Merge branch 'hexagon' into 'master'

Support Mace to Hexagon

See merge request !849

mace/core/runtime/hexagon/hexagon_control_wrapper.cc

@@ -324,7 +324,9 @@ void HexagonControlWrapper::GetPerfInfo() {
     }
     ++node_type_counters[node_type].first;
     node_type_counters[node_type].second += node_id_counters[node_id];
-    total_duration += node_id_counters[node_id];
+    if (node_type.compare("Const") != 0) {
+      total_duration += node_id_counters[node_id];
+    }
   }
 
   for (auto &node_type_counter : node_type_counters) {
@@ -332,7 +334,7 @@ void HexagonControlWrapper::GetPerfInfo() {
               << ", time: " << node_type_counter.second.first
               << ", duration: " << node_type_counter.second.second;
   }
-  LOG(INFO) << "total duration: " << total_duration;
+  LOG(INFO) << "total duration: " << std::fixed << total_duration;
 }
 
 void HexagonControlWrapper::ResetPerfInfo() {

mace/proto/mace.proto

@@ -36,7 +36,9 @@ message ConstTensor {
   optional int64 data_size = 7;
   optional float scale = 8;
   optional int32 zero_point = 9;
-  optional bool quantized = 10 [default = false];
+  optional float minval = 10;
+  optional float maxval = 11;
+  optional bool quantized = 12 [default = false];
 
   optional uint32 node_id = 100;
 }

mace/python/tools/BUILD

@@ -10,18 +10,19 @@ py_library(
     name = "converter_lib",
     srcs = [
         "convert_util.py",
-        "graph_util.py",
-        "converter_tool/tf_dsp_converter.py",
         "converter_tool/base_converter.py",
+        "converter_tool/caffe_converter.py",
+        "converter_tool/hexagon_converter.py",
         "converter_tool/shape_inference.py",
         "converter_tool/tensorflow_converter.py",
-        "converter_tool/caffe_converter.py",
+        "converter_tool/tf_dsp_converter.py",
         "converter_tool/transformer.py",
+        "graph_util.py",
     ],
     srcs_version = "PY2AND3",
     deps = [
-        ":quantization_lib",
         ":memory_optimizer",
+        ":quantization_lib",
         "//mace/proto:mace_py",
         "//third_party/caffe:caffe_py",
     ],

mace/python/tools/converter.py

@@ -132,15 +132,30 @@ def main(unused_args):
         option.add_input_node(input_node)
 
     output_node_names = FLAGS.output_node.split(',')
+    output_node_shapes = FLAGS.output_shape.split(':')
+    if len(output_node_names) != len(output_node_shapes):
+        raise Exception('output node count and shape count do not match.')
     for i in six.moves.range(len(output_node_names)):
         output_node = cvt.NodeInfo()
         output_node.name = output_node_names[i]
+        output_node.shape = parse_int_array_from_str(output_node_shapes[i])
         option.add_output_node(output_node)
 
+    if FLAGS.check_node != '':
+        check_node_names = FLAGS.check_node.split(',')
+        check_node_shapes = FLAGS.check_shape.split(':')
+        if len(check_node_names) != len(check_node_shapes):
+            raise Exception('check node count and shape count do not match.')
+        for i in six.moves.range(len(check_node_names)):
+            check_node = cvt.NodeInfo()
+            check_node.name = check_node_names[i]
+            check_node.shape = parse_int_array_from_str(check_node_shapes[i])
+            option.add_check_node(check_node)
+
     option.build()
 
     print("Transform model to one that can better run on device")
-    if FLAGS.runtime == 'dsp':
+    if FLAGS.runtime == 'dsp' and not option.quantize:
         mace_check(FLAGS.platform == 'tensorflow',
                    'DSP only supports tensorflow')
         from mace.python.tools.converter_tool import tf_dsp_converter
@@ -172,7 +187,7 @@ def main(unused_args):
                 FLAGS.data_type, cvt.DeviceType.GPU.value)
             mace_gpu_transformer = transformer.Transformer(
                 option, output_graph_def)
-            output_graph_def = mace_gpu_transformer.run()
+            output_graph_def, _ = mace_gpu_transformer.run()
             six.print_("start optimize gpu memory.")
             memory_optimizer.optimize_gpu_memory(output_graph_def)
             six.print_("GPU memory optimization done.")
@@ -183,7 +198,7 @@ def main(unused_args):
             option.disable_transpose_filters()
             mace_cpu_transformer = transformer.Transformer(
                 option, cpu_graph_def)
-            cpu_graph_def = mace_cpu_transformer.run()
+            cpu_graph_def, _ = mace_cpu_transformer.run()
             print("start optimize cpu memory.")
             memory_optimizer.optimize_cpu_memory(cpu_graph_def)
             print("CPU memory optimization done.")
@@ -206,13 +221,21 @@ def main(unused_args):
                 FLAGS.data_type, option.device)
             mace_transformer = transformer.Transformer(
                 option, output_graph_def)
-            output_graph_def = mace_transformer.run()
+            output_graph_def, quantize_activation_info = mace_transformer.run()
+
+            if FLAGS.runtime == 'dsp':
+                from mace.python.tools.converter_tool import hexagon_converter
+                converter = hexagon_converter.HexagonConverter(
+                    option, output_graph_def, quantize_activation_info)
+                output_graph_def = converter.run()
 
             print("start optimize memory.")
             if FLAGS.runtime == 'gpu':
                 memory_optimizer.optimize_gpu_memory(output_graph_def)
             elif FLAGS.runtime == 'cpu':
                 memory_optimizer.optimize_cpu_memory(output_graph_def)
+            elif FLAGS.runtime == 'dsp':
+                pass
             else:
                 mace_check(False, "runtime only support [gpu|cpu|dsp]")
 
@@ -272,6 +295,8 @@ def parse_args():
         help="e.g., input_node")
     parser.add_argument(
         "--output_node", type=str, default="softmax", help="e.g., softmax")
+    parser.add_argument(
+        "--check_node", type=str, default="softmax", help="e.g., softmax")
     parser.add_argument(
         "--template_dir", type=str, default="", help="template path")
     parser.add_argument(
@@ -297,6 +322,10 @@ def parse_args():
         "--input_shape", type=str, default="", help="input shape.")
     parser.add_argument(
         "--input_range", type=str, default="", help="input range.")
+    parser.add_argument(
+        "--output_shape", type=str, default="", help="output shape.")
+    parser.add_argument(
+        "--check_shape", type=str, default="", help="check shape.")
     parser.add_argument(
         "--platform", type=str, default="tensorflow", help="tensorflow/caffe")
     parser.add_argument(

mace/python/tools/converter_tool/base_converter.py

@@ -40,6 +40,7 @@ class PaddingMode(Enum):
     VALID = 0
     SAME = 1
     FULL = 2
+    NA = 3
 
 
 class PoolingType(Enum):
@@ -270,6 +271,7 @@ class ConverterOption(object):
     def __init__(self):
         self._input_nodes = {}
         self._output_nodes = {}
+        self._check_nodes = {}
         self._data_type = mace_pb2.DT_FLOAT
         self._device = DeviceType.CPU.value
         self._winograd = 0
@@ -287,6 +289,10 @@ class ConverterOption(object):
     def output_nodes(self):
         return self._output_nodes
 
+    @property
+    def check_nodes(self):
+        return self._check_nodes
+
     @property
     def data_type(self):
         return self._data_type
@@ -335,6 +341,14 @@ class ConverterOption(object):
     def add_output_node(self, output_node):
         self._output_nodes[output_node.name] = output_node
 
+    @check_nodes.setter
+    def check_nodes(self, check_nodes):
+        for node in check_nodes:
+            self.check_nodes[node.name] = node
+
+    def add_check_node(self, check_node):
+        self._check_nodes[check_node.name] = check_node
+
     @data_type.setter
     def data_type(self, data_type):
         self._data_type = data_type

mace/python/tools/converter_tool/hexagon_converter.py

+# Copyright 2018 Xiaomi, Inc.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from mace.proto import mace_pb2
+from mace.python.tools.converter_tool import base_converter
+from mace.python.tools.converter_tool.base_converter import ConverterUtil
+from mace.python.tools.converter_tool.base_converter import EltwiseType
+from mace.python.tools.converter_tool.base_converter import MaceKeyword
+from mace.python.tools.converter_tool.base_converter import MaceOp
+from mace.python.tools.converter_tool.base_converter import PaddingMode
+from mace.python.tools.converter_tool.base_converter import PoolingType
+from mace.python.tools.convert_util import mace_check
+from mace.python.tools import graph_util
+
+import copy
+from operator import mul
+
+
+class HexagonOps(object):
+    def __init__(self):
+        self.hexagon_ops = {
+            'INPUT': 'INPUT',
+            'OUTPUT': 'OUTPUT',
+            'Quantize': 'Quantize',
+            'Dequantize': 'Dequantize',
+            'Concat': 'QuantizedConcat_8',
+            'Conv2D': 'Supernode_8x8p32to8',
+            'DepthwiseConv2d': 'DepthwiseSupernode_8x8p32to8',
+            'ResizeBilinear': 'QuantizedResizeBilinear_8',
+            'SpaceToBatchND': 'SpaceToBatchND_8',
+            'BatchToSpaceND': 'BatchToSpaceND_8',
+            'Softmax': 'QuantizedSoftmax_8',
+            'Eltwise': 'Eltwise',
+            'Pooling': 'Pooling',
+            'Identity': 'Nop',
+            'Squeeze': 'Nop',
+        }
+
+    def has_op(self, tf_op):
+        return tf_op in self.hexagon_ops
+
+    def map_nn_op(self, tf_op):
+        if tf_op not in self.hexagon_ops:
+            raise Exception('Could not map nn op for: ', tf_op)
+        return self.hexagon_ops[tf_op]
+
+
+padding_mode = {
+    PaddingMode.NA: 0,
+    PaddingMode.SAME: 1,
+    PaddingMode.VALID: 2
+}
+
+
+def get_tensor_name_from_op(op_name, port):
+    return op_name + ':' + str(port)
+
+
+def get_op_and_port_from_tensor(tensor_name):
+    if ':' in tensor_name:
+        op, port = tensor_name.split(':')
+        port = int(port)
+    else:
+        op = tensor_name
+        port = 0
+    return op, port
+
+
+class HexagonConverter(base_converter.ConverterInterface):
+    def __init__(self, option, model, quantize_activation_info):
+        self._option = option
+        self._model = model
+        self._hexagon_ops = HexagonOps()
+        self._consts = {}
+        self._quantize_activation_info = quantize_activation_info
+
+    def run(self):
+        mace_check(len(self._option.input_nodes) == 1
+                   and len(self._option.output_nodes) == 1,
+                   'dsp only support single input and output')
+
+        for tensor in self._model.tensors:
+            self._consts[tensor.name] = tensor
+
+        # convert op node
+        self.convert_ops()
+
+        self.add_input_output_node()
+
+        self._model = graph_util.sort_mace_graph(self._model, '__output__')
+
+        self.add_node_id()
+
+        return self._model
+
+    def convert_ops(self):
+        print("Convert mace graph to hexagon.")
+        for op in self._model.op:
+            if not self._hexagon_ops.has_op(op.type):
+                raise Exception('Unsupported op: ', op)
+            print('Op: ', op.name, op.type)
+            for i in range(len(op.input)):
+                if ':' not in op.input[i]:
+                    node_name = op.input[i]
+                    op.input[i] += ':0'
+                    if node_name in self._quantize_activation_info:
+                        self._quantize_activation_info[op.input[i]] = \
+                            self._quantize_activation_info[node_name]
+
+            if op.type == MaceOp.Conv2D.name \
+                    or op.type == MaceOp.DepthwiseConv2d.name:
+                mace_check(len(op.input) == 3,
+                           "Missing bias of Conv or Depthwise Conv.")
+                bias = op.input.pop()
+                self.add_min_max_const_node(op, op.input[0])
+                self.add_min_max_const_node(op, op.input[1])
+                strides_arg = ConverterUtil.get_arg(op, 'strides')
+                mace_check(strides_arg is not None,
+                           "Missing strides of Conv or Depthwise Conv.")
+                strides = self.add_shape_const_node(
+                    op, [1, strides_arg.ints[0], strides_arg.ints[1], 1],
+                    MaceKeyword.mace_strides_str)
+                op.input.extend([strides, bias])
+                self.add_min_max_const_node(op, bias)
+                self.add_min_max_const_node(
+                    op, op.output[0], True, True, False)
+            elif op.type == MaceOp.Eltwise.name:
+                self.add_min_max_const_node(op, op.input[0])
+                self.add_min_max_const_node(op, op.input[1])
+                self.add_min_max_const_node(
+                    op, op.output[0], True, True, False)
+            elif op.type == MaceOp.BatchToSpaceND.name \
+                    or op.type == MaceOp.SpaceToBatchND.name:
+                strides_arg = ConverterUtil.get_arg(
+                    op, MaceKeyword.mace_space_batch_block_shape_str)
+                strides_tensor = self._model.tensors.add()
+                strides_tensor.name = op.name + '/strides:0'
+                strides_tensor.data_type = mace_pb2.DT_INT32
+                strides_tensor.dims.extend([1, 1, 1, len(strides_arg.ints)])
+                strides_tensor.int32_data.extend(strides_arg.ints)
+                if op.type == MaceOp.BatchToSpaceND.name:
+                    pad_arg = ConverterUtil.get_arg(
+                        op, MaceKeyword.mace_batch_to_space_crops_str)
+                else:
+                    pad_arg = ConverterUtil.get_arg(
+                        op, MaceKeyword.mace_paddings_str)
+                pad_tensor = self._model.tensors.add()
+                pad_tensor.name = op.name + '/pad:0'
+                pad_tensor.data_type = mace_pb2.DT_INT32
+                pad_tensor.dims.extend([1, 1, len(pad_arg.ints) / 2, 2])
+                pad_tensor.int32_data.extend(pad_arg.ints)
+                op.input.extend([strides_tensor.name, pad_tensor.name])
+                self.add_min_max_const_node(op, op.input[0])
+            elif op.type == MaceOp.Pooling.name:
+                self.add_min_max_const_node(op, op.input[0])
+                window_arg = ConverterUtil.get_arg(
+                    op, MaceKeyword.mace_kernel_str)
+                window_tensor = self._model.tensors.add()
+                window_tensor.name = op.name + '/window:0'
+                window_tensor.data_type = mace_pb2.DT_INT32
+                window_tensor.dims.extend(
+                    [1, window_arg.ints[0], window_arg.ints[1], 1])
+                strides_arg = ConverterUtil.get_arg(
+                    op, MaceKeyword.mace_strides_str)
+                strides_tensor = self._model.tensors.add()
+                strides_tensor.name = op.name + '/strides:0'
+                strides_tensor.data_type = mace_pb2.DT_INT32
+                strides_tensor.dims.extend(
+                    [1, strides_arg.ints[0], strides_arg.ints[1], 1])
+                op.input.extend([window_tensor.name, strides_tensor.name])
+            elif op.type == MaceOp.ResizeBilinear.name:
+                newdim_arg = ConverterUtil.get_arg(
+                    op, MaceKeyword.mace_resize_size_str)
+                newdim_tensor = self._model.tensors.add()
+                newdim_tensor.name = op.name + '/newdim:0'
+                newdim_tensor.data_type = mace_pb2.DT_INT32
+                newdim_tensor.dims.extend([len(newdim_arg.ints)])
+                newdim_tensor.int32_data.extend(newdim_arg.ints)
+                op.input.extend([newdim_tensor.name])
+                self.add_min_max_const_node(op, op.input[0])
+            elif op.type == MaceOp.Concat.name:
+                inputs = copy.deepcopy(op.input)
+                for ipt in inputs:
+                    self.add_min_max_const_node(op, ipt, True, False)
+                for ipt in inputs:
+                    self.add_min_max_const_node(op, ipt, False, True)
+                dim_arg = ConverterUtil.get_arg(
+                    op, MaceKeyword.mace_axis_str)
+                dim_tensor = self._model.tensors.add()
+                dim_tensor.name = op.name + '/dim:0'
+                dim_tensor.data_type = mace_pb2.DT_INT32
+                dim_tensor.dims.extend([1])
+                dim_tensor.int32_data.extend([dim_arg.i])
+                op.input.insert(0, dim_tensor.name)
+            elif op.type in [MaceOp.Softmax.name,
+                             MaceOp.Dequantize.name]:
+                self.add_min_max_const_node(op, op.input[0])
+
+            if op.type != MaceOp.Dequantize.name:
+                min_output_shape = op.output_shape.add()
+                min_output_shape.dims.extend([1])
+                max_output_shape = op.output_shape.add()
+                max_output_shape.dims.extend([1])
+                op.output_type.extend(
+                    [mace_pb2.DT_UINT8, mace_pb2.DT_FLOAT, mace_pb2.DT_FLOAT])
+            for i in range(len(op.output_shape)):
+                out_max_byte_size = reduce(mul, op.output_shape[i].dims)
+                if op.output_type[i] == mace_pb2.DT_FLOAT:
+                    out_max_byte_size *= 4
+                op.out_max_byte_size.extend([out_max_byte_size])
+
+            op.padding = padding_mode[PaddingMode.NA]
+            arg = ConverterUtil.get_arg(op, MaceKeyword.mace_padding_str)
+            if arg is not None:
+                op.padding = padding_mode[PaddingMode(arg.i)]
+
+            if (op.type == MaceOp.Eltwise.name
+                    and ConverterUtil.get_arg(
+                        op, MaceKeyword.mace_element_type_str).i
+                    == EltwiseType.SUM.value):
+                op.type = 'QuantizedAdd_8p8to8'
+            elif op.type == MaceOp.Pooling.name:
+                pooling_type_arg = ConverterUtil.get_arg(
+                    op, MaceKeyword.mace_pooling_type_str)
+                if PoolingType(pooling_type_arg.i) == PoolingType.AVG:
+                    op.type = 'QuantizedAvgPool_8'
+                else:
+                    op.type = 'QuantizedMaxPool_8'
+            else:
+                op.type = self._hexagon_ops.map_nn_op(op.type)
+
+    def add_min_max(self, name, val):
+        if name not in self._consts:
+            tensor = self._model.tensors.add()
+            self._consts[name] = tensor
+            tensor.name = name
+            tensor.data_type = mace_pb2.DT_FLOAT
+            tensor.dims.extend([1])
+            tensor.float_data.extend([val])
+
+    def add_min_max_const_node(
+            self, this_op, tensor_name, add_min=True, add_max=True,
+            diff_port=True):
+        op, port = get_op_and_port_from_tensor(tensor_name)
+        mace_check(port == 0, 'port should be 0 to add min max tensor then.')
+        if tensor_name in self._quantize_activation_info:
+            quantize_info = self._quantize_activation_info[tensor_name]
+            minval = quantize_info.minval
+            maxval = quantize_info.maxval
+            is_activation = True
+        elif tensor_name in self._consts:
+            tensor = self._consts[tensor_name]
+            minval = tensor.minval
+            maxval = tensor.maxval
+            is_activation = False
+        else:
+            raise Exception('Quantize info not found: ', tensor_name)
+
+        if add_min:
+            if is_activation and diff_port:
+                min_tensor_name = op + ':1'
+            else:
+                min_tensor_name = op + '_min:0'
+                self.add_min_max(min_tensor_name, minval)
+            this_op.input.extend([min_tensor_name])
+        if add_max:
+            if is_activation and diff_port:
+                max_tensor_name = op + ':2'
+            else:
+                max_tensor_name = op + '_max:0'
+                self.add_min_max(max_tensor_name, maxval)
+            this_op.input.extend([max_tensor_name])
+
+    def add_shape_const_node(self, op, values, name):
+        tensor = self._model.tensors.add()
+        node_name = op.name + '/' + name
+        tensor.name = node_name + ':0'
+        tensor.data_type = mace_pb2.DT_INT32
+        tensor.dims.extend(values)
+        return tensor.name
+
+    def add_input_output_node(self):
+        input_node = self._option.input_nodes.values()[0]
+        op_def = self._model.op.add()
+        op_def.name = '__input__'
+        op_def.type = 'INPUT'
+        shape = op_def.output_shape.add()
+        shape.dims.extend(input_node.shape)
+        op_def.output_type.extend([mace_pb2.DT_FLOAT])
+        out_max_byte_size = reduce(mul, shape.dims)
+        op_def.out_max_byte_size.extend([out_max_byte_size])
+        for op in self._model.op:
+            if op.name == input_node.name:
+                del op.input[0]
+                input_name = op_def.name + ':0'
+                op.input.extend([input_name])
+                self._consts[input_name] = \
+                    self._quantize_activation_info[input_node.name]
+                self.add_min_max_const_node(op, input_name)
+                del self._consts[input_name]
+                break
+
+        output_node = None
+        if not self._option.check_nodes:
+            output_name = self._option.output_nodes.values()[0].name
+        else:
+            output_name = self._option.check_nodes.values()[0].name
+        for op in self._model.op:
+            if op.name.startswith(MaceKeyword.mace_output_node_name) \
+                    and op.name.find(output_name) != -1:
+                output_node = op
+                break
+        mace_check(output_node is not None,
+                   "mace_output_node_* not found.")
+        op_def = self._model.op.add()
+        op_def.name = '__output__'
+        op_def.type = 'OUTPUT'
+        op_def.input.extend([get_tensor_name_from_op(output_node.name, 0)])
+
+    def add_node_id(self):
+        node_id_counter = 0
+        node_id_map = {}
+        for tensor in self._model.tensors:
+            tensor.node_id = node_id_counter
+            node_id_counter += 1
+            tensor_op, port = get_op_and_port_from_tensor(tensor.name)
+            node_id_map[tensor_op] = tensor.node_id
+
+        for op in self._model.op:
+            op.node_id = node_id_counter
+            node_id_counter += 1
+            node_id_map[op.name] = op.node_id
+            for ipt in op.input:
+                if ipt.startswith(MaceKeyword.mace_input_node_name):
+                    ipt = ipt[len(MaceKeyword.mace_input_node_name + '_'):]
+                op_name, port = get_op_and_port_from_tensor(ipt)
+                node_id = node_id_map[op_name]
+                node_input = op.node_input.add()
+                node_input.node_id = node_id
+                node_input.output_port = int(port)

mace/python/tools/converter_tool/transformer.py

@@ -134,7 +134,9 @@ class Transformer(base_converter.ConverterInterface):
                 changed = transformer()
                 if not changed:
                         break
-        return self._model
+
+        self.add_check_nodes()
+        return self._model, self._quantize_activation_info
 
     def filter_format(self):
         filter_format_value = ConverterUtil.get_arg(self._model,
@@ -284,12 +286,20 @@ class Transformer(base_converter.ConverterInterface):
             input_info = net.input_info.add()
             input_info.name = input_node.name
             input_info.dims.extend(input_node.shape)
+            if self._option.quantize:
+                input_info.data_type = mace_pb2.DT_FLOAT
+            else:
+                input_info.data_type = self._option.data_type
 
         for output_node in self._option.output_nodes.values():
             output_info = net.output_info.add()
             output_info.name = output_node.name
             output_info.dims.extend(
                 self._producer[output_node.name].output_shape[0].dims)
+            if self._option.quantize:
+                output_info.data_type = mace_pb2.DT_FLOAT
+            else:
+                output_info.data_type = self._option.data_type
 
         return False
 
@@ -904,6 +914,8 @@ class Transformer(base_converter.ConverterInterface):
                 consumer_op = self._consumers[op.output[0]][0]
                 if consumer_op.type == MaceOp.BiasAdd.name:
                     print("Fold biasadd: %s(%s)" % (op.name, op.type))
+                    op.name = consumer_op.name
+                    op.output[0] = consumer_op.output[0]
                     op.input.append(consumer_op.input[1])
                     self.replace_quantize_info(op, consumer_op)
                     self.safe_remove_node(consumer_op, op)
@@ -1306,6 +1318,11 @@ class Transformer(base_converter.ConverterInterface):
                     transposed_deconv_filter.add(op.input[1])
 
             self.set_filter_format(FilterFormat.OHWI)
+        elif self._option.quantize and \
+                self._option.device == DeviceType.HEXAGON.value:
+            print("Transpose filters to HWIO/HWIM")
+            mace_check(filter_format == FilterFormat.HWIO,
+                       "HEXAGON only support HWIO/HWIM filter format.")
         else:
             print("Transpose filters to OIHW/MIHW")
             # transpose filter to OIHW/MIHW for tensorflow (HWIO/HWIM)
@@ -1795,16 +1812,23 @@ class Transformer(base_converter.ConverterInterface):
                         check_deconv = len(ops[0].input) >= 4\
                                        and ops[0].input[3] == tensor.name
             if check_conv or check_deconv:
-                conv_op = ops[0]
-                scale_input = self._quantize_activation_info[
-                    conv_op.input[0]].scale
-                if conv_op.input[1] not in self._quantized_tensor:
-                    self.quantize_tensor(self._consts[conv_op.input[1]])
-                scale_filter = self._consts[conv_op.input[1]].scale
-                scale = scale_input * scale_filter
-
-                quantized_tensor = quantize_util.quantize_with_scale_and_zero(
-                    tensor.float_data, scale, 0)
+                if self._option.device == DeviceType.CPU.value:
+                    conv_op = ops[0]
+                    scale_input = self._quantize_activation_info[
+                        conv_op.input[0]].scale
+                    if conv_op.input[1] not in self._quantized_tensor:
+                        self.quantize_tensor(self._consts[conv_op.input[1]])
+                    scale_filter = self._consts[conv_op.input[1]].scale
+                    scale = scale_input * scale_filter
+                    quantized_tensor = \
+                        quantize_util.quantize_with_scale_and_zero(
+                            tensor.float_data, scale, 0)
+                elif self._option.device == DeviceType.HEXAGON.value:
+                    quantized_tensor = \
+                        quantize_util.quantize_bias_for_hexagon(
+                            tensor.float_data)
+                else:
+                    mace_check(False, "wrong device.")
                 tensor.data_type = mace_pb2.DT_INT32
             else:
                 quantized_tensor = quantize_util.quantize(tensor.float_data)
@@ -1814,6 +1838,8 @@ class Transformer(base_converter.ConverterInterface):
             tensor.int32_data.extend(quantized_tensor.data)
             tensor.scale = quantized_tensor.scale
             tensor.zero_point = quantized_tensor.zero
+            tensor.minval = quantized_tensor.minval
+            tensor.maxval = quantized_tensor.maxval
             tensor.quantized = True
             self._quantized_tensor.update([tensor.name])
 
@@ -1828,8 +1854,8 @@ class Transformer(base_converter.ConverterInterface):
         return False
 
     def add_quantize_info(self, op, minval, maxval):
-        scale, zero = quantize_util.adjust_range(minval, maxval,
-                                                 non_zero=False)
+        scale, zero, minval, maxval = \
+            quantize_util.adjust_range(minval, maxval, non_zero=False)
         quantize_info = op.quantize_info.add()
         quantize_info.minval = minval
         quantize_info.maxval = maxval
@@ -1928,8 +1954,9 @@ class Transformer(base_converter.ConverterInterface):
                     tensor_name, minmax = line.split("@@")[:2]
                     min_val, max_val = [float(i) for i in
                                         minmax.strip().split(",")]
-                    scale, zero = quantize_util.adjust_range(min_val, max_val,
-                                                             non_zero=False)
+                    scale, zero, min_val, max_val = \
+                        quantize_util.adjust_range(
+                            min_val, max_val, non_zero=False)
                     activation_info = mace_pb2.QuantizeActivationInfo()
                     activation_info.minval = min_val
                     activation_info.maxval = max_val
@@ -1954,6 +1981,7 @@ class Transformer(base_converter.ConverterInterface):
         for op in self._model.op:
             if op.type in [MaceOp.Pooling.name,
                            MaceOp.Squeeze.name,
+                           MaceOp.Reshape.name,
                            MaceOp.ResizeBilinear.name,
                            MaceOp.BatchToSpaceND.name,
                            MaceOp.SpaceToBatchND.name]:
@@ -2012,12 +2040,13 @@ class Transformer(base_converter.ConverterInterface):
             if input_node.name not in self._quantize_activation_info:
                 print("Input range %s: %s" % (input_node.name,
                                               str(input_node.range)))
-                scale, zero = quantize_util.adjust_range(input_node.range[0],
-                                                         input_node.range[1],
-                                                         non_zero=False)
+                scale, zero, minval, maxval = \
+                    quantize_util.adjust_range(input_node.range[0],
+                                               input_node.range[1],
+                                               non_zero=False)
                 quantize_info = mace_pb2.QuantizeActivationInfo()
-                quantize_info.minval = input_node.range[0]
-                quantize_info.maxval = input_node.range[1]
+                quantize_info.minval = minval
+                quantize_info.maxval = maxval
                 quantize_info.scale = scale
                 quantize_info.zero_point = zero
                 self._quantize_activation_info[input_node.name] = quantize_info
@@ -2049,3 +2078,32 @@ class Transformer(base_converter.ConverterInterface):
         arg.name = MaceKeyword.mace_opencl_mem_type
         arg.i = mace_pb2.GPU_IMAGE if self._option.cl_mem_type == "image"\
             else mace_pb2.GPU_BUFFER
+
+    def add_check_nodes(self):
+        if self._option.check_nodes:
+            mace_check(len(self._option.check_nodes) == 1,
+                       "Only support one check node now.")
+            check_node = None
+            for i in six.moves.range(len(self._model.op)):
+                if self._model.op[i].name in self._option.check_nodes:
+                    check_node = self._model.op[i]
+                    del self._model.op[i+1:]
+                    break
+            mace_check(check_node is not None, "check node not found.")
+            output_name = \
+                MaceKeyword.mace_output_node_name + '_' + check_node.name
+            op_def = self._model.op.add()
+            op_def.name = self.normalize_op_name(output_name)
+            op_def.type = MaceOp.Dequantize.name
+            op_def.input.extend([check_node.output[0]])
+            op_def.output.extend([output_name])
+            output_shape = op_def.output_shape.add()
+            output_shape.dims.extend(check_node.output_shape[0].dims)
+            ConverterUtil.add_data_type_arg(op_def, mace_pb2.DT_UINT8)
+            op_def.output_type.extend([mace_pb2.DT_FLOAT])
+
+            del self._model.output_info[:]
+            output_info = self._model.output_info.add()
+            output_info.name = check_node.name
+            output_info.dims.extend(check_node.output_shape[0].dims)
+            output_info.data_type = mace_pb2.DT_FLOAT

mace/python/tools/graph_util.py

@@ -63,7 +63,6 @@ def sort_mace_graph(graph_def, output_name):
     for node in graph_def.op:
         nodes_map[node.name] = node
     sort_mace_node(nodes_map[output_name], nodes_map, ordered_nodes_map)
-    sorted_graph = mace_pb2.NetDef()
-    sorted_graph.tensors.extend(graph_def.tensors)
-    sorted_graph.op.extend([node for node in ordered_nodes_map.values()])
-    return sorted_graph
+    del graph_def.op[:]
+    graph_def.op.extend([node for node in ordered_nodes_map.values()])
+    return graph_def

mace/python/tools/model_saver.py

@@ -20,6 +20,7 @@ import hashlib
 from enum import Enum
 
 from mace.proto import mace_pb2
+from mace.python.tools.convert_util import mace_check
 from jinja2 import Environment, FileSystemLoader
 
 GENERATED_NAME = set()
@@ -107,11 +108,9 @@ class TensorInfo:
         self.id = id
         self.data_type = tensor.data_type
         if tensor.data_type == mace_pb2.DT_HALF:
-            self.data_type = mace_pb2.DT_HALF
             self.data = bytearray(
                 np.array(tensor.float_data).astype(np.float16).tobytes())
         elif tensor.data_type == mace_pb2.DT_FLOAT:
-            self.data_type = mace_pb2.DT_FLOAT
             self.data = bytearray(
                 np.array(tensor.float_data).astype(np.float32).tobytes())
         elif tensor.data_type == mace_pb2.DT_INT32:
@@ -139,7 +138,7 @@ def update_tensor_infos(net_def, runtime, data_type):
         tensor_info = TensorInfo(counter, tensor)
         tensor_infos.append(tensor_info)
         # align
-        if tensor_info.data_type != 'DT_UINT8' and offset % 4 != 0:
+        if tensor_info.data_type != mace_pb2.DT_UINT8 and offset % 4 != 0:
             padding = 4 - offset % 4
             offset += padding
 
@@ -162,10 +161,11 @@ def extract_model_data(net_def):
     for tensor in net_def.tensors:
         tensor_info = TensorInfo(counter, tensor)
         # align
-        if tensor_info.data_type != mace_pb2.DT_UINT8 and offset % 4 != 0:
-            padding = 4 - offset % 4
-            model_data.extend(bytearray([0] * padding))
-            offset += padding
+        mace_check(offset <= tensor.offset,
+                   "Current offset should be <= tensor.offset")
+        if offset < tensor.offset:
+            model_data.extend(bytearray([0] * (tensor.offset - offset)))
+            offset = tensor.offset
         model_data.extend(tensor_info.data)
         offset += len(tensor_info.data)
         counter += 1

mace/python/tools/quantization/quantize_util.py

@@ -7,6 +7,8 @@ class QuantizedData(object):
         self._data = None
         self._scale = 0
         self._zero = 0
+        self._minval = 0.0
+        self._maxval = 0.0
 
     @property
     def data(self):
@@ -20,6 +22,14 @@ class QuantizedData(object):
     def zero(self):
         return self._zero
 
+    @property
+    def minval(self):
+        return self._minval
+
+    @property
+    def maxval(self):
+        return self._maxval
+
     @data.setter
     def data(self, data):
         self._data = data
@@ -32,6 +42,14 @@ class QuantizedData(object):
     def zero(self, zero):
         self._zero = zero
 
+    @minval.setter
+    def minval(self, minval):
+        self._minval = minval
+
+    @maxval.setter
+    def maxval(self, maxval):
+        self._maxval = maxval
+
 
 def adjust_range(in_min, in_max, non_zero):
     out_max = max(0.0, in_max)
@@ -54,7 +72,7 @@ def adjust_range(in_min, in_max, non_zero):
     else:
         zero_int = 255
 
-    return scale, zero_int
+    return scale, zero_int, -zero_int*scale, (255-zero_int)*scale
 
 
 def cal_multiplier_and_shift(scale):
@@ -94,13 +112,34 @@ def quantize(data):
     np_data = np.array(data).astype(float)
     in_min = np_data.min()
     in_max = np_data.max()
-    scale, zero = adjust_range(in_min, in_max, non_zero=True)
+    scale, zero, out_min, out_max = adjust_range(in_min, in_max, non_zero=True)
     output = np.clip((np.round(zero + data / scale).astype(int)), 0, 255)
 
     quantized_data = QuantizedData()
     quantized_data.data = output
     quantized_data.scale = scale
     quantized_data.zero = zero
+    quantized_data.minval = out_min
+    quantized_data.maxval = out_max
+    return quantized_data
+
+
+def quantize_bias_for_hexagon(data):
+    np_data = np.array(data).astype(float)
+    max_val = max(abs(np_data.min()), abs(np_data.max()))
+    in_min = -max_val
+    in_max = max_val
+    scale = (in_max - in_min) / 2**32
+    zero = 0
+    output = np.clip((np.round(zero + data / scale).astype(long)),
+                     -2**31, 2**31 - 1)
+
+    quantized_data = QuantizedData()
+    quantized_data.data = output
+    quantized_data.scale = scale
+    quantized_data.zero = zero
+    quantized_data.minval = in_min
+    quantized_data.maxval = in_max
     return quantized_data
 
 

tools/converter.py

@@ -190,6 +190,8 @@ class YAMLKeyword(object):
     input_ranges = 'input_ranges'
     output_tensors = 'output_tensors'
     output_shapes = 'output_shapes'
+    check_tensors = 'check_tensors'
+    check_shapes = 'check_shapes'
     runtime = 'runtime'
     data_type = 'data_type'
     input_data_types = 'input_data_types'
@@ -460,6 +462,16 @@ def format_model_config(flags):
                     subgraph[key] = [value]
                 subgraph[key] = [str(v) for v in subgraph[key]]
 
+            for key in [YAMLKeyword.check_tensors,
+                        YAMLKeyword.check_shapes]:
+                value = subgraph.get(key, "")
+                if value != "":
+                    if not isinstance(value, list):
+                        subgraph[key] = [value]
+                    subgraph[key] = [str(v) for v in subgraph[key]]
+                else:
+                    subgraph[key] = []
+
             input_data_types = subgraph.get(YAMLKeyword.input_data_types, "")
             if input_data_types:
                 if not isinstance(input_data_types, list):
@@ -787,10 +799,13 @@ def convert_model(configs, cl_mem_type):
             model_config[YAMLKeyword.weight_sha256_checksum],
             ",".join(subgraphs[0][YAMLKeyword.input_tensors]),
             ",".join(subgraphs[0][YAMLKeyword.output_tensors]),
+            ",".join(subgraphs[0][YAMLKeyword.check_tensors]),
             runtime,
             model_name,
             ":".join(subgraphs[0][YAMLKeyword.input_shapes]),
             ":".join(subgraphs[0][YAMLKeyword.input_ranges]),
+            ":".join(subgraphs[0][YAMLKeyword.output_shapes]),
+            ":".join(subgraphs[0][YAMLKeyword.check_shapes]),
             model_config[YAMLKeyword.nnlib_graph_mode],
             embed_model_data,
             model_config[YAMLKeyword.winograd],
@@ -1216,6 +1231,12 @@ def run_specific_target(flags, configs, target_abi,
         for runtime in runtime_list:
             device_type = parse_device_type(runtime)
             # run for specified soc
+            if not subgraphs[0][YAMLKeyword.check_tensors]:
+                output_nodes = subgraphs[0][YAMLKeyword.output_tensors]
+                output_shapes = subgraphs[0][YAMLKeyword.output_shapes]
+            else:
+                output_nodes = subgraphs[0][YAMLKeyword.check_tensors]
+                output_shapes = subgraphs[0][YAMLKeyword.check_shapes]
             run_output = sh_commands.tuning_run(
                 abi=target_abi,
                 serialno=serial_num,
@@ -1225,9 +1246,9 @@ def run_specific_target(flags, configs, target_abi,
                 embed_model_data=embed_model_data,
                 model_output_dir=model_output_dir,
                 input_nodes=subgraphs[0][YAMLKeyword.input_tensors],
-                output_nodes=subgraphs[0][YAMLKeyword.output_tensors],
+                output_nodes=output_nodes,
                 input_shapes=subgraphs[0][YAMLKeyword.input_shapes],
-                output_shapes=subgraphs[0][YAMLKeyword.output_shapes],
+                output_shapes=output_shapes,
                 mace_model_dir=mace_model_dir,
                 model_tag=model_name,
                 device_type=device_type,
@@ -1261,7 +1282,8 @@ def run_specific_target(flags, configs, target_abi,
                     model_config[YAMLKeyword.weight_sha256_checksum])
 
                 validate_type = device_type
-                if model_config[YAMLKeyword.quantize] == 1:
+                if model_config[YAMLKeyword.quantize] == 1 \
+                        and device_type == DeviceType.CPU:
                     validate_type = device_type + "_QUANTIZE"
 
                 sh_commands.validate_model(
@@ -1272,9 +1294,9 @@ def run_specific_target(flags, configs, target_abi,
                     platform=model_config[YAMLKeyword.platform],
                     device_type=device_type,
                     input_nodes=subgraphs[0][YAMLKeyword.input_tensors],
-                    output_nodes=subgraphs[0][YAMLKeyword.output_tensors],
+                    output_nodes=output_nodes,
                     input_shapes=subgraphs[0][YAMLKeyword.input_shapes],
-                    output_shapes=subgraphs[0][YAMLKeyword.output_shapes],
+                    output_shapes=output_shapes,
                     model_output_dir=model_output_dir,
                     phone_data_dir=PHONE_DATA_DIR,
                     input_data_types=subgraphs[0][YAMLKeyword.input_data_types],  # noqa

tools/sh_commands.py

@@ -551,10 +551,13 @@ def gen_model_code(model_codegen_dir,
                    weight_sha256_checksum,
                    input_nodes,
                    output_nodes,
+                   check_nodes,
                    runtime,
                    model_tag,
                    input_shapes,
                    input_ranges,
+                   output_shapes,
+                   check_shapes,
                    dsp_mode,
                    embed_model_data,
                    winograd,
@@ -581,11 +584,14 @@ def gen_model_code(model_codegen_dir,
               "--weight_checksum=%s" % weight_sha256_checksum,
               "--input_node=%s" % input_nodes,
               "--output_node=%s" % output_nodes,
+              "--check_node=%s" % check_nodes,
               "--runtime=%s" % runtime,
               "--template=%s" % "mace/python/tools",
               "--model_tag=%s" % model_tag,
               "--input_shape=%s" % input_shapes,
               "--input_range=%s" % input_ranges,
+              "--output_shape=%s" % output_shapes,
+              "--check_shape=%s" % check_shapes,
               "--dsp_mode=%s" % dsp_mode,
               "--embed_model_data=%s" % embed_model_data,
               "--winograd=%s" % winograd,