Merge branch 'onnx' into 'master'

Fix ONNX padding and add transpose half

See merge request deep-computing/mace!1271

mace/ops/transpose.cc

@@ -65,6 +65,8 @@ class TransposeOp : public Operation {
 void RegisterTranspose(OpRegistry *op_registry) {
   MACE_REGISTER_OP(op_registry, "Transpose", TransposeOp,
                    DeviceType::CPU, float);
+  MACE_REGISTER_OP(op_registry, "Transpose", TransposeOp,
+                   DeviceType::CPU, half);
   MACE_REGISTER_BF16_OP(op_registry, "Transpose", TransposeOp,
                         DeviceType::CPU);
 }

tools/python/transform/onnx_converter.py

@@ -1236,6 +1236,8 @@ class OnnxConverter(base_converter.ConverterInterface):
             paddings_arg = op.arg.add()
             paddings_arg.name = MaceKeyword.mace_paddings_str
             paddings_value = node.attrs['pads']
+            paddings_value = np.asarray(paddings_value).reshape(
+                (2, -1)).transpose().reshape(-1).tolist()
             paddings_arg.ints.extend(paddings_value)
         if 'value' in node.attrs:
             constant_value_arg = op.arg.add()