Merge branch 'fix-tensor-reuse' into 'master'

Fix transpose filter repeatly when different ops reuse same tensor.

See merge request !754

mace/python/tools/converter_tool/transformer.py

@@ -983,6 +983,7 @@ class Transformer(base_converter.ConverterInterface):
     def transpose_filters(self):
         net = self._model
         filter_format = self.filter_format()
+        transposed_filter = set()
 
         if self._option.quantize:
             print("Transpose filters to OHWI")
@@ -995,14 +996,16 @@ class Transformer(base_converter.ConverterInterface):
                            "filter format: %s" % filter_format.name)
 
             for op in net.op:
-                if op.type == MaceOp.Conv2D.name \
-                        or op.type == MaceOp.Deconv2D.name:
+                if (op.type == MaceOp.Conv2D.name
+                        or op.type == MaceOp.Deconv2D.name) \
+                        and op.input[1] not in transposed_filter:
                     filter = self._consts[op.input[1]]
                     filter_data = np.array(filter.float_data).reshape(
                         filter.dims)
                     filter_data = filter_data.transpose(transpose_order)
                     filter.float_data[:] = filter_data.flat
                     filter.dims[:] = filter_data.shape
+                    transposed_filter.add(op.input[1])
 
             self.set_filter_format(FilterFormat.OHWI)
         else:
@@ -1010,24 +1013,29 @@ class Transformer(base_converter.ConverterInterface):
             # transpose filter to OIHW/MIHW for tensorflow (HWIO/HWIM)
             if filter_format == FilterFormat.HWIO:
                 for op in net.op:
-                    if op.type == MaceOp.Conv2D.name \
-                            or op.type == MaceOp.Deconv2D.name \
-                            or op.type == MaceOp.DepthwiseConv2d.name:
+                    if (op.type == MaceOp.Conv2D.name
+                            or op.type == MaceOp.Deconv2D.name
+                            or op.type == MaceOp.DepthwiseConv2d.name) \
+                            and op.input[1] not in transposed_filter:
                         filter = self._consts[op.input[1]]
                         filter_data = np.array(filter.float_data).reshape(
                             filter.dims)
                         filter_data = filter_data.transpose(3, 2, 0, 1)
                         filter.float_data[:] = filter_data.flat
                         filter.dims[:] = filter_data.shape
-                    if (op.type == MaceOp.MatMul.name and
-                                ConverterUtil.get_arg(op, MaceKeyword.mace_winograd_filter_transformed) is not None):  # noqa
+                        transposed_filter.add(op.input[1])
+                    if (op.type == MaceOp.MatMul.name
+                            and (ConverterUtil.get_arg(op, MaceKeyword.mace_winograd_filter_transformed) is not None)  # noqa
+                            and op.input[1] not in transposed_filter):
                         filter = self._consts[op.input[0]]
                         filter_data = np.array(filter.float_data).reshape(
                             filter.dims)
                         filter_data = filter_data.transpose(3, 2, 0, 1)
                         filter.float_data[:] = filter_data.flat
                         filter.dims[:] = filter_data.shape
-                    if op.type == MaceOp.FullyConnected.name:
+                        transposed_filter.add(op.input[0])
+                    if op.type == MaceOp.FullyConnected.name \
+                            and op.input[1] not in transposed_filter:
                         weight = self._consts[op.input[1]]
                         if len(weight.dims) == 4:
                             weight_data = np.array(weight.float_data).reshape(
@@ -1035,16 +1043,19 @@ class Transformer(base_converter.ConverterInterface):
                             weight_data = weight_data.transpose(3, 2, 0, 1)
                             weight.float_data[:] = weight_data.flat
                             weight.dims[:] = weight_data.shape
+                            transposed_filter.add(op.input[1])
 
                 self.set_filter_format(FilterFormat.OIHW)
             for op in net.op:
-                if op.type == MaceOp.Deconv2D.name:
+                if op.type == MaceOp.Deconv2D.name \
+                        and op.input[1] not in transposed_filter:
                     filter = self._consts[op.input[1]]
                     filter_data = np.array(filter.float_data).reshape(
                         filter.dims)
                     filter_data = filter_data.transpose(1, 0, 2, 3)
                     filter.float_data[:] = filter_data.flat
                     filter.dims[:] = filter_data.shape
+                    transposed_filter.add(op.input[1])
 
         return False