Merge pull request #2 from PaddlePaddle/develop

pull

README.md

@@ -9,13 +9,13 @@ 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 >= 3.5  
+python == 2.7 | python >= 3.5  
 paddlepaddle >= 1.5.0  
 
 **按需安装以下依赖**  
 tensorflow ： tensorflow == 1.14.0  
 caffe ： 无  
-onnx ： onnx == 1.5.0  pytorch == 1.1.0
+onnx ： onnx == 1.5.0  onnxruntime == 0.4.0
 
 ## 安装
 ### 安装方式一（推荐）

setup.py

@@ -23,4 +23,9 @@ 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',
+            'onnx_infer=x2paddle.onnx_infer:main'
+        ]
+    })

x2paddle/__init__.py

-__version__ = "0.4.5"
+__version__ = "0.5.0"

x2paddle/convert.py

@@ -154,7 +154,7 @@ def onnx2paddle(model_path, save_dir):
     model = ONNXDecoder(model_path)
 
     from x2paddle.op_mapper.onnx_op_mapper import ONNXOpMapper
-    mapper = ONNXOpMapper(model)
+    mapper = ONNXOpMapper(model, save_dir)
 
     from x2paddle.optimizer.onnx_optimizer import ONNXOptimizer
     optimizer = ONNXOptimizer(mapper)

x2paddle/decoder/onnx_decoder.py

@@ -17,7 +17,6 @@ from x2paddle.core.fluid_code import FluidCode
 from onnx.checker import ValidationError
 from onnx.checker import check_model
 from onnx.utils import polish_model
-from onnx.version_converter import convert_version
 from onnx import helper
 from onnx.helper import get_attribute_value, make_attribute
 from onnx.shape_inference import infer_shapes
@@ -26,9 +25,11 @@ from onnx.numpy_helper import to_array
 from onnx import AttributeProto, TensorProto, GraphProto
 from collections import OrderedDict as Dict
 import onnx
+from onnx.helper import ValueInfoProto
 import numpy as np
 from copy import deepcopy
 import logging as _logging
+import os
 
 default_op_domain = 'ai.onnx'
 _logger = _logging.getLogger(__name__)
@@ -47,6 +48,7 @@ class ONNXGraphNode(GraphNode):
         self.weight_inputs = list()
         self.out_shapes = list()
         self.dtype = None
+        self.which_child = {}
 
     def get_attr_map(self):
         """
@@ -60,10 +62,9 @@ class ONNXGraphNode(GraphNode):
     @property
     def value(self):
         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]
+        return self.attr_map['value']
 
     def get_attribute_value2(self, attr):
         """
@@ -105,29 +106,39 @@ class ONNXGraphDataNode(GraphNode):
         self.fluid_code = FluidCode()
         self.weight = None
         self.embeded_as = None
+        self.which_child = {}
 
     @property
     def out_shapes(self):
+        if isinstance(self.layer, ValueInfoProto):
             values = self.layer.type.tensor_type.shape.dim
             out_shapes = list()
             out_shapes.append([dim.dim_value for dim in values])
             return out_shapes
+        else:
+            values = self.layer.dims
+            out_shapes = list()
+            out_shapes.append(values)
+            return out_shapes
 
     @property
     def dtype(self):
+        if isinstance(self.layer, ValueInfoProto):
             dtype = self.layer.type.tensor_type.elem_type
             return TENSOR_TYPE_TO_NP_TYPE[dtype]
+        else:
+            dtype = self.layer.data_type
+            return TENSOR_TYPE_TO_NP_TYPE[dtype]
 
 
 class ONNXGraph(Graph):
-    def __init__(self, graph, onnx_model):
-        super(ONNXGraph, self).__init__(graph)
+    def __init__(self, onnx_model):
+        super(ONNXGraph, self).__init__(onnx_model.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.value_infos = self.inferred_model_value_info(self.model)
         self.results_of_inference = dict()
 
     def get_inner_nodes(self):
@@ -165,22 +176,9 @@ 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:
             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:
@@ -191,20 +189,40 @@ class ONNXGraph(Graph):
                     is_global_input=is_place_holder)
 
         #set data node's weight
-        for name, weight in self.graph_weights(self.model):
+        for initializer in self.model.initializer:
+            name = initializer.name
+            weight = to_array(initializer)
             if name in self.node_map:
                 if isinstance(self.node_map[name], ONNXGraphDataNode):
                     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].weight = weight
+                self.node_map[name].embeded_as = []
 
         #generate connection between nodes for topo
         for layer_name, node in self.node_map.items():
             if isinstance(node, ONNXGraphNode):
                 for idx, in_node in enumerate(node.layer.input):
                     if in_node not in self.node_map:
+                        flag = 0
+                        for nd in self.model.node:
+                            for idx, opt in enumerate(nd.output):
+                                if opt == in_node:
+                                    self.connect(nd.name, layer_name)
+                                    flag = 1
+                                    node.which_child[nd.name] = idx
+                                    self.node_map[nd.name].index = 0
+                                    break
+                            if flag == 1:
+                                break
+                        if flag == 0:
                             raise Exception(
-                            'input[{}] of node[{}] does not exist in node_map'.
-                            format(in_node, layer_name))
+                                'input[{}] of node[{}] does not exist in node_map'
+                                .format(in_node, layer_name))
                     else:
                         self.connect(in_node, layer_name)
         #generate topo
@@ -212,13 +230,16 @@ class ONNXGraph(Graph):
 
         self.input_nodes = self.place_holder_nodes
 
-    def get_nodes(self, names, copy=False):
-        """
-        get nodes by more than one name
-        """
-        nodes = []
-        for name in names:
-            nodes.add(self.get_node(name, copy=copy))
+    def get_input_node(self, node, idx=0, copy=False):
+        if len(node.which_child) == 0:
+            ipt_node = super(ONNXGraph, self).get_node(node.inputs[idx], copy)
+            return ipt_node
+
+        else:
+            ipt_node = super(ONNXGraph, self).get_node(node.inputs[idx], copy)
+            if ipt_node.layer_name in node.which_child:
+                ipt_node.index = node.which_child[ipt_node.layer_name]
+            return ipt_node
 
     def graph_weights(self, graph):
         """
@@ -270,50 +291,6 @@ class ONNXGraph(Graph):
             }
         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):
@@ -334,8 +311,8 @@ class ONNXDecoder(object):
         self.standardize_variable_name(model.graph)
 
         self.model = model
-        graph_def = model.graph
-        self.onnx_graph = ONNXGraph(graph_def, model)
+        graph = model.graph
+        self.onnx_graph = ONNXGraph(model)
         self.onnx_graph.build()
 
     def build_value_refs(self, nodes):
@@ -476,7 +453,7 @@ class ONNXDecoder(object):
 
         if name == '':
             raise ValueError('name should not be empty')
-        for s in ' .*?\\/-:':  #
+        for s in ' .*?\\/-:':
             name = name.replace(s, '_')
         return '_' + name
 
@@ -499,46 +476,3 @@ 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])
-
-    def split_model(self, model, outputs=None):
-        """
-        Takes a model and changes its outputs.
-        """
-        if outputs is None:
-            raise RuntimeError("outputs is None")
-        if outputs == model.graph.output[0].name:
-            return model
-        nodes = model.graph.node
-        keep_nodes = []
-
-        # all the nodes we need to keep.
-        for node in nodes:
-            if outputs in node.output:
-                keep_nodes.append(node)
-                break
-            keep_nodes.append(node)
-
-        infer_shapes = onnx.shape_inference.infer_shapes(model)
-
-        var_out = []
-        for value_info in infer_shapes.graph.value_info:
-            if value_info.name == outputs:
-                var_out.append(value_info)
-                break
-
-        graph = helper.make_graph(keep_nodes, model.graph.name,
-                                  model.graph.input, var_out,
-                                  model.graph.initializer)
-
-        onnx_model = helper.make_model(graph)
-        onnx_model.ir_version = model.ir_version
-        onnx_model.producer_name = model.producer_name
-        onnx_model.producer_version = model.producer_version
-        onnx_model.domain = model.domain
-        onnx_model.model_version = model.model_version
-        onnx_model.doc_string = model.doc_string
-
-        if len(onnx_model.graph.input) != len(model.graph.input):
-            raise RuntimeError("Input mismatch {} != {}".format(
-                len(onnx_model.input), len(model.input)))
-        return onnx_model

x2paddle/onnx_infer.py

+import os
+import sys
+import numpy as np
+import onnx
+import json
+import argparse
+from six import text_type as _text_type
+
+
+def arg_parser():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--save_dir",
+                        "-s",
+                        type=_text_type,
+                        default=None,
+                        help="define save_dir")
+    return parser
+
+
+def main():
+    try:
+        import onnxruntime as rt
+        version = rt.__version__
+        if version != '0.4.0':
+            print("onnxruntime==0.4.0 is required")
+            return
+    except:
+        print(
+            "onnxruntime is not installed, use \"pip install onnxruntime==0.4.0\"."
+        )
+        return
+    parser = arg_parser()
+    args = parser.parse_args()
+
+    save_dir = args.save_dir
+    model_dir = os.path.join(save_dir, 'onnx_model_infer.onnx')
+    data_dir = os.path.join(save_dir, 'input_data.npy')
+
+    model = onnx.load(model_dir)
+    sess = rt.InferenceSession(model_dir)
+
+    inputs = np.load(data_dir, allow_pickle=True)
+    data_dir
+    inputs_dict = {}
+    for i, ipt in enumerate(inputs):
+        inputs_dict[sess.get_inputs()[i].name] = ipt
+
+    res = sess.run(None, input_feed=inputs_dict)
+    for idx, value_info in enumerate(model.graph.output):
+        np.save(os.path.join(save_dir, value_info.name), res[idx])
+
+
+if __name__ == "__main__":
+    main()

x2paddle/op_mapper/caffe_op_mapper.py

@@ -954,6 +954,13 @@ class CaffeOpMapper(OpMapper):
         inputs_node = []
         for i in range(len(node.inputs)):
             input = self.graph.get_bottom_node(node, idx=i, copy=True)
+            if i == 1 and op == 'DetectionOutput':
+                input = self.graph.get_bottom_node(node, idx=i, copy=True)
+                print(input.layer_type)
+                while input is not None and input.layer_type != 'Softmax':
+                    input = self.graph.get_bottom_node(input, idx=0, copy=True)
+                assert input is not None, 'This kind of DetectionOutput is not supported!'
+                input = self.graph.get_bottom_node(input, idx=0, copy=True)
             inputs_node.append(input)
         node.fluid_code.add_layer(func.__code__.co_name,
                                   inputs=inputs_node,

x2paddle/op_mapper/onnx_custom_layer/InstanceNormalization.py

@@ -22,8 +22,9 @@ def InstanceNormalization_shape(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),
+    input_ = inputs[0]
+    mean = fluid.layers.reduce_mean(input_, dim=[2, 3], keep_dim=True)
+    var = fluid.layers.reduce_mean(fluid.layers.square(input_ - mean),
                                    dim=[2, 3],
                                    keep_dim=True)
     if name is not None:
@@ -36,13 +37,13 @@ def InstanceNormalization_layer(inputs, name=None):
                                    initializer=fluid.initializer.Constant(0.0),
                                    trainable=True)
     scale = fluid.layers.create_parameter(attr=scale_param,
-                                          shape=inputs.shape[1:2],
+                                          shape=input_.shape[1:2],
                                           dtype="float32")
     offset = fluid.layers.create_parameter(attr=offset_param,
-                                           shape=inputs.shape[1:2],
+                                           shape=input_.shape[1:2],
                                            dtype="float32")
 
-    tmp = fluid.layers.elementwise_mul(x=(inputs - mean), y=scale, axis=1)
+    tmp = fluid.layers.elementwise_mul(x=(input_ - mean), y=scale, axis=1)
     tmp = tmp / fluid.layers.sqrt(var + epsilon)
     tmp = fluid.layers.elementwise_add(tmp, offset, axis=1)
     return tmp
@@ -56,4 +57,5 @@ def InstanceNormalization_weights(name, data=None):
 register(kind='InstanceNormalization',
          shape=InstanceNormalization_shape,
          layer=InstanceNormalization_layer,
+         child_func=None,
          weights=InstanceNormalization_weights)

x2paddle/op_mapper/onnx_custom_layer/__init__.py

@@ -95,6 +95,17 @@ def make_custom_layer(node):
     return inspect.getsource(layer_func), layer_func
 
 
+def make_custom_child_func(node):
+    """ get the code which implement the custom layer function
+    """
+    layer_type = node.layer_type
+    child_func = custom_layers[layer_type]['child_func']
+    if child_func is None:
+        return None, child_func
+    import inspect
+    return inspect.getsource(child_func), child_func
+
+
 def deal_weights(node, data=None):
     """ deal the weights of the custom layer
     """

x2paddle/op_mapper/onnx_custom_layer/register.py

@@ -17,7 +17,7 @@
 g_custom_layers = {}
 
 
-def register(kind, shape, layer, weights):
+def register(kind, shape, layer, child_func, weights):
     """ register a custom layer or a list of custom layers
 
     Args:
@@ -48,6 +48,7 @@ def register(kind, shape, layer, weights):
         g_custom_layers[k] = {
             'shape': shape,
             'layer': layer,
+            'child_func': child_func,
             'weights': weights
         }
 

x2paddle/op_mapper/onnx_directly_map.py

@@ -29,9 +29,6 @@ default_op_mapping = {
     'Gather': ['gather', ['X'], ['Out'],
                dict(axis='')],
     'Shape': ['shape', ['X'], ['Out']],
-    'Mul': ['elementwise_mul', ['X', 'Y'], ['Out'],
-            dict(),
-            dict(axis=-1)],
     'Clip': [
         'clip', ['X'], ['Out'],
         dict(),
@@ -42,6 +39,7 @@ default_op_mapping = {
                              dtype=_np.uint8).view(_np.float32)),
         )
     ],
+    'Ceil': ['ceil', ['X'], ['Out']],
     'ReduceMean': [
         'reduce_mean', ['X'], ['Out'],
         dict(axes='dim', keepdims='keep_dim'),
@@ -52,7 +50,11 @@ default_op_mapping = {
         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)
+    ],
     #active function
     'Relu': ['relu', ['X'], ['Out']],
     'LeakyRelu': ['leaky_relu', ['X'], ['Out'],
@@ -66,9 +68,6 @@ default_op_mapping = {
     ],
     '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'),
@@ -78,8 +77,8 @@ default_op_mapping = {
     'Softplus': ['softplus', ['X'], ['Out']],
     'Exp': ['exp', ['X'], ['Out']],
     'Softmax': ['softmax', ['X'], ['Out'],
-                dict(axis=''),
-                dict(axis=1)],
+                dict(), dict(axis=1)],
+    'Sqrt': ['sqrt', ['X'], ['Out']],
 }
 
 activefunc_op_mapping = {

x2paddle/op_mapper/tf_op_mapper.py

@@ -785,6 +785,9 @@ class TFOpMapper(OpMapper):
         start = self.graph.get_node(node.layer.input[0], copy=True)
         limit = self.graph.get_node(node.layer.input[1], copy=True)
         delta = self.graph.get_node(node.layer.input[2], copy=True)
+        self.add_omit_nodes(start.layer_name, node.layer_name)
+        self.add_omit_nodes(limit.layer_name, node.layer_name)
+        self.add_omit_nodes(delta.layer_name, node.layer_name)
         if start.layer_type == "Const":
             start = start.value
         else:
@@ -797,9 +800,6 @@ class TFOpMapper(OpMapper):
             delta = delta.value
         else:
             delta = self.decoder.infer_tensor(delta)
-        self.add_omit_nodes(start.layer_name, node.layer_name)
-        self.add_omit_nodes(limit.layer_name, node.layer_name)
-        self.add_omit_nodes(delta.layer_name, node.layer_name)
 
         inputs = {"start": start, "end": limit, "step": delta}
         attr = {"dtype": string(node.dtype)}

x2paddle/op_mapper/tf_op_mapper_nhwc.py

@@ -74,7 +74,7 @@ class TFOpMapperNHWC(OpMapper):
         unsupported_ops = set()
         sys.stderr.write("Total nodes: {}\n".format(len(self.graph.topo_sort)))
         for i, node_name in enumerate(self.graph.topo_sort):
-            sys.stderr.write("\rConverting node {} ...     ".format(i))
+            sys.stderr.write("\rConverting node {} ...     ".format(i + 1))
             node = self.graph.get_node(node_name)
             op = node.layer_type
             if op in self.directly_map_ops:
@@ -99,7 +99,7 @@ class TFOpMapperNHWC(OpMapper):
             for op in unsupported_ops:
                 print("========== {} ============".format(op))
             sys.exit(-1)
-        sys.stderr.write("\nDone\n")
+        sys.stderr.write("\nDone!\n")
 
     def add_omit_nodes(self, in_node_name, out_node_name):
         in_node = self.graph.get_node(in_node_name)