，同时固化输入大小的方式，继续尝试转换，见如下命令，转换过程中，根据提示，输入相应tensor的固化shape大小
 ```
 x2paddle -f tensorflow -m tf.pb -s pd-model --without_data_format_optimization --define_input_shape
 ```
 
-#### without_data_format_optimization : 关闭NHWC优化
-TensorFlow的CV模型，大多的输入格式为`NHWC`，而Paddle目前仅支持`NCHW`，如若直接转换，需要在conv2d、pool2d等操作前后添加transpose解决，这样会带来性能的损耗。X2Paddle在模型转换过程中，对此问题进行了优化，避免transpose操作带来的性能问题，但目前仅在部分模型上进行了测试，不一定适用于其它模型，因此，如若模型转换存在问题时，我们建议你关闭NHWC的优化。
-
-在模型转换时添加参数 --without_data_format_optimization
-```
-x2paddle -f tensorflow -m tf.pb -s pd-model --without_data_format_optimization
-```
-
-### define_input_shape : 固定Batch大小
-受限于不同框架的运行机制，在转换过程中，Batch维度也有一定可能会带来模型转换失败的问题。可以尝试固定Batch维度后再转换
-
-在模型转换时添加参数 --define_input_shape
-```
-x2paddle -f tensorflow -m tf.pb -s pd-model --define_input_shape
-```
-如原tensorflow模型的输入shape为`[None, 224, 224, 3]`，可添加参数后，根据提示，把输入的shape修改为`[2, 224, 224, 3]`
+> 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。  
+> 2. X2Paddle默认情况，TensorFlow模型转换后得到的Paddle模型为`NCHW`的输入格式。但在指定`--withou_data_format_optimization`后，转换后的Paddle模型输入格式也同样为`NHWC`。

README.md

@@ -4,24 +4,19 @@
 X2Paddle支持将其余深度学习框架训练得到的模型，转换至PaddlePaddle模型。  
 X2Paddle is a toolkit for converting trained model to PaddlePaddle from other deep learning frameworks.
 
-## 更新历史
-2019.08.05  
-1. 统一tensorflow/caffe/onnx模型转换代码和对外接口
-2. 解决上一版caffe2fluid无法转换多分支模型的问题
-3. 解决Windows上保存模型无法加载的问题
-4. 新增optimizer，优化代码结构，合并conv、batch_norm的bias和激活函数  
-
-**如果你需要之前版本的tensorflow2fluid/caffe2fluid/onnx2fluid，可以继续访问release-0.3分支，获取之前版本的代码使用。**
+## 转换模型库
+X2Paddle在多个主流的CV模型上，测试过TensorFlow/Caffe/ONNX模型的转换，可以在[X2Paddle-Model-Zoo](x2paddle_model_zoo.md)查看我们的模型测试列表。如果你在新的模型上进行了测试转换，也欢迎继续补充该列表；如若无法转换，可通过ISSUE反馈给我们，我们会尽快跟进。
 
 ## 环境依赖
 
 python >= 3.5  
 paddlepaddle >= 1.5.0  
 
-**以下依赖只需对应安装自己需要的即可**  
-转换tensorflow模型 ： tensorflow == 1.14.0  
-转换caffe模型 ： caffe == 1.0.0  
-转换onnx模型 ： onnx == 1.5.0  pytorch == 1.1.0
+**按需安装以下依赖**  
+tensorflow ： tensorflow == 1.14.0  
+caffe ： caffe == 1.0.0  
+onnx ： onnx == 1.5.0  pytorch == 1.1.0
+
 ## 安装
 ### 安装方式一（推荐）
 使用最新的代码版本，可使用如下方式进行安装  
@@ -63,18 +58,36 @@ x2paddle --framework=onnx --model=onnx_model.onnx --save_dir=pd_model
 |--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 |
-|--model | 当framework为tensorflow/pmmx时，该参数指定tensorflow的pb模型文件或onnx模型路径 |
-|--caffe_proto | [可选]由caffe.proto编译成caffe_pb2.py文件的存放路径，当存在自定义Layer时使用，默认为None |
+|--model | 当framework为tensorflow/onnx时，该参数指定tensorflow的pb模型文件或onnx模型路径 |
+|--caffe_proto | **[可选]** 由caffe.proto编译成caffe_pb2.py文件的存放路径，当存在自定义Layer时使用，默认为None |
+|--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) |
+
 
 ## 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  
 `inference_model`中保存了序列化的模型结构和参数，可直接使用paddle的接口进行加载，见[load_inference_model](https://www.paddlepaddle.org.cn/documentation/docs/zh/1.5/api_guides/low_level/inference.html#api-guide-inference)
 
+## 小工具
+X2Paddle提供了工具解决如下问题，详见[tools/README.md](tools/README.md)
+1. 检测模型是否在PaddleLite中支持  
+2. 合并模型参数文件
+
 ## 相关文档
 1. [X2Paddle使用过程中常见问题](FAQ.md)  
 2. [如何导出TensorFlow的pb模型](export_tf_model.md)
-3. [X2Paddle测试模型库](x2paddle_model_zoo.md)
+3. [X2Paddle测试模型库](x2paddle_model_zoo.md)  
+4. [PyTorch模型导出为ONNX模型](pytorch_to_onnx.md)
+
+## 更新历史
+2019.08.05  
+1. 统一tensorflow/caffe/onnx模型转换代码和对外接口
+2. 解决上一版caffe2fluid无法转换多分支模型的问题
+3. 解决Windows上保存模型无法加载的问题
+4. 新增optimizer，优化代码结构，合并conv、batch_norm的bias和激活函数  
+
+**如果你需要之前版本的tensorflow2fluid/caffe2fluid/onnx2fluid，可以继续访问release-0.3分支，获取之前版本的代码使用。**
 
 
 ## Acknowledgements

pytorch_to_onnx.md

+## PyTorch模型导出为ONNX模型
+
+目前onnx2paddle主要支持onnx operator version 9。 用户可通过如下示例代码，将torchvison或者自己开发写的模型转换成onnx model:
+```
+#coding: utf-8
+import torch
+import torchvision
+
+# 指定输入大小的shape
+dummy_input = torch.randn(1, 3, 224, 224)
+
+# 构建pytorch model，并载入模型参数
+resnet18 = torchvision.models.resnet18(pretrained=True)
+
+# 导出resnet18.onnx模型文件
+torch.onnx.export(resnet18, dummy_input, "resnet18.onnx",verbose=True)
+
+```

tools/README.md

+### 一、PaddleLite部署
+使用X2Paddle转换后的模型均可以使用Paddle Fluid进行预测。但对于PaddleLite上的部署，则需要检查模型中的OP是否都在PaddleLite中支持。使用`check_for_lite.py`可以进行检查。
+
+```
+python tools/check_for_lite.py paddle_model/inference_model/__model__
+```
+
+
+### 二、模型参数合并
+X2Paddle转换后产出的路径下包括两个目录，  
+1. `model_with_code`: 包含保存的参数文件和模型python代码文件，供用户debug  
+2. `inference_model`: 参数文件和序列化的模型结构文件，供用户预测部署  
+
+其中在`inference_model`中，X2Paddle将每个参数独立保存在不同的文件中（文件名和参数名一致），用户可使用`merge_params.py`将参数文件合并成一个文件使用
+```
+python tools/merge_params.py paddle_model/inference_model  new_model_dir
+```
+合并参数后的模型保存在`new_model_dir`中

x2paddle/__init__.py

-__version__ = "0.4.3"
+__version__ = "0.4.5"

x2paddle/convert.py

@@ -142,14 +142,17 @@ def onnx2paddle(model_path, save_dir):
     except:
         print("onnx is not installed, use \"pip install onnx==1.5.0\".")
         return
+    print("Now translating model from onnx to paddle.")
 
     from x2paddle.decoder.onnx_decoder import ONNXDecoder
-    from x2paddle.op_mapper.onnx_op_mapper import ONNXOpMapper
-    from x2paddle.optimizer.onnx_optimizer import ONNXOptimizer
-    print("Now translating model from onnx to paddle.")
     model = ONNXDecoder(model_path)
+
+    from x2paddle.op_mapper.onnx_op_mapper import ONNXOpMapper
     mapper = ONNXOpMapper(model)
+
+    from x2paddle.optimizer.onnx_optimizer import ONNXOptimizer
     optimizer = ONNXOptimizer(mapper)
+
     optimizer.delete_redundance_code()
     mapper.save_inference_model(save_dir)
 

x2paddle/decoder/onnx_decoder.py

@@ -23,6 +23,7 @@ from onnx.helper import get_attribute_value, make_attribute
 from onnx.shape_inference import infer_shapes
 from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
 from onnx.numpy_helper import to_array
+from onnx import AttributeProto, TensorProto, GraphProto
 from collections import OrderedDict as Dict
 import onnx
 import numpy as np
@@ -44,7 +45,7 @@ class ONNXGraphNode(GraphNode):
         self.attr_map = self.get_attr_map()
         self.dtype_map = {1: "float32", 3: "int32", 9: "int64"}
         self.weight_inputs = list()
-        self.out_shapes = None
+        self.out_shapes = list()
         self.dtype = None
 
     def get_attr_map(self):
@@ -58,11 +59,10 @@ class ONNXGraphNode(GraphNode):
 
     @property
     def value(self):
-        assert 'Constant' in self.layer_type, "Only Constant node has value."
-
-        attr = self.layer.attr['value']
-        if 'value' in self.attr_map:
-            return default
+        assert 'Constant' in self.layer_type, "Only Constant | ConstantOfShape node has value."
+        attr = self.layer.attribute['value']
+        if 'value' not in self.attr_map:
+            return None
         return self.attr_map[name]
 
     def get_attribute_value2(self, attr):
@@ -110,23 +110,26 @@ class ONNXGraphDataNode(GraphNode):
     def out_shapes(self):
         values = self.layer.type.tensor_type.shape.dim
         out_shapes = list()
-        out_shapes = [dim.dim_value for dim in values]
+        out_shapes.append([dim.dim_value for dim in values])
         return out_shapes
 
     @property
     def dtype(self):
         dtype = self.layer.type.tensor_type.elem_type
-
         return TENSOR_TYPE_TO_NP_TYPE[dtype]
 
 
 class ONNXGraph(Graph):
-    def __init__(self, model):
-        super(ONNXGraph, self).__init__(model)
+    def __init__(self, graph, onnx_model):
+        super(ONNXGraph, self).__init__(graph)
+        self.onnx_model = onnx_model
         self.initializer = {}
         self.place_holder_nodes = list()
         self.get_place_holder_nodes()
 
+        self.value_infos = self.inferred_model_value_info(graph)
+        self.results_of_inference = dict()
+
     def get_inner_nodes(self):
         """
         generate inner node of ONNX model
@@ -162,17 +165,22 @@ class ONNXGraph(Graph):
         """
         build topo_sort of ONNX model
         """
+        data_node = self.place_holder_nodes[0]
+        value_info = self.value_infos[data_node]
+        input_shape = value_info['shape']
+        self.get_results_of_inference(self.onnx_model, input_shape)
         for layer in self.model.node:
-            self.node_map[layer.name] = ONNXGraphNode(layer)
-
-        #set op node's dtype and out_shapes
-        for item in self.model.value_info:
-            if item.name in self.node_map:
-                self.node_map[item.name].dtype = TENSOR_TYPE_TO_NP_TYPE[
-                    item.type.tensor_type.elem_type]
-                self.node_map[item.name].out_shapes = [
-                    dim.dim_value for dim in item.type.tensor_type.shape.dim
-                ]
+            node = ONNXGraphNode(layer)
+            self.node_map[layer.name] = node
+            for opt in layer.output:
+                if opt in self.value_infos:
+                    value_info = self.value_infos[opt]
+                    node.dtype = value_info['dtype']
+                    node.out_shapes.append(value_info['shape'])
+                else:
+                    _, dtype, shape = self.get_dynamic_shape(opt)
+                    node.dtype = dtype
+                    node.out_shapes.append(shape)
 
         for layer in self.model.input:
             if layer.name not in self.node_map:
@@ -199,7 +207,6 @@ class ONNXGraph(Graph):
                             format(in_node, layer_name))
                     else:
                         self.connect(in_node, layer_name)
-
         #generate topo
         super(ONNXGraph, self).build()
 
@@ -227,31 +234,108 @@ class ONNXGraph(Graph):
             weight = to_array(initializer)
             yield name, weight
 
+    def inferred_model_value_info(self, graph):
+        """
+        collect value/type info for an ONNX model
+        """
+        assert isinstance(graph,
+                          onnx.GraphProto), 'model is not a ModelProto instance'
+
+        value_info = Dict()
+        for item in graph.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': 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
+
+    def get_results_of_inference(self, model, shape):
+        try:
+            import torch
+            version = torch.__version__
+            if '1.1.0' not in version:
+                print("your model have dynamic graph, torch==1.1.0 is required")
+                return
+        except:
+            print(
+                "your model have dynamic graph, we use caff2 to inference graph, please use \"pip install torch==1.1.0\"."
+            )
+            return
+        from x2paddle.decoder.onnx_backend import prepare
+
+        np_images = np.random.rand(shape[0], shape[1], shape[2],
+                                   shape[3]).astype('float32')
+
+        outputs = []
+        for node in model.graph.node:
+            value_info = helper.make_tensor_value_info(node.name,
+                                                       TensorProto.UNDEFINED,
+                                                       [])
+            outputs.append(value_info)
+
+        while len(outputs) > 0:
+            tmp_outputs = outputs[:254]
+            model.graph.ClearField('output')
+            model.graph.output.MergeFrom(tmp_outputs)
+            prepared_backend = prepare(model,
+                                       device='CPU',
+                                       no_check_UNSAFE=True)
+            res = prepared_backend.run(inputs=np_images)
+            for idx, info in enumerate(tmp_outputs):
+                self.results_of_inference[info.name] = res[idx]
+            outputs = outputs[254:]
+        return
+
+    def get_dynamic_shape(self, layer):
+        """
+        get dynamic shape from caffe2.backend
+        """
+        output = self.results_of_inference[layer]
+        return output.tolist(), output.dtype, output.shape
+
 
 class ONNXDecoder(object):
     def __init__(self, onnx_model):
         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 main support convert onnx model opset_verison == 9,'
                 'opset_verison of your onnx model is %d < 9,'
                 'some operator may cannot convert.',
                 model.opset_import[0].version)
-        check_model(model)
-
-        model = polish_model(model)
 
+        check_model(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_def = model.graph
-
-        self.onnx_graph = ONNXGraph(graph_def)
+        self.onnx_graph = ONNXGraph(graph_def, model)
         self.onnx_graph.build()
 
     def build_value_refs(self, nodes):
@@ -334,9 +418,13 @@ class ONNXDecoder(object):
                                                     output_name, output_refs)
             else:
                 processed = -1
-
             if processed > 0:
                 nodes_to_remove.append(node_idx)
+                for value_info in ret.graph.value_info:
+                    for output in node.output:
+                        if value_info.name == output:
+                            ret.graph.value_info.remove(value_info)
+
                 print('skip op {}: {} -> {} -> {}'.format(
                     node_idx, input_name, node.op_type, output_name))
             elif processed == 0:
@@ -396,7 +484,6 @@ class ONNXDecoder(object):
         """
         standardize variable name for paddle's code
         """
-
         for initializer in graph.initializer:
             initializer.name = self.make_variable_name(initializer.name)
         for ipt in graph.input:
@@ -455,43 +542,3 @@ class ONNXDecoder(object):
             raise RuntimeError("Input mismatch {} != {}".format(
                 len(onnx_model.input), len(model.input)))
         return onnx_model
-
-    def get_dynamic_shape_from_caffe2(self, layer, input_shapes):
-        """
-        get dynamic shape from caffe2.backend
-        """
-        try:
-            import torch
-            version = torch.__version__
-            if '1.1.0' not in version:
-                print("your model have dynamic graph, torch==1.1.0 is required")
-                return
-        except:
-            print(
-                "your model have dynamic graph, we use caff2 to inference graph, please use \"pip install torch==1.1.0\"."
-            )
-            return
-        from caffe2.python.onnx.backend import prepare
-        shape = input_shapes[0]
-        np_images = np.random.rand(shape[0], shape[1], shape[2],
-                                   shape[3]).astype('float32')
-        num_onnx = self.split_model(self.model, layer)
-        prepared_backend = prepare(num_onnx, device='CPU')
-        output = prepared_backend.run(inputs=np_images)
-        return output[0].tolist()
-
-    def get_dynamic_shape_from_onnx(self, layer, input_shapes):
-        """
-        get dynamic shape from onnxruntime
-        """
-        import onnxruntime as rt
-        from onnxruntime.backend import prepare
-        import numpy as np
-        num_onnx = self.split_model(self.model, layer)
-        sess = prepare(num_onnx)
-        shape = input_shapes[0]
-        print(shape)
-        np_images = np.random.rand(shape[0], shape[1], shape[2],
-                                   shape[3]).astype('float32')
-        output = sess.run(model=sess, inputs=np_images)
-        return output[0].tolist()

x2paddle/decoder/tf_decoder.py

@@ -176,7 +176,7 @@ class TFGraph(Graph):
     def _remove_identity_node(self):
         identity_node = list()
         for node_name, node in self.node_map.items():
-            if node.layer_type == "Identity":
+            if node.layer_type == "Identity" or node.layer_type == "StopGradient":
                 identity_node.append(node_name)
 
         for node_name in identity_node:
@@ -374,3 +374,38 @@ class TFDecoder(object):
             return results[0].tolist()
         else:
             raise Exception("Couldn't infer a stable shape shape tensor value")
+
+    def infer_tensor_shape(self, graph_node):
+        if hasattr(graph_node, "index"):
+            tensor_name = graph_node.layer.name + ":{}".format(graph_node.index)
+        else:
+            tensor_name = graph_node.layer.name + ":0"
+        feed = dict()
+        batch_size = [2, 3, 5]
+        shapes = list()
+        for b in batch_size:
+            for input_name, info in self.input_info.items():
+                (shape, dtype) = cp.deepcopy(info)
+                input_tensor = self.sess.graph.get_tensor_by_name(input_name +
+                                                                  ":0")
+                if shape.count(-1) > 0:
+                    shape[shape.index(-1)] = b
+                feed[input_tensor] = numpy.random.random_sample(shape)
+            output_tensor = self.sess.graph.get_tensor_by_name(tensor_name)
+            shape = self.sess.run([output_tensor], feed)[0].shape
+            shapes.append(numpy.array(shape))
+
+        compare01 = (shapes[0] == shapes[1])
+        compare12 = (shapes[1] == shapes[2])
+
+        if compare01.all() and compare12.all():
+            return shape[0].tolist()
+
+        if (compare01 == compare12).all():
+            index = numpy.argwhere(compare01 == False).flatten()
+            if index.shape[0] != 1:
+                raise Exception("There's not only one unstable dimension")
+            if index[0] != 0:
+                raise Exception("Batch size not in the first dimension")
+            shapes[0][0] = -1
+            return shapes[0].tolist()

x2paddle/op_mapper/caffe_op_mapper.py

@@ -308,7 +308,7 @@ class CaffeOpMapper(OpMapper):
             'pool_padding': pad,
             'ceil_mode': ceil_mode,
             'pool_type': string(pool_type),
-            'exclusive': True,
+            'exclusive': False,
             'global_pooling': global_pool,
             'name': string(node.layer_name)
         }

x2paddle/op_mapper/onnx_custom_layer/InstanceNormalization.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 .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
+    mean = fluid.layers.reduce_mean(inputs, dim=[2, 3], keep_dim=True)
+    var = fluid.layers.reduce_mean(fluid.layers.square(inputs - mean),
+                                   dim=[2, 3],
+                                   keep_dim=True)
+    if name is not None:
+        scale_name = name + "_scale"
+        offset_name = name + "_offset"
+    scale_param = fluid.ParamAttr(name=scale_name,
+                                  initializer=fluid.initializer.Constant(1.0),
+                                  trainable=True)
+    offset_param = fluid.ParamAttr(name=offset_name,
+                                   initializer=fluid.initializer.Constant(0.0),
+                                   trainable=True)
+    scale = fluid.layers.create_parameter(attr=scale_param,
+                                          shape=inputs.shape[1:2],
+                                          dtype="float32")
+    offset = fluid.layers.create_parameter(attr=offset_param,
+                                           shape=inputs.shape[1:2],
+                                           dtype="float32")
+
+    tmp = fluid.layers.elementwise_mul(x=(inputs - 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,
+         weights=InstanceNormalization_weights)

x2paddle/op_mapper/onnx_custom_layer/__init__.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 .register import get_registered_layers
+#custom layer import begins
+
+from . import InstanceNormalization
+#custom layer import ends
+
+custom_layers = get_registered_layers()
+
+
+def set_args(f, params):
+    """ set args for function 'f' using the parameters in node.layer.param
+    Args:
+        f (function): a python function object
+        params (object): a object contains attributes needed by f's arguments
+    Returns:
+        arg_names (list): a list of argument names
+        kwargs (dict): a dict contains needed arguments
+    """
+    argc = f.__code__.co_argcount
+    arg_list = f.__code__.co_varnames[0:argc]
+    kwargs = {}
+    for arg_name in arg_list:
+        if hasattr(params, arg_name) and params is not None:
+            kwargs[arg_name] = getattr(params, arg_name)
+    return arg_list, kwargs
+
+
+def has_layer(layer_type):
+    """ test whether this layer exists in custom layer
+    """
+    return layer_type in custom_layers
+
+
+def get_params(layer, layer_type):
+    import re
+    if layer_type.lower() == "deconvolution" or layer_type.lower(
+    ) == "convolutiondepthwise":
+        param_name = '_'.join(('convolution', 'param'))
+    elif layer_type.lower() == "normalize":
+        param_name = '_'.join(('norm', 'param'))
+    elif len(layer_type) - len(re.sub("[A-Z]", "", layer_type)) >= 2:
+        s = ''
+        tmp_name = ''
+        for i, ch in enumerate(layer_type):
+            if i == 0:
+                s += ch.lower()
+                continue
+            elif ch.isupper() and layer_type[i - 1].islower():
+                tmp_name += (s + '_')
+                s = ''
+            s += ch.lower()
+        tmp_name += s
+        param_name = '_'.join((tmp_name, 'param'))
+    else:
+        param_name = '_'.join((layer_type.lower(), 'param'))
+    return getattr(layer, param_name, None)
+
+
+def compute_output_shape(node):
+    """ compute the output shape of custom layer
+    """
+    layer_type = node.layer_type
+    assert layer_type in custom_layers, "layer[%s] not exist in custom layers" % (
+        layer_type)
+    shape_func = custom_layers[layer_type]['shape']
+    layer = node.layer
+    params = get_params(layer, layer_type)
+    arg_names, kwargs = set_args(shape_func, params)
+    input_shape = node.input_shape
+    return shape_func(input_shape, **kwargs)
+
+
+def make_custom_layer(node):
+    """ get the code which implement the custom layer function
+    """
+    layer_type = node.layer_type
+    assert layer_type in custom_layers, "layer[%s] not exist in custom layers" % (
+        layer_type)
+    layer_func = custom_layers[layer_type]['layer']
+    import inspect
+    return inspect.getsource(layer_func), layer_func
+
+
+def deal_weights(node, data=None):
+    """ deal the weights of the custom layer
+    """
+    layer_type = node.layer_type
+    weights_func = custom_layers[layer_type]['weights']
+    name = node.layer_name
+    return weights_func(name, data)

x2paddle/op_mapper/onnx_custom_layer/register.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.
+""" this module provides 'register' for registering customized layers
+"""
+
+g_custom_layers = {}
+
+
+def register(kind, shape, layer, weights):
+    """ register a custom layer or a list of custom layers
+
+    Args:
+        @kind (str or list): type name of the layer
+        @shape (function): a function to generate the shape of layer's output
+        @layer (function): a function to generate the paddle code of layer
+        @weights (function): a function to deal with weights data
+
+    Returns:
+        None
+    """
+    assert type(shape).__name__ == 'function', 'shape should be a function'
+    assert type(layer).__name__ == 'function', 'layer should be a function'
+
+    if type(kind) is str:
+        kind = [kind]
+    else:
+        assert type(
+            kind
+        ) is list, 'invalid param "kind" for register, not a list or str'
+
+    for k in kind:
+        assert type(
+            k) is str, 'invalid param "kind" for register, not a list of str'
+        assert k not in g_custom_layers, 'this type[%s] has already been registered' % (
+            k)
+        print('register layer[%s]' % (k))
+        g_custom_layers[k] = {
+            'shape': shape,
+            'layer': layer,
+            'weights': weights
+        }
+
+
+def get_registered_layers():
+    return g_custom_layers

x2paddle/op_mapper/onnx_directly_map.py

@@ -24,6 +24,7 @@ default_op_mapping_field_values['DEFAULTS'] = dict()
 default_op_mapping_field_values['INPUT_PERM'] = None
 default_op_mapping_field_values['OUTPUT_PERM'] = None
 default_op_mapping_field_values['FILL_NAME_FIELD'] = True
+
 default_op_mapping = {
     'Gather': ['gather', ['X'], ['Out'],
                dict(axis='')],
@@ -46,8 +47,44 @@ default_op_mapping = {
         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)
+    ],
+
+    #active function
+    'Relu': ['relu', ['X'], ['Out']],
+    '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'),
+        dict(alpha=1.)
+    ],
+    'Tanh': ['tanh', ['X'], ['Out']],
+    'Sigmoid': ['sigmoid', ['X'], ['Out']],
+    'Pow': ['elementwise_pow', ['X', 'Y'], ['Out'],
+            dict(),
+            dict(axis=-1)],  # TODO: pow for scalar exponent
+    'HardSigmoid': [
+        '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(axis=''),
+                dict(axis=1)],
+}
+
+activefunc_op_mapping = {
     'LeakyRelu': ['leaky_relu', ['X'], ['Out'],
-                  dict(), dict(alpha=.01)]
+                  dict(), dict(alpha=.01)],
 }
 
 default_ioa_constraint = {

x2paddle/op_mapper/tf_op_mapper.py

@@ -58,6 +58,7 @@ class TFOpMapper(OpMapper):
         'Exp': ['exp'],
         'Rsqrt': ['rsqrt'],
         'swish_f32': ['swish'],
+        'Tanh': ['tanh'],
         'LeakyRelu': ['leaky_relu', {
             'alpha': 'alpha'
         }]
@@ -188,6 +189,10 @@ class TFOpMapper(OpMapper):
             if y_shape[index] != x_shape[index]:
                 is_sub_seq = False
         if not is_sub_seq:
+            if x_shape.count(-1) > 2:
+                x_shape = self.decoder.infer_tensor_shape(x_input)
+            if y_shape.count(-1) > 2:
+                y_shape = self.decoder.infer_tensor_shape(y_input)
             x_expand_times = [1] * len(x_shape)
             y_expand_times = [1] * len(y_shape)
             x_need_expand = False
@@ -913,6 +918,12 @@ class TFOpMapper(OpMapper):
         self.add_omit_nodes(kernel.layer_name, node.layer_name)
         self.add_omit_nodes(out_shape.layer_name, node.layer_name)
 
+        if out_shape.layer_type == "Const":
+            out_shape = out_shape.value.tolist()
+        else:
+            out_shape = self.decoder.infer_shape_tensor(out_shape,
+                                                        node.out_shapes[0])
+
         in_shape = input.out_shapes[0]
         if in_shape.count(-1) > 2:
             in_shape = self.decoder.infer_tensor(input).shape
@@ -920,7 +931,7 @@ class TFOpMapper(OpMapper):
         if k_size.count(-1) > 2:
             k_size = self.decoder.infer_tensor(kernel).shape
 
-        pad_mode = node.get_attr("padding")
+        pad_mode = node.get_attr("padding").decode()
         strides = node.get_attr("strides")
         dilations = node.get_attr("dilations")
         data_format = node.get_attr("data_format").decode()
@@ -963,6 +974,22 @@ class TFOpMapper(OpMapper):
                                   output=node,
                                   param_attr=attr)
 
+        if pad_mode == "SAME":
+            if node.tf_data_format == "NHWC":
+                out_shape = [out_shape[i] for i in [0, 3, 1, 2]]
+            for i in range(4):
+                if out_shape[i] < 0:
+                    out_shape[i] = 999999
+            attr = {
+                "axes": [0, 1, 2, 3],
+                "starts": [0, 0, 0, 0],
+                "ends": out_shape
+            }
+            node.fluid_code.add_layer("slice",
+                                      inputs=node,
+                                      output=node,
+                                      param_attr=attr)
+
     def Max(self, node):
         input = self.graph.get_node(node.layer.input[0], copy=True)
         reduce_idx = self.graph.get_node(node.layer.input[1], copy=True)
@@ -1142,3 +1169,17 @@ class TFOpMapper(OpMapper):
                                       inputs=None,
                                       output=node,
                                       param_attr=attr)
+
+    def SquaredDifference(self, node):
+        x = self.graph.get_node(node.layer.input[0], copy=True)
+        y = self.graph.get_node(node.layer.input[1], copy=True)
+        inputs = {"x": x, "y": y}
+        node.fluid_code.add_layer("elementwise_sub",
+                                  inputs=inputs,
+                                  output=node,
+                                  param_attr=None)
+        inputs = {"x": node, "y": node}
+        node.fluid_code.add_layer("elementwise_mul",
+                                  inputs=inputs,
+                                  output=node,
+                                  param_attr=None)

x2paddle/optimizer/onnx_optimizer.py

@@ -14,7 +14,6 @@
 
 # TODO useless node remove
 from x2paddle.op_mapper.onnx_op_mapper import ONNXOpMapper
-from x2paddle.core.util import *
 
 
 class ONNXOptimizer(object):

x2paddle_model_zoo.md

-目前X2Paddle支持40+的TensorFlow OP，40+的Caffe Layer，覆盖了大部分CV分类模型常用的操作。我们在如下模型列表中测试了X2Paddle的转换。
-
-受限于不同框架的差异，部分模型可能会存在目前无法转换的情况，如TensorFlow中包含控制流的模型，NLP模型等。对于CV常见的模型，如若您发现无法转换或转换失败，存在较大diff等问题，欢迎通过[ISSUE反馈](https://github.com/PaddlePaddle/X2Paddle/issues/new)的方式告知我们(模型名，代码实现或模型获取方式)，我们会即时跟进：）
-
-# TensorFlow
-
-| 模型 | 代码 |
-|------|----------|
-| SqueezeNet | [code](https://github.com/tensorflow/tpu/blob/master/models/official/squeezenet/squeezenet_model.py)|
-| MobileNet_V1 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md) |
-| MobileNet_V2 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet) |
-| 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_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/blob/master/research/slim/nets/vgg.py) |
-| ResNet_V1_101 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/resnet_v1.py) |
-| ResNet_V2_101 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/resnet_v2.py) |
-
-# Caffe
+# X2Paddle模型测试库
+> 目前X2Paddle支持40+的TensorFlow OP，40+的Caffe Layer，覆盖了大部分CV分类模型常用的操作。我们在如下模型列表中测试了X2Paddle的转换。
+
+**注：** 受限于不同框架的差异，部分模型可能会存在目前无法转换的情况，如TensorFlow中包含控制流的模型，NLP模型等。对于CV常见的模型，如若您发现无法转换或转换失败，存在较大diff等问题，欢迎通过[ISSUE反馈](https://github.com/PaddlePaddle/X2Paddle/issues/new)的方式告知我们(模型名，代码实现或模型获取方式)，我们会及时跟进：）
+
+## TensorFlow
+
+| 模型 | 代码 | 备注 |
+|------|----------|------|
+| SqueezeNet | [code](https://github.com/tensorflow/tpu/blob/master/models/official/squeezenet/squeezenet_model.py)|-|
+| MobileNet_V1 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-|
+| MobileNet_V2 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-|
+| 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_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) |-|
+| ResNet_V1_101 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-|
+| ResNet_V2_101 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-|
+| UNet | [code1](https://github.com/jakeret/tf_unet )/[code2](https://github.com/lyatdawn/Unet-Tensorflow) |-|
+|MTCNN | [code](https://github.com/AITTSMD/MTCNN-Tensorflow) |-|
+|YOLO-V3| [code](https://github.com/YunYang1994/tensorflow-yolov3) | 转换需要关闭NHWC->NCHW的优化，见[文档Q2](FAQ.md) |
+|Inception_V4| [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) | - |
+|Inception_ResNet_V2| [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) | - |
+
+## Caffe
 
 | 模型 | 代码 |
 |-------|--------|
@@ -29,36 +35,24 @@
 | mNASNet | [code](https://github.com/LiJianfei06/MnasNet-caffe) |
 | MTCNN | [code](https://github.com/kpzhang93/MTCNN_face_detection_alignment/tree/master/code/codes/MTCNNv1/model) |
 
-# ONNX
+## 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|
 |-------|--------|---------|
-| 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|
-| Resnet101 | [torchvison.model.resnet101](https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py) |9|
-| Vgg11 | [torchvison.model.vgg11](https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py) |9|
-| Vgg11_bn | [torchvison.model.vgg11_bn](https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py) |9|
-| Vgg19| [torchvison.model.vgg19](https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py) |9|
-| Densenet121 | [torchvison.model.densenet121](https://github.com/pytorch/vision/blob/master/torchvision/models/densenet.py) |9|
-| Alexnet | [torchvison.model.alexnet](https://github.com/pytorch/vision/blob/master/torchvision/models/alexnet.py) |9|
-| Shufflenet | [onnx official](https://github.com/onnx/models/tree/master/vision/classification/shufflenet) |9|
-| Inception_v2 | [onnx official](https://github.com/onnx/models/tree/master/vision/classification/inception_and_googlenet/inception_v2) |9|
-| Mobilenet_v2 | [pytorch(personal practice)](https://github.com/tonylins/pytorch-mobilenet-v2) |9|
-
-目前onnx2paddle主要支持onnx operator version 9；  
-如何将torchvison或者个人开发者写的pytroch model转换成onnx model:
-```
-import torch
-import torchvision
-
-#根据不同模型调整输入的shape
-dummy_input = torch.randn(1, 3, 224, 224)
-
-#预训练后的pytorch model
-resnet18 = torchvision.models.resnet18(pretrained=True)
-
-#"resnet18.onnx"为onnx model的存储路径，1.1
-torch.onnx.export(resnet18, dummy_input, "resnet18.onnx",verbose=True)
+| 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|
+| ResNet101 | [torchvison.model.resnet101](https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py) |9|
+| VGG11 | [torchvison.model.vgg11](https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py) |9|
+| VGG11_bn | [torchvison.model.vgg11_bn](https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py) |9|
+| VGG19| [torchvison.model.vgg19](https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py) |9|
+| DenseNet121 | [torchvison.model.densenet121](https://github.com/pytorch/vision/blob/master/torchvision/models/densenet.py) |9|
+| AlexNet | [torchvison.model.alexnet](https://github.com/pytorch/vision/blob/master/torchvision/models/alexnet.py) |9|
+| ShuffleNet | [onnx official](https://github.com/onnx/models/tree/master/vision/classification/shufflenet) |9|
+| Inception_V2 | [onnx official](https://github.com/onnx/models/tree/master/vision/classification/inception_and_googlenet/inception_v2) |9|
+| MobileNet_V2 | [pytorch(personal practice)](https://github.com/tonylins/pytorch-mobilenet-v2) |9|
+| 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|
 
-```