broadcast無しのMulの変換に対応

sony · Mar 15, 2018 · b125910 · b125910
1 parent d8161c9
commit b125910
Show file tree

Hide file tree

Showing 3 changed files with 41 additions and 2 deletions.
diff --git a/python/src/nnabla/utils/converter/onnx/exporter.py b/python/src/nnabla/utils/converter/onnx/exporter.py
@@ -35,7 +35,8 @@
     "Add2": "Add",
     # optype that gets converted
     "Identity": "Dropout",
-    "Affine": "Gemm"
+    "Affine": "Gemm",
+    "Mul2": "Mul"
 }
 
 

diff --git a/python/src/nnabla/utils/converter/onnx/reader.py b/python/src/nnabla/utils/converter/onnx/reader.py
@@ -44,7 +44,8 @@
     "AveragePool": "AveragePooling",
     "Sum": "Add2",
     "Gemm": "Affine",
-    "Add": "Add2"
+    "Add": "Add2",
+    "Mul": "Mul2"
 }
 
 
@@ -412,6 +413,35 @@ def convert_to_functions(node, base_name, initializers, func_counter):
                 raise ValueError("broadcasting is currently not supported for {}".format(node.op_type))
                 # Add2 broadcasts by default so we do nothing here
                 #pass
+    elif node.op_type == "Mul":
+        # We need the input buffer's dimension information here
+        # in order to reshape the bias vector correctly.
+        # Therefore we cannot support broadcasting unless we get an operator like ReshapeTo
+        # which allows reshaping without shape specification.
+        reshaped_postfix = "_reshaped"
+        input = node.input[:]
+        for attr in node.attribute:
+            if attr.name == "axis":
+                pass
+                ## Reshape the input bias so it fits
+                ## the specifed axis's broadcasted shape
+                #rin = node.input[1]
+                #rout = rin+reshaped_postfix
+                #rs = nnabla_pb2.Function()
+                #rs.type = "Reshape"
+                #set_function_name(rs, node.name, base_name, func_counter)
+                #rs.input.extend([rin])
+                #rs.output.extend([rout])
+                ## Calculate the reshaped size for the bias.
+                ## We calculate this from the input buffer's dimension and
+                ## the specified axis.
+                #rp = rs.reshape_param
+                #rp.shape.dim.extend(reshaped)
+                #input[1] = rout  # rewire input to reshaped input
+            elif attr.name == "broadcast":
+                raise ValueError("broadcasting is currently not supported for {}".format(node.op_type))
+                # Mul2 broadcasts by default so we do nothing here
+                #pass
     func_list.append(func)
     return func_list
 

diff --git a/python/test/utils/test_conversion.py b/python/test/utils/test_conversion.py
@@ -240,6 +240,14 @@ def test_nnp_onnx_conversion_add_no_broadcast(tmpdir, nnp_fixture):
     convert_nnp_to_onnx_and_compare(
         tmpdir, TEST_DATA_DIR, "add_no_broadcast.nnp", "add_no_broadcast.onnx", "out_data_1", "exec_0")
 
+def test_onnx_nnp_conversion_mul_no_broadcast(tmpdir, nnp_fixture):
+    convert_onnx_to_nnp_and_compare(
+        tmpdir, TEST_DATA_DIR, "mul_no_broadcast.onnx", "mul_no_broadcast.nnp", "out_data_1", "exec_0")
+
+def test_nnp_onnx_conversion_mul_no_broadcast(tmpdir, nnp_fixture):
+    convert_nnp_to_onnx_and_compare(
+        tmpdir, TEST_DATA_DIR, "mul_no_broadcast.nnp", "mul_no_broadcast.onnx", "out_data_1", "exec_0")
+
 def test_onnx_nnp_conversion_squeezenet(tmpdir, nnp_fixture):
     onnx_dir = TEST_DATA_DIR
     onnx_name = "squeezenet.onnx"