Merge pull request #899 from wjj19950828/Add_onnx_tests

Add input_shape_dict and rm raw_input

README.md

@@ -115,6 +115,7 @@ x2paddle --framework=caffe --prototxt=deploy.prototxt --weight=deploy.caffemodel
 | --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/onnx时，该参数指定tensorflow的pb模型文件或onnx模型路径 |
+| --input_shape_dict   | **[可选]** For ONNX, 定义ONNX模型输入大小 |
 | --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 |
 | --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](./docs/inference_model_convertor/FAQ.md) |
 | --enable_code_optim  | **[可选]** For PyTorch, 是否对生成代码进行优化，默认为False |

x2paddle/convert.py

@@ -73,6 +73,13 @@ def arg_parser():
         action="store_true",
         default=False,
         help="define input shape for tf model")
+    parser.add_argument(
+       "--input_shape_dict",
+       "-isd",
+       type=_text_type,
+       default=None,
+       help="define input shapes, e.g --input_shape_dict=\"{'image':[1, 3, 608, 608]}\" or" \
+       "--input_shape_dict=\"{'image':[1, 3, 608, 608], 'im_shape': [1, 2], 'scale_factor': [1, 2]}\"")
     parser.add_argument(
         "--convert_torch_project",
         "-tp",
@@ -265,6 +272,7 @@ def caffe2paddle(proto_file,
 
 def onnx2paddle(model_path,
                 save_dir,
+                input_shape_dict=None,
                 convert_to_lite=False,
                 lite_valid_places="arm",
                 lite_model_type="naive_buffer",
@@ -292,7 +300,7 @@ def onnx2paddle(model_path,
 
     from x2paddle.decoder.onnx_decoder import ONNXDecoder
     from x2paddle.op_mapper.onnx2paddle.onnx_op_mapper import ONNXOpMapper
-    model = ONNXDecoder(model_path, enable_onnx_checker)
+    model = ONNXDecoder(model_path, input_shape_dict, enable_onnx_checker)
     mapper = ONNXOpMapper(model)
     mapper.paddle_graph.build()
     logging.info("Model optimizing ...")
@@ -481,6 +489,7 @@ def main():
             onnx2paddle(
                 args.model,
                 args.save_dir,
+                input_shape_dict=args.input_shape_dict,
                 convert_to_lite=args.to_lite,
                 lite_valid_places=args.lite_valid_places,
                 lite_model_type=args.lite_model_type,

x2paddle/decoder/onnx_decoder.py

@@ -173,9 +173,12 @@ def dtype(self):
 
 
 class ONNXGraph(Graph):
-    def __init__(self, onnx_model):
+    def __init__(self, onnx_model, input_shape_dict):
         super(ONNXGraph, self).__init__(onnx_model)
         self.fixed_input_shape = {}
+        if input_shape_dict is not None:
+            for k, v in eval(input_shape_dict).items():
+                self.fixed_input_shape["x2paddle_" + k] = v
         self.initializer = {}
         self.place_holder_nodes = list()
         self.value_infos = {}
@@ -216,45 +219,13 @@ def get_symbolic_shape(self, dims):
                 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 NameError:
-                            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()
         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):
@@ -416,7 +387,7 @@ def allocate_shapes(self):
 
 
 class ONNXDecoder(object):
-    def __init__(self, onnx_model, enable_onnx_checker):
+    def __init__(self, onnx_model, input_shape_dict, enable_onnx_checker):
         onnx_model = onnx.load(onnx_model)
         print('model ir_version: {}, op version: {}'.format(
             onnx_model.ir_version, onnx_model.opset_import[0].version))
@@ -427,7 +398,7 @@ def __init__(self, onnx_model, enable_onnx_checker):
 
         onnx_model = self.optimize_model_skip_op(onnx_model)
         onnx_model = self.optimize_node_name(onnx_model)
-        self.graph = ONNXGraph(onnx_model)
+        self.graph = ONNXGraph(onnx_model, input_shape_dict)
 
     def build_value_refs(self, nodes):
         """