def test_generic_array(generic_array):
tf_value = TensorProtoConverter.get_tf_value(generic_array)
assert type(
tf_value).__name__ == TensorProtoConverter.__tfproto_type__.__name__
generic = TensorProtoConverter.get_generic_value(tf_value)
assert isinstance(generic, type(generic_array))
assert (generic.np_array == generic_array.np_array).all()
def _build_param_ops(self, onnx_graph, ugraph, op_types_cnt,
tensor_names_map):
"""Find all tensors in initialization list in onnx_graph, normally constants
Note that this method will update op_types_cnt and tensor_names_map **inplace**
"""
# find Const ops
params_dict = {}
# FIXME: avoid using internal api of other library
dict_items = _onnx_initializer_to_input_dict_items(
onnx_graph.initializer)
for name, tf_tensor in dict_items:
params_dict[name] = AttrValueConverter.GenericType(
value_name='value',
value=TensorProtoConverter.get_generic_value(
tf_tensor.op.get_attr('value')))
# build Const ops
for tensor_name, tensor_value in params_dict.items():
cnt = op_types_cnt['Const']
node_name = self._format_node_name(tensor_name, 'Const', cnt)
op_types_cnt['Const'] += 1
tensor_names_map[tensor_name] = TensorInfo(
name=self._format_tensor_name('', node_name, 0),
op_name=node_name,
dtype=tensor_value.value.dtype,
shape=list(tensor_value.value.np_array.shape),
ugraph=ugraph)
OperationInfo(name=node_name,
input_tensors=[],
output_tensors=[tensor_names_map[tensor_name]],
op_type='Const',
lib_name='onnx',
ugraph=ugraph,
op_attr={'value': tensor_value})
def test_tf_tensor_quint8(tf_quint8_tensor):
np_array = TensorProtoConverter.get_generic_value(tf_quint8_tensor)
assert np_array.dtype[0] == np.dtype('uint8')
tf_value = TensorProtoConverter.get_tf_value(np_array)
assert tf_value.tensor_content == tf_quint8_tensor.tensor_content
assert tf_value.tensor_shape == tf_quint8_tensor.tensor_shape
def generic_array():
np_array = np.random.randn(3, 3).astype(np.float32)
return TensorProtoConverter.__utensor_generic_type__(np_array=np_array)