padsの仕様を少し勘違いしていたので修正

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

@@ -107,6 +107,11 @@ def set_kernel_parameter(node, kp):
             if attr.type != AttributeProto.INTS:
                 raise ValueError("Only INTS are supported for pads in {}"
                                  .format(node.op_type))
+            if len(attr.ints) > 4:
+                # pads with more than 4 (NCHW) dimension means
+                # it has a start and end specified.
+                # NNabla does not support different padding for start and end.
+                raise ValueError("NNabla does not support different padding for start and end of each axis")
             pads.extend(attr.ints)
             dims.append(len(pads))
         elif attr.name == "kernel_shape":
@@ -119,20 +124,21 @@ def set_kernel_parameter(node, kp):
             raise ValueError("Unsupported attribute {} was specified at {}"
                              .format(attr.name, node.op_type))
     # NNabla requires for the dimensions of strides, pads, kernels to match.
-    # We align the dimensions for all three attributes to the shortest one
+    # We align the dimensions for all three attributes to the shortest one.
+    # We use the dimensions from the end (hence the negative dim).
     dim = min(dims)
     if strides:
-        kp.stride.dim.extend(strides[:dim])
+        kp.stride.dim.extend(strides[-dim:])
     if pads:
-        kp.pad.dim.extend(pads[:dim])
+        kp.pad.dim.extend(pads[-dim:])
     else:
         # In case we don't have padding set,
         # we set zero padding just in case NNabla does not set the
         # default padding values correctly (such as in AveragePooling).
         # This code should not be needed if NNabla handles default values correctly.
         kp.pad.dim.extend([0]*dim)
     if kernel:
-        kp.kernel.dim.extend(kernel[:dim])
+        kp.kernel.dim.extend(kernel[-dim:])
 
 
 def update_function_counter(func_type, func_counter, count):

python/test/utils/conversion/test_conversion.py

@@ -29,6 +29,10 @@
 
 TEST_DATA_DIR = "nnabla-sample-data/conversion_data"
 
+def print_buffer_shape(net):
+    for k, v in net.functions.items():
+        out = v.outputs[0]
+        print(out.name, net.variables[out.name].variable_instance.shape)
 
 def run_executor(nn_net, exec_name):
     """Run specified executor and return its network"""
@@ -51,10 +55,14 @@ def convert_onnx_to_nnp_and_compare(
         if show_onnx:
             print(model)
         c2out = None
+        rep = oc2.prepare(model)
         if type(in_img) is np.ndarray:
-            c2out = oc2.run_model(model, [in_img])
+            c2out = rep.run([in_img])
         else:
-            c2out = oc2.run_model(model, [])
+            c2out = rep.run([])
+        #for k in rep.workspace.Blobs():
+        #    v = rep.workspace.FetchBlob(k)
+        #    print(k, v.shape)
         backend_out = c2out[out_name]
     elif backend == "cntk":
         n = cntkf.Function.load(path, format=cntk.ModelFormat.ONNX)
@@ -80,13 +88,14 @@ def convert_onnx_to_nnp_and_compare(
     p = os.path.join(str(nnpdir), nnp_name)
     nnpex.export_nnp(p)
     # read exported nnp and run network
-    # pdb.set_trace()
     nn_net = nnload.load([p])
     if type(in_img) is np.ndarray:
         net = nn_net.executors[exec_name].network
         in_data = net.variables[in_name]
         in_data.variable_instance.d = in_img
+    # pdb.set_trace()
     exe = run_executor(nn_net, exec_name)
+    #print_buffer_shape(exe)
     # in_data = exe.variables["in_data_0"]
     # print(in_data.variable_instance.d)
     nnout = exe.variables[out_name].variable_instance.d