def convert_quantize_layer(self, keras_op):
op = self._mace_net_def.op.add()
op.name = keras_op.name
op.type = MaceOp.Identity.name
op.input.append(get_input(keras_op).name)
op.output.append(get_output(keras_op).name)
output_shape = op.output_shape.add()
output_shape.dims.extend(keras_shape2list(get_output(keras_op).shape))
ConverterUtil.add_data_type_arg(op, mace_pb2.DT_FLOAT)
ConverterUtil.add_data_format_arg(op, DataFormat.NHWC)
output_min = keras_op.weights[0].numpy()
output_max = keras_op.weights[1].numpy()
self.add_quantize_info(op, output_min, output_max)
return op
def convert(model_file, output_dir, layers):
mace_check(os.path.isdir(output_dir),
"Output directory '" + output_dir + "' does not exist!")
multi_net_def = init_multi_net_def(model_file)
multi_net_info = MultiNetDefInfo(multi_net_def, layers)
output_configs = {ModelKeys.subgraphs: []}
while multi_net_info.index_valid():
# omit BatchToSpaceND and op before that due to changed graph
cur_op = multi_net_info.get_current_op()
next_op = multi_net_info.get_next_op()
if cur_op.type == MaceOp.BatchToSpaceND.name or \
cur_op.type == HexagonOp.BatchToSpaceND_8.name or \
(cur_op.type == MaceOp.Quantize.name or
cur_op.type == HexagonOp.QuantizeINPUT_f_to_8.name or
cur_op.type == HexagonOp.INPUT.name) or \
(cur_op.type == HexagonOp.DequantizeOUTPUT_8tof.name or
cur_op.type == HexagonOp.OUTPUT.name) or \
(next_op is not None and
(next_op.type == MaceOp.BatchToSpaceND.name or
next_op.type == HexagonOp.BatchToSpaceND_8.name)) or \
cur_op.name.startswith(MaceKeyword.mace_output_node_name):
multi_net_info.StartIndexIncrement()
continue
multi_net = copy.deepcopy(multi_net_def)
net_defs = multi_net.net_def
# remove unused net_def
net = None
cur_net_idx = multi_net_info.get_current_net_idx()
for net_def in net_defs[:]:
if net_def.infer_order == cur_net_idx:
net = net_def
elif net_def.infer_order > cur_net_idx:
net_defs.remove(net_def)
del multi_net.output_tensor[:]
data_format = net.output_info[0].data_format
# remove unsued op
cur_op_idx = multi_net_info.get_current_op_idx()
is_hexagon = multi_net_info.is_hexagon(cur_net_idx)
if is_hexagon:
# reuse dequantize op and it's min/max tensor's node_id
del net.op[(cur_op_idx + 1):-1]
else:
del net.op[(cur_op_idx + 1):]
del net.output_info[:]
op = net.op[cur_op_idx]
multi_net_info.StartIndexIncrement()
output_tensors = []
output_shapes = []
op_name = op.name
if str(op.name).startswith(MaceKeyword.mace_output_node_name):
continue
is_quantize = multi_net_info.is_quantize(cur_net_idx)
if is_quantize:
op.name = MaceKeyword.mace_output_node_name + '_' + op.name
if is_hexagon:
if len(op.output) != 1:
print("Skip %s(%s)" % (op.name, op.type))
continue
for i in range(len(op.output)):
output_tensors.append(str(op.output[i]))
output_shapes.append(",".join(
[str(dim) for dim in op.output_shape[i].dims]))
# modify output info
multi_net.output_tensor.append(op.output[i])
output_info = net.output_info.add()
output_info.name = op.output[i]
output_info.data_format = data_format
output_info.dims.extend(op.output_shape[i].dims)
output_info.data_type = mace_pb2.DT_FLOAT
if is_quantize:
output_info.scale = op.quantize_info[0].scale
output_info.zero_point = op.quantize_info[0].zero_point
# modify output op
if is_quantize:
output_name = op.output[i]
new_output_name = \
MaceKeyword.mace_output_node_name + '_' + op.output[i]
op.output[i] = new_output_name
if not is_hexagon:
dequantize_op = net.op.add()
dequantize_op.name = normalize_op_name(output_name)
dequantize_op.type = MaceOp.Dequantize.name
dequantize_op.input.append(new_output_name)
dequantize_op.output.append(output_name)
output_shape = dequantize_op.output_shape.add()
output_shape.dims.extend(op.output_shape[i].dims)
dequantize_op.output_type.append(mace_pb2.DT_FLOAT)
ConverterUtil.add_data_type_arg(dequantize_op,
mace_pb2.DT_UINT8)
else:
dequantize_op = net.op[-1]
dequantize_op.name = normalize_op_name(output_name)
del dequantize_op.input[:]
del dequantize_op.output[:]
dequantize_op.input.append(new_output_name)
dequantize_op.node_input[0].node_id = op.node_id
dequantize_op.output.append(output_name)
if dequantize_op.type == \
HexagonOp.DequantizeOUTPUT_8tof.name:
input_min = new_output_name[:-1] + '1'
input_max = new_output_name[:-1] + '2'
dequantize_op.input.extend([input_min, input_max])
dequantize_op.node_input[1].node_id = op.node_id
dequantize_op.node_input[2].node_id = op.node_id
del dequantize_op.node_input[3:]
else:
del dequantize_op.node_input[1:]
model_path = save_model_to_proto(multi_net, normalize_op_name(op_name),
output_dir)
output_config = {
ModelKeys.model_file_path: str(model_path),
ModelKeys.output_tensors: output_tensors,
ModelKeys.output_shapes: output_shapes,
"output_data_formats": [DataFormat.NHWC.name] * len(output_shapes)
}
output_configs[ModelKeys.subgraphs].append(output_config)
output_configs_path = output_dir + "outputs.yml"
with open(output_configs_path, "w") as f:
yaml.dump(output_configs, f, default_flow_style=False)
def convert(model_file, output_dir, layers):
mace_check(os.path.isfile(model_file),
"Input graph file '" + model_file + "' does not exist!")
mace_check(os.path.isdir(output_dir),
"Output directory '" + output_dir + "' does not exist!")
net_def = mace_pb2.NetDef()
with open(model_file, "rb") as f:
net_def.ParseFromString(f.read())
is_quantize = ConverterUtil.get_arg(net_def,
MaceKeyword.mace_quantize_flag_arg_str)
is_quantize = False if is_quantize is None else is_quantize.i == 1
is_hexagon = False
index = 0
end_index = len(net_def.op)
if is_quantize:
while index < end_index:
# omit op quantize
if net_def.op[index].type == MaceOp.Quantize.name or \
net_def.op[index].type == \
HexagonOp.QuantizeINPUT_f_to_8.name or \
net_def.op[index].type == HexagonOp.INPUT.name:
index += 1
# omit op dequantize
elif net_def.op[end_index - 1].type == MaceOp.Dequantize.name or \
net_def.op[end_index - 1].type == \
HexagonOp.DequantizeOUTPUT_8tof.name or \
net_def.op[end_index - 1].type == HexagonOp.OUTPUT.name:
end_index -= 1
else:
break
mace_check(
0 < index < end_index < len(net_def.op),
"Wrong number of op quantize(%d) or dequantize(%d)." %
(index, len(net_def.op) - end_index))
if net_def.op[-1].type == HexagonOp.DequantizeOUTPUT_8tof.name or \
net_def.op[-1].type == HexagonOp.OUTPUT.name:
is_hexagon = True
index, end_index = handle_index(index, end_index, layers)
data_format = net_def.output_info[0].data_format
output_configs = {"subgraphs": []}
while index < end_index:
# omit BatchToSpaceND and op before that due to changed graph
if net_def.op[index].type == MaceOp.BatchToSpaceND.name or \
net_def.op[index].type == HexagonOp.BatchToSpaceND_8.name or \
(index + 1 < end_index and
(net_def.op[index + 1].type == MaceOp.BatchToSpaceND.name or
net_def.op[index + 1].type == HexagonOp.BatchToSpaceND_8.name)): # noqa
index += 1
continue
net = copy.deepcopy(net_def)
if is_hexagon:
# reuse dequantize op and it's min/max tensor's node_id
del net.op[index + 1:-1]
else:
del net.op[index + 1:]
del net.output_info[:]
op = net.op[index]
index += 1
output_tensors = []
output_shapes = []
op_name = op.name
if str(op.name).startswith(MaceKeyword.mace_output_node_name):
continue
if is_quantize:
op.name = MaceKeyword.mace_output_node_name + '_' + op.name
if is_hexagon:
if len(op.output) != 1:
print("Skip %s(%s)" % (op.name, op.type))
continue
for i in range(len(op.output)):
output_tensors.append(str(op.output[i]))
output_shapes.append(",".join(
[str(dim) for dim in op.output_shape[i].dims]))
# modify output info
output_info = net.output_info.add()
output_info.name = op.output[i]
output_info.data_format = data_format
output_info.dims.extend(op.output_shape[i].dims)
output_info.data_type = mace_pb2.DT_FLOAT
if is_quantize:
output_info.scale = op.quantize_info[0].scale
output_info.zero_point = op.quantize_info[0].zero_point
# modify output op
if is_quantize:
output_name = op.output[i]
new_output_name = \
MaceKeyword.mace_output_node_name + '_' + op.output[i]
op.output[i] = new_output_name
if not is_hexagon:
dequantize_op = net.op.add()
dequantize_op.name = normalize_op_name(output_name)
dequantize_op.type = MaceOp.Dequantize.name
dequantize_op.input.append(new_output_name)
dequantize_op.output.append(output_name)
output_shape = dequantize_op.output_shape.add()
output_shape.dims.extend(op.output_shape[i].dims)
dequantize_op.output_type.append(mace_pb2.DT_FLOAT)
ConverterUtil.add_data_type_arg(dequantize_op,
mace_pb2.DT_UINT8)
else:
dequantize_op = net.op[-1]
dequantize_op.name = normalize_op_name(output_name)
del dequantize_op.input[:]
del dequantize_op.output[:]
dequantize_op.input.append(new_output_name)
dequantize_op.node_input[0].node_id = op.node_id
dequantize_op.output.append(output_name)
if dequantize_op.type == \
HexagonOp.DequantizeOUTPUT_8tof.name:
input_min = new_output_name[:-1] + '1'
input_max = new_output_name[:-1] + '2'
dequantize_op.input.extend([input_min, input_max])
dequantize_op.node_input[1].node_id = op.node_id
dequantize_op.node_input[2].node_id = op.node_id
del dequantize_op.node_input[3:]
else:
del dequantize_op.node_input[1:]
model_path = save_model_to_proto(net, normalize_op_name(op_name),
output_dir)
output_config = {
"model_file_path": str(model_path),
"output_tensors": output_tensors,
"output_shapes": output_shapes,
"output_data_formats": [DataFormat.NHWC.name]
}
output_configs["subgraphs"].append(output_config)
output_configs_path = output_dir + "outputs.yml"
with open(output_configs_path, "w") as f:
yaml.dump(output_configs, f, default_flow_style=False)
