Merge branch 'transpose_matmul_weight' into 'master'

Add transpose weight of Matmul in model transformer.

See merge request !749

mace/python/tools/converter_tool/base_converter.py

@@ -195,6 +195,7 @@ class TransformerRule(Enum):
     QUANTIZE_NODES = 23
     ADD_QUANTIZE_TENSOR_RANGE = 24
     QUANTIZE_WEIGHTS = 25
+    TRANSPOSE_MATMUL_WEIGHT = 26
 
 
 class ConverterInterface(object):
@@ -345,6 +346,8 @@ class ConverterOption(object):
                 TransformerRule.FOLD_ACTIVATION,
                 TransformerRule.TRANSFORM_GLOBAL_CONV_TO_FC,
                 TransformerRule.RESHAPE_FC_WEIGHT,
+                # Transpose the weight of matmul if necessary
+                TransformerRule.TRANSPOSE_MATMUL_WEIGHT,
                 # Model data format related transformation
                 TransformerRule.TRANSPOSE_FILTERS,
                 TransformerRule.TRANSPOSE_DATA_FORMAT,

mace/python/tools/converter_tool/transformer.py

@@ -64,6 +64,9 @@ class Transformer(base_converter.ConverterInterface):
             TransformerRule.FOLD_BIASADD: self.fold_biasadd,
             TransformerRule.FLATTEN_ATROUS_CONV: self.flatten_atrous_conv,
             TransformerRule.FOLD_ACTIVATION: self.fold_activation,
+            # TODO(liuqi): should move to transpose_filter
+            TransformerRule.TRANSPOSE_MATMUL_WEIGHT:
+                self.transpose_matmul_weight,
             TransformerRule.TRANSPOSE_FILTERS: self.transpose_filters,
             TransformerRule.TRANSPOSE_DATA_FORMAT: self.transpose_data_format,
             TransformerRule.ADD_IN_OUT_TENSOR_INFO:
@@ -954,6 +957,29 @@ class Transformer(base_converter.ConverterInterface):
 
         return False
 
+    def transpose_matmul_weight(self):
+        if self._option.device != DeviceType.CPU.value:
+            return False
+        net = self._model
+        transpose_arg_names = [MaceKeyword.mace_transpose_a_str,
+                               MaceKeyword.mace_transpose_b_str]
+        for op in net.op:
+            if op.type == MaceOp.MatMul.name:  # noqa
+                for i in range(len(op.input)):
+                    input = op.input[i]
+                    if input in self._consts \
+                            and len(self._consts[input].dims) == 2:
+                        arg = ConverterUtil.get_arg(op, transpose_arg_names[i])
+                        if arg is not None and arg.i == 1:
+                            print 'convert matmul'
+                            filter = self._consts[input]
+                            filter_data = np.array(filter.float_data).reshape(
+                                filter.dims)
+                            filter_data = filter_data.transpose(1, 0)
+                            filter.float_data[:] = filter_data.flat
+                            filter.dims[:] = filter_data.shape
+                            arg.i = 0
+
     def transpose_filters(self):
         net = self._model
         filter_format = self.filter_format()

tools/converter.py

@@ -457,6 +457,7 @@ def format_model_config(flags):
                            "'%s' is necessary in subgraph" % key)
                 if not isinstance(value, list):
                     subgraph[key] = [value]
+                subgraph[key] = [str(v) for v in subgraph[key]]
 
             input_data_types = subgraph.get(YAMLKeyword.input_data_types, "")
             if input_data_types:
@@ -507,6 +508,8 @@ def format_model_config(flags):
                 subgraph[YAMLKeyword.input_ranges] = [input_ranges]
             else:
                 subgraph[YAMLKeyword.input_ranges] = input_ranges
+            subgraph[YAMLKeyword.input_ranges] =\
+                [str(v) for v in subgraph[YAMLKeyword.input_ranges]]
 
         for key in [YAMLKeyword.limit_opencl_kernel_time,
                     YAMLKeyword.nnlib_graph_mode,

tools/generate_data.py

@@ -44,19 +44,22 @@ def generate_input_data(input_file, input_node, input_shape, input_ranges,
                         input_data_type):
     input_names = [name for name in input_node.split(',')]
     input_shapes = [shape for shape in input_shape.split(':')]
+
     if input_ranges:
         input_ranges = [r for r in input_ranges.split(':')]
     else:
-        input_ranges = [[-1, 1]] * len(input_names)
+        input_ranges = ["-1,1"] * len(input_names)
     if input_data_type:
         input_data_types = [data_type
                             for data_type in input_data_type.split(',')]
     else:
         input_data_types = ['float32'] * len(input_names)
+
     assert len(input_names) == len(input_shapes) == len(input_ranges) == len(input_data_types)  # noqa
     for i in range(len(input_names)):
         shape = [int(x) for x in input_shapes[i].split(',')]
-        generate_data(input_names[i], shape, input_file, input_ranges[i],
+        input_range = [float(x) for x in input_ranges[i].split(',')]
+        generate_data(input_names[i], shape, input_file, input_range,
                       input_data_types[i])
     print "Generate input file done."