Seluのサポートを追加し、未対応のattributeが来たときのエラー処理も追加

python/src/nnabla/utils/converter/onnx/ONNXOpCoverage.md

@@ -83,7 +83,7 @@ that indicates if each operator can be converted to NNP.
 |ReduceSumSquare|:black_heart:||
 |Relu|:green_heart:||
 |Reshape|:yellow_heart:|implementing|
-|Selu|:black_heart:||
+|Selu|:green_heart:||
 |Sigmoid|:green_heart:||
 |Slice|:black_heart:||
 |Softmax|:yellow_heart:|Supporting 2D input only|

python/src/nnabla/utils/converter/onnx/exporter.py

@@ -41,6 +41,7 @@
     "BatchMatmul": "MatMul",
     "LogicalNot": "Not",
     "ELU": "Elu",
+    "SELU": "Selu",
     # optype that gets converted
     "Identity": "Dropout",
     "Affine": "Gemm",
@@ -199,6 +200,11 @@ def convert_to_nodes(func, variables, input_types, output_types):
         # Store the input/output tensor's name and convert it to boolean
         input_types[n.input[0]] = TensorProto.BOOL
         output_types[n.output[0]] = TensorProto.BOOL
+    elif func.type == "SELU":
+        sp = func.selu_param
+        a = onnx.helper.make_attribute("alpha", sp.alpha)
+        g = onnx.helper.make_attribute("gamma", sp.scale)
+        n.attribute.extend([a, g])
     nl.append(n)
     return nl
 

python/src/nnabla/utils/converter/onnx/reader.py

@@ -56,6 +56,7 @@
     "LeakyRelu": "LeakyReLU",
     "Not": "LogicalNot",
     "Elu": "ELU",
+    "Selu": "SELU",
     # Constant does not get converted to a function
     # but we list it here so we can accept it
     "Constant": ""
@@ -456,6 +457,9 @@ def convert_to_functions(pb, network, node, base_name, initializers,
                 raise ValueError("broadcasting is currently not supported for {}".format(node.op_type))
                 # Add2 broadcasts by default so we do nothing here
                 #pass
+            else:
+                raise ValueError("Unsupported attribute {} was specified at {}"
+                                 .format(attr.name, node.op_type))
         func_list.append(func)
     elif node.op_type == "Mul":
         # We need the input buffer's dimension information here
@@ -486,6 +490,9 @@ def convert_to_functions(pb, network, node, base_name, initializers,
                 raise ValueError("broadcasting is currently not supported for {}".format(node.op_type))
                 # Mul2 broadcasts by default so we do nothing here
                 #pass
+            else:
+                raise ValueError("Unsupported attribute {} was specified at {}"
+                                 .format(attr.name, node.op_type))
         func_list.append(func)
     elif node.op_type == "Constant":
         # Convert a Constant node as an input parameter and not a function
@@ -510,6 +517,9 @@ def convert_to_functions(pb, network, node, base_name, initializers,
                 v.initializer.type = "Constant"
                 v.initializer.multiplier = 1.0
                 param_list.append(v)
+            else:
+                raise ValueError("Unsupported attribute {} was specified at {}"
+                                 .format(attr.name, node.op_type))
         # We do not add any function to the list here
         # since the node is converted as a parameter
     elif node.op_type == "Reshape":
@@ -522,6 +532,9 @@ def convert_to_functions(pb, network, node, base_name, initializers,
                     raise ValueError("Only INTS is supported for shape in {} op_type".format(node.op_type))
                 rp.shape.dim.extend(attr.ints)
                 shape_found = True
+            else:
+                raise ValueError("Unsupported attribute {} was specified at {}"
+                                 .format(attr.name, node.op_type))
         if len(func.input) == 2:
             # Shape comes as input for Reshape-5.
             # NNabla reshape excepts a single input (data),
@@ -555,6 +568,9 @@ def convert_to_functions(pb, network, node, base_name, initializers,
                 if attr.type != AttributeProto.INTS:
                     raise ValueError("Only INTS is supported for perm in {} op_type".format(node.op_type))
                 tp.axes.extend(attr.ints)
+            else:
+                raise ValueError("Unsupported attribute {} was specified at {}"
+                                 .format(attr.name, node.op_type))
         func_list.append(func)
     elif node.op_type == "LeakyRelu":
         lrp = func.leaky_relu_param
@@ -564,6 +580,9 @@ def convert_to_functions(pb, network, node, base_name, initializers,
                 if attr.type != AttributeProto.FLOAT:
                     raise ValueError("Only FLOAT is supported for alpha in {} op_type".format(node.op_type))
                 lrp.alpha = attr.f
+            else:
+                raise ValueError("Unsupported attribute {} was specified at {}"
+                                 .format(attr.name, node.op_type))
         func_list.append(func)
     elif node.op_type == "Elu":
         ep = func.elu_param
@@ -573,6 +592,26 @@ def convert_to_functions(pb, network, node, base_name, initializers,
                 if attr.type != AttributeProto.FLOAT:
                     raise ValueError("Only FLOAT is supported for alpha in {} op_type".format(node.op_type))
                 ep.alpha = attr.f
+            else:
+                raise ValueError("Unsupported attribute {} was specified at {}"
+                                 .format(attr.name, node.op_type))
+        func_list.append(func)
+    elif node.op_type == "Selu":
+        sp = func.selu_param
+        sp.alpha = 1.6732  # Default value for ONNX
+        sp.scale = 1.0507
+        for attr in node.attribute:
+            if attr.name == "alpha":
+                if attr.type != AttributeProto.FLOAT:
+                    raise ValueError("Only FLOAT is supported for alpha in {} op_type".format(node.op_type))
+                sp.alpha = attr.f
+            elif attr.name == "gamma":
+                if attr.type != AttributeProto.FLOAT:
+                    raise ValueError("Only FLOAT is supported for gamma in {} op_type".format(node.op_type))
+                sp.scale = attr.f
+            else:
+                raise ValueError("Unsupported attribute {} was specified at {}"
+                                 .format(attr.name, node.op_type))
         func_list.append(func)
     else:
         # Simply add the function for all other conversions

python/test/utils/conversion/test_conversion.py

@@ -401,6 +401,17 @@ def test_nnp_onnx_conversion_elu(tmpdir, nnp_fixture):
     convert_nnp_to_onnx_and_compare(
         tmpdir, TEST_DATA_DIR, "elu.nnp", "elu.onnx", "out_data_1", "exec_0")
 
+
+def test_onnx_nnp_conversion_selu(tmpdir, nnp_fixture):
+    convert_onnx_to_nnp_and_compare(
+        tmpdir, TEST_DATA_DIR, "selu.onnx", "selu.nnp", "out_data_1", "exec_0")
+
+
+def test_nnp_onnx_conversion_selu(tmpdir, nnp_fixture):
+    convert_nnp_to_onnx_and_compare(
+        tmpdir, TEST_DATA_DIR, "selu.nnp", "selu.onnx", "out_data_1", "exec_0")
+
+
 def test_onnx_nnp_conversion_squeezenet(tmpdir, nnp_fixture):
     img = np.random.rand(1, 3, 224, 224).astype(np.float32)
     convert_onnx_to_nnp_and_compare(