def test_compile_simple_fp16(self):
with TemporaryDirectory() as model_dir:
_save_saved_model_file(model_dir)
compiled_1 = compiler.compile_source(
source=SavedModelFile(model_path=model_dir),
config=Config(input_formats=[]))
compiled_2 = compiler.compile_source(
source=SavedModelFile(model_path=model_dir),
config=Config(optimization=True,
supported_types=[tf.float16],
input_formats=['']))
self.assertLess(len(compiled_1.tflite_model),
len(compiled_2.tflite_model))
def test_compile_select_tf_op(self):
with TemporaryDirectory() as model_dir:
_save_saved_model_file(model_dir)
compiled = compiler.compile_source(
source=SavedModelFile(model_path=model_dir),
config=Config(supported_ops=[tf.lite.OpsSet.SELECT_TF_OPS],
input_formats=['channels_first']))
self.assertIsInstance(compiled.tflite_model, bytes)
def test_compile_simple(self):
with TemporaryDirectory() as model_dir:
_save_saved_model_file(model_dir)
compiled = compiler.compile_source(
source=SavedModelFile(model_path=model_dir),
config=Config(
input_formats=['channels_first', 'channels_last']))
self.assertIsInstance(compiled.tflite_model, bytes)
def test_from_json(self):
self.assertEqual(Config(input_names=[], input_formats=[]),
Config.from_json({}))
self.assertEqual(Config(input_names=['x', 'y'], input_formats=[]),
Config.from_json({'input_names': ['x', 'y']}))
self.assertEqual(
Config(input_names=['x', 'y'], input_formats=['channels_first']),
Config.from_json({
'input_names': ['x', 'y'],
'input_formats': ['channels_first']
}))
self.assertEqual(
Config(input_names=['x', 'y'],
input_formats=['channels_first', 'channels_last']),
Config.from_json({
'input_names': ['x', 'y'],
'input_formats': ['channels_first', 'channels_last']
}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=False,
supported_types=None), Config.from_json({}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=False,
supported_types=None),
Config.from_json({'optimization': False}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=True,
supported_types=None),
Config.from_json({'optimization': True}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=False,
supported_types=[tf.float16]),
Config.from_json({'supported_types': ['float16']}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=False,
supported_types=[tf.float16, tf.float32]),
Config.from_json({'supported_types': ['float16', 'float32']}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=True,
supported_types=[tf.float16, tf.float32]),
Config.from_json({
'optimization': True,
'supported_types': ['float16', 'float32']
}))
self.assertEqual(
Config(input_names=[], input_formats=[], supported_ops=None),
Config.from_json({}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
supported_ops=[tf.lite.OpsSet.TFLITE_BUILTINS_INT8]),
Config.from_json({'supported_ops': ['TFLITE_BUILTINS_INT8']}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
supported_ops=[
tf.lite.OpsSet.SELECT_TF_OPS,
tf.lite.OpsSet.TFLITE_BUILTINS_INT8
]),
Config.from_json(
{'supported_ops': ['SELECT_TF_OPS', 'TFLITE_BUILTINS_INT8']}))
self.assertEqual(
Config(input_names=[], input_formats=[],
inference_input_type=None), Config.from_json({}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
inference_input_type=tf.float32),
Config.from_json({'inference_input_type': 'float32'}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
inference_output_type=None), Config.from_json({}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
inference_output_type=tf.float32),
Config.from_json({'inference_output_type': 'float32'}))
def test_invalid_data_type(self):
with self.assertRaises(ValueError):
Config.from_json({'supported_types': ['foobar']})
with self.assertRaises(ValueError):
Config.from_json({'supported_types': ['as_dtype']})
def test_from_env(self):
self.assertEqual(Config(input_names=[], input_formats=[]),
Config.from_env({}))
self.assertEqual(Config(input_names=['x', 'y'], input_formats=[]),
Config.from_env({'INPUT_NAMES': 'x,y'}))
self.assertEqual(
Config(input_names=['x', 'y'], input_formats=['channels_first']),
Config.from_env({
'INPUT_NAMES': 'x,y',
'INPUT_FORMATS': 'channels_first'
}))
self.assertEqual(
Config(input_names=['x', 'y'],
input_formats=['channels_first', 'channels_last']),
Config.from_env({
'INPUT_NAMES': 'x,y',
'INPUT_FORMATS': 'channels_first,channels_last'
}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=False,
supported_types=None), Config.from_env({}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=False,
supported_types=None),
Config.from_env({'OPTIMIZATION': '0'}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=True,
supported_types=None),
Config.from_env({'OPTIMIZATION': '1'}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=False,
supported_types=[tf.float16]),
Config.from_env({'SUPPORTED_TYPES': 'float16'}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=False,
supported_types=[tf.float16, tf.float32]),
Config.from_env({'SUPPORTED_TYPES': 'float16,float32'}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
optimization=True,
supported_types=[tf.float16, tf.float32]),
Config.from_env({
'OPTIMIZATION': '1',
'SUPPORTED_TYPES': 'float16,float32'
}))
self.assertEqual(
Config(input_names=[], input_formats=[], supported_ops=None),
Config.from_env({}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
supported_ops=[tf.lite.OpsSet.TFLITE_BUILTINS_INT8]),
Config.from_env({'SUPPORTED_OPS': 'TFLITE_BUILTINS_INT8'}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
supported_ops=[
tf.lite.OpsSet.SELECT_TF_OPS,
tf.lite.OpsSet.TFLITE_BUILTINS_INT8
]),
Config.from_env(
{'SUPPORTED_OPS': 'SELECT_TF_OPS,TFLITE_BUILTINS_INT8'}))
self.assertEqual(
Config(input_names=[], input_formats=[],
inference_input_type=None), Config.from_env({}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
inference_input_type=tf.float32),
Config.from_env({'INFERENCE_INPUT_TYPE': 'float32'}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
inference_output_type=None), Config.from_env({}))
self.assertEqual(
Config(input_names=[],
input_formats=[],
inference_output_type=tf.float32),
Config.from_env({'INFERENCE_OUTPUT_TYPE': 'float32'}))
