def run_sequential_model(model_fn, input_shape):
def get_input(shape):
_input = np.random.uniform(0, 1, shape).astype(dtype="float32")
return _input
def save_and_reload(_model):
with tempfile.TemporaryDirectory() as model_path:
tf.saved_model.save(_model, model_path)
loaded = tf.saved_model.load(model_path)
func = loaded.signatures["serving_default"]
frozen_func = convert_variables_to_constants_v2(func)
return frozen_func
def model_graph(model, input_shape):
_input = get_input(input_shape)
f = save_and_reload(model(input_shape))
_output = run_tf_code(f, _input)
gdef = f.graph.as_graph_def(add_shapes=True)
return gdef, _input, _output
compare_tf_tvm(*model_graph(model_fn, input_shape), runtime="vm")
def run_model_graph(TestClass):
compare_tf_tvm(*_model_graph(TestClass), runtime="vm")
def run_func_graph(TestClass, runtime="vm"):
compare_tf_tvm(*_function_graph(TestClass), runtime=runtime)
def run_model_graph(TestClass, outputs=None):
compare_tf_tvm(*_model_graph(TestClass), runtime="vm", output_tensors=outputs)
def run_func_graph(TestClass, runtime="vm", outputs=None):
compare_tf_tvm(*_function_graph(TestClass), runtime=runtime, output_tensors=outputs)