Merge branch 'master' into 'master'

Support inception_v3 neon

See merge request !345

mace/kernels/arm/conv_2d.cc

@@ -159,7 +159,6 @@ void Conv2dFunctor<DeviceType::NEON, float>::operator()(const Tensor *input,
   auto filter_data = filter->data<float>();
   auto bias_data = bias == nullptr ? nullptr : bias->data<float>();
   auto output_data = output->mutable_data<float>();
-  memset(output_data, 0, sizeof(float) * batch * channels * height * width);
 
   if (USE_WINOGRAD && filter_h == 3 && filter_w == 3 && stride_h == 1
     && stride_w == 1
@@ -301,6 +300,7 @@ void Conv2dFunctor<DeviceType::NEON, float>::operator()(const Tensor *input,
     std::vector<index_t> extra_output_shape
       {batch, channels, extra_output_height, extra_output_width};
     padded_output_.Resize(extra_output_shape);
+    padded_output_.Clear();
     pad_output_ptr = &padded_output_;
   }
   float *pad_output_data = pad_output_ptr->mutable_data<float>();

mace/ops/concat.cc

@@ -13,7 +13,6 @@ void Register_Concat(OperatorRegistry *op_registry) {
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ConcatOp<DeviceType::CPU, float>);
-
   REGISTER_OPERATOR(op_registry, OpKeyBuilder("Concat")
                                      .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
@@ -24,6 +23,11 @@ void Register_Concat(OperatorRegistry *op_registry) {
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     ConcatOp<DeviceType::OPENCL, half>);
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Concat")
+                                     .Device(DeviceType::NEON)
+                                     .TypeConstraint<float>("T")
+                                     .Build(),
+                    ConcatOp<DeviceType::NEON, float>);
 }
 
 }  // namespace ops

mace/python/tools/tf_converter_lib.py

@@ -663,7 +663,10 @@ class TFConverter(object):
     op_def.output.extend([output.name for output in op.outputs])
     axis_arg = op_def.arg.add()
     axis_arg.name = 'axis'
-    axis_arg.i = get_input_tensor(op, len(op.inputs) - 1).eval().astype(np.int32)
+    axis = get_input_tensor(op, len(op.inputs) - 1).eval().astype(np.int32)
+    if self.device == 'neon' and axis == 3:
+      axis = 1
+    axis_arg.i = axis
     self.add_output_shape(op.outputs, op_def)
     self.resolved_ops[op.name] = 1
     self.unused_tensor.add(get_input_tensor(op, len(op.inputs) - 1).name)