def testFlexWithDoubleOp(self):
# Create a graph that has one double op.
saved_model_dir = os.path.join(self.get_temp_dir(), 'model2')
with ops.Graph().as_default():
with session.Session() as sess:
in_tensor = array_ops.placeholder(shape=[1, 4],
dtype=dtypes.int32,
name='input')
out_tensor = double_op.double(in_tensor)
inputs = {'x': in_tensor}
outputs = {'z': out_tensor}
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
converter = lite.TFLiteConverterV2.from_saved_model(saved_model_dir)
converter.target_spec.supported_ops = set([lite.OpsSet.SELECT_TF_OPS])
converter.target_spec.experimental_select_user_tf_ops = ['Double']
tflite_model = converter.convert()
self.assertTrue(tflite_model)
self.assertIn('FlexDouble',
tflite_test_util.get_ops_list(tflite_model))
# Check the model works with TensorFlow ops.
interpreter = Interpreter(model_content=tflite_model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
test_input = np.array([[1.0, 2.0, 3.0, 4.0]], dtype=np.int32)
interpreter.set_tensor(input_details[0]['index'], test_input)
interpreter.invoke()
output_details = interpreter.get_output_details()
expected_output = np.array([[2.0, 4.0, 6.0, 8.0]], dtype=np.int32)
output_data = interpreter.get_tensor(output_details[0]['index'])
self.assertTrue((expected_output == output_data).all())
def _createSavedModelWithCustomOp(self):
custom_opdefs_str = (
'name: \'CustomAdd\' input_arg: {name: \'Input1\' type: DT_FLOAT} '
'input_arg: {name: \'Input2\' type: DT_FLOAT} output_arg: {name: '
'\'Output\' type: DT_FLOAT}')
# Create a graph that has one add op.
new_graph = graph_pb2.GraphDef()
with ops.Graph().as_default():
with session.Session() as sess:
in_tensor = array_ops.placeholder(shape=[1, 16, 16, 3],
dtype=dtypes.float32,
name='input')
out_tensor = in_tensor + in_tensor
inputs = {'x': in_tensor}
outputs = {'z': out_tensor}
new_graph.CopyFrom(sess.graph_def)
# Rename Add op name to CustomAdd.
for node in new_graph.node:
if node.op.startswith('Add'):
node.op = 'CustomAdd'
del node.attr['T']
# Register custom op defs to import modified graph def.
register_custom_opdefs([custom_opdefs_str])
# Store saved model.
saved_model_dir = self._getFilepath('model')
with ops.Graph().as_default():
with session.Session() as sess:
import_graph_def(new_graph, name='')
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
return (saved_model_dir, custom_opdefs_str)
def _createSimpleSavedModel(self, shape):
"""Create a simple savedmodel on the fly."""
saved_model_dir = os.path.join(self.get_temp_dir(), "simple_savedmodel")
with session.Session() as sess:
in_tensor = array_ops.placeholder(shape=shape, dtype=dtypes.float32)
out_tensor = in_tensor + in_tensor
inputs = {"x": in_tensor}
outputs = {"y": out_tensor}
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
return saved_model_dir
def _createSimpleSavedModel(self, shape):
"""Create a simple SavedModel on the fly."""
saved_model_dir = os.path.join(self.get_temp_dir(), "simple_savedmodel")
with session.Session() as sess:
in_tensor = array_ops.placeholder(shape=shape, dtype=dtypes.float32)
out_tensor = in_tensor + in_tensor
inputs = {"x": in_tensor}
outputs = {"y": out_tensor}
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
return saved_model_dir
def test_need_flex_ops(self):
def create_graph_with_custom_add(opname='CustomAdd'):
custom_opdefs_str = (
'name: \'' + opname +
'\' input_arg: {name: \'Input1\' type: DT_FLOAT} '
'input_arg: {name: \'Input2\' type: DT_FLOAT} output_arg: {name: '
'\'Output\' type: DT_FLOAT}')
# Create a graph that has one add op.
new_graph = graph_pb2.GraphDef()
with ops.Graph().as_default():
with session.Session() as sess:
in_tensor = array_ops.placeholder(
shape=[1, 16, 16, 3], dtype=dtypes.float32, name='input')
out_tensor = in_tensor + in_tensor
inputs = {'x': in_tensor}
outputs = {'z': out_tensor}
new_graph.CopyFrom(sess.graph_def)
# Rename Add op name to opname.
for node in new_graph.node:
if node.op.startswith('Add'):
node.op = opname
del node.attr['T']
# Register custom op defs to import modified graph def.
register_custom_opdefs([custom_opdefs_str])
return (new_graph, inputs, outputs)
new_graph, inputs, outputs = create_graph_with_custom_add()
# Import to load the custom opdef.
saved_model_dir = os.path.join(self.get_temp_dir(), 'model')
with ops.Graph().as_default():
with session.Session() as sess:
import_graph_def(new_graph, name='')
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
converter = lite.TFLiteConverterV2.from_saved_model(saved_model_dir)
self.convert_and_check_location_info(
converter,
converter_error_data_pb2.ConverterErrorData.NAMELOC,
expected_sources='add')
exported_error = metrics._gauge_conversion_errors.get_cell(
'CONVERT_TF_TO_TFLITE_MODEL', 'CONVERT_SAVED_MODEL', 'tf.CustomAdd',
'ERROR_NEEDS_CUSTOM_OPS').value()
self.assertEqual(
exported_error,
"\'tf.CustomAdd\' op is neither a custom op nor a flex op\n"
"Error code: ERROR_NEEDS_CUSTOM_OPS"
)
def testFloat(self):
saved_model_dir = os.path.join(self.get_temp_dir(), 'simple_savedmodel')
with session.Session().as_default() as sess:
in_tensor_1 = array_ops.placeholder(
shape=[1, 16, 16, 3], dtype=dtypes.float32, name='inputB')
in_tensor_2 = array_ops.placeholder(
shape=[1, 16, 16, 3], dtype=dtypes.float32, name='inputA')
out_tensor = in_tensor_1 + in_tensor_2
inputs = {'x': in_tensor_1, 'y': in_tensor_2}
outputs = {'z': out_tensor}
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
model_coverage.test_saved_model(saved_model_dir)
def testSavedModel(self):
saved_model_dir = self._getFilepath('model')
with ops.Graph().as_default():
with session.Session() as sess:
in_tensor = array_ops.placeholder(
shape=[1, 16, 16, 3], dtype=dtypes.float32, name='inputB')
out_tensor = in_tensor + in_tensor
inputs = {'x': in_tensor}
outputs = {'z': out_tensor}
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
flags_str = '--saved_model_dir={}'.format(saved_model_dir)
self._run(flags_str, should_succeed=True)
def _createSavedModelTwoInputArrays(self, shape):
"""Create a simple SavedModel."""
saved_model_dir = os.path.join(self.get_temp_dir(), "simple_savedmodel")
with session.Session() as sess:
in_tensor_1 = array_ops.placeholder(
shape=shape, dtype=dtypes.float32, name="inputB")
in_tensor_2 = array_ops.placeholder(
shape=shape, dtype=dtypes.float32, name="inputA")
out_tensor = in_tensor_1 + in_tensor_2
inputs = {"x": in_tensor_1, "y": in_tensor_2}
outputs = {"z": out_tensor}
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
return saved_model_dir
def _createSavedModel(self, shape):
"""Create a simple SavedModel."""
saved_model_dir = os.path.join(self.get_temp_dir(), 'simple_savedmodel')
with session.Session() as sess:
in_tensor_1 = array_ops.placeholder(
shape=shape, dtype=dtypes.float32, name='inputB')
in_tensor_2 = array_ops.placeholder(
shape=shape, dtype=dtypes.float32, name='inputA')
out_tensor = in_tensor_1 + in_tensor_2
inputs = {'x': in_tensor_1, 'y': in_tensor_2}
outputs = {'z': out_tensor}
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
return saved_model_dir
def testFloat(self):
saved_model_dir = os.path.join(self.get_temp_dir(), 'simple_savedmodel')
with session.Session().as_default() as sess:
in_tensor_1 = array_ops.placeholder(
shape=[1, 16, 16, 3], dtype=dtypes.float32, name='inputB')
in_tensor_2 = array_ops.placeholder(
shape=[1, 16, 16, 3], dtype=dtypes.float32, name='inputA')
out_tensor = in_tensor_1 + in_tensor_2
inputs = {'x': in_tensor_1, 'y': in_tensor_2}
outputs = {'z': out_tensor}
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
model_coverage.test_saved_model(saved_model_dir)
def _createSavedModelTwoInputArrays(self, shape):
"""Create a simple SavedModel."""
saved_model_dir = os.path.join(self.get_temp_dir(), "simple_savedmodel")
with session.Session() as sess:
in_tensor_1 = array_ops.placeholder(
shape=shape, dtype=dtypes.float32, name="inputB")
in_tensor_2 = array_ops.placeholder(
shape=shape, dtype=dtypes.float32, name="inputA")
out_tensor = in_tensor_1 + in_tensor_2
inputs = {"x": in_tensor_1, "y": in_tensor_2}
outputs = {"z": out_tensor}
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
return saved_model_dir
def _createSavedModel(self, shape):
"""Create a simple SavedModel."""
saved_model_dir = os.path.join(self.get_temp_dir(), 'simple_savedmodel')
with session.Session() as sess:
in_tensor_1 = array_ops.placeholder(
shape=shape, dtype=dtypes.float32, name='inputB')
in_tensor_2 = array_ops.placeholder(
shape=shape, dtype=dtypes.float32, name='inputA')
out_tensor = in_tensor_1 + in_tensor_2
inputs = {'x': in_tensor_1, 'y': in_tensor_2}
outputs = {'z': out_tensor}
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
return saved_model_dir
def _createV1SavedModel(self, shape):
"""Create a simple SavedModel."""
saved_model_dir = os.path.join(self.get_temp_dir(), 'simple_savedmodel')
with tf.Graph().as_default():
with tf.compat.v1.Session() as sess:
in_tensor_1 = tf.compat.v1.placeholder(
shape=shape, dtype=tf.float32, name='inputB')
in_tensor_2 = tf.compat.v1.placeholder(
shape=shape, dtype=tf.float32, name='inputA')
variable_node = tf.Variable(1.0, name='variable_node')
out_tensor = in_tensor_1 + in_tensor_2 * variable_node
inputs = {'x': in_tensor_1, 'y': in_tensor_2}
outputs = {'z': out_tensor}
sess.run(tf.compat.v1.variables_initializer([variable_node]))
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
return saved_model_dir
def testFlexWithCustomOp(self):
new_graph, inputs, outputs = self._createGraphWithCustomOp(
opname='CustomAdd4')
# Import to load the custom opdef.
saved_model_dir = os.path.join(self.get_temp_dir(), 'model')
with ops.Graph().as_default():
with session.Session() as sess:
import_graph_def(new_graph, name='')
saved_model.simple_save(sess, saved_model_dir, inputs, outputs)
converter = lite.TFLiteConverterV2.from_saved_model(saved_model_dir)
converter.target_spec.supported_ops = set([lite.OpsSet.SELECT_TF_OPS])
converter.target_spec.experimental_select_user_tf_ops = ['CustomAdd4']
tflite_model = converter.convert()
self.assertIn('FlexCustomAdd4', tflite_test_util.get_ops_list(tflite_model))
