Ruslan NugmanovandEvgenya Stepyreva
4601b31bd0
* keras to tflite version of tests * Update tests/layer_tests/common/tf2_layer_test_class.py Co-authored-by: Evgenya Stepyreva <eva.my.link@gmail.com> * moved out tf utility functions from modules with tf_layer_test classes to tf_utils module * moved out tf utility functions from modules with tf_layer_test classes to tf_utils module and tflite_utils modules Co-authored-by: Evgenya Stepyreva <eva.my.link@gmail.com>
87 lines
3.5 KiB
Python
87 lines
3.5 KiB
Python
# Copyright (C) 2022 Intel Corporation
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
|
|
import os
|
|
|
|
from common.layer_test_class import CommonLayerTest
|
|
from common.utils.tflite_utils import get_tflite_results, save_tf2_saved_model_to_tflite
|
|
|
|
|
|
def save_to_tf2_savedmodel(tf2_model, path_to_saved_tf2_model):
|
|
import tensorflow as tf
|
|
assert int(tf.__version__.split('.')[0]) >= 2, "TensorFlow 2 must be used for this suite validation"
|
|
tf.keras.models.save_model(tf2_model, path_to_saved_tf2_model, save_format='tf')
|
|
assert os.path.isdir(path_to_saved_tf2_model), "the model haven't been saved " \
|
|
"here: {}".format(path_to_saved_tf2_model)
|
|
return path_to_saved_tf2_model
|
|
|
|
|
|
class CommonTF2LayerTest(CommonLayerTest):
|
|
input_model_key = "saved_model_dir"
|
|
|
|
def produce_model_path(self, framework_model, save_path):
|
|
if not getattr(self, 'tflite', False):
|
|
return save_to_tf2_savedmodel(framework_model, save_path)
|
|
else:
|
|
self.input_model_key = 'input_model'
|
|
tf2_saved_model = save_to_tf2_savedmodel(framework_model, save_path)
|
|
return save_tf2_saved_model_to_tflite(tf2_saved_model)
|
|
|
|
def get_framework_results(self, inputs_dict, model_path):
|
|
if not getattr(self, 'tflite', False):
|
|
return self.get_tf2_keras_results(inputs_dict, model_path)
|
|
else:
|
|
# get results from tflite
|
|
return get_tflite_results(self.use_new_frontend, self.use_old_api, inputs_dict, model_path)
|
|
|
|
def get_tf2_keras_results(self, inputs_dict, model_path):
|
|
import tensorflow as tf
|
|
import numpy as np
|
|
|
|
# load a model
|
|
loaded_model = tf.keras.models.load_model(model_path, custom_objects=None)
|
|
|
|
# prepare input
|
|
input_for_model = []
|
|
|
|
# order inputs based on input names in tests
|
|
# since TF2 Keras model accepts a list of tensors for prediction
|
|
for input_name in sorted(inputs_dict):
|
|
input_value = inputs_dict[input_name]
|
|
# convert NCHW to NHWC layout of tensor rank greater 3
|
|
if self.use_old_api:
|
|
if len(input_value.shape) == 4:
|
|
input_value = np.transpose(input_value, (0, 2, 3, 1))
|
|
elif len(input_value.shape) == 5:
|
|
input_value = np.transpose(input_value, (0, 2, 3, 4, 1))
|
|
input_for_model.append(input_value)
|
|
if len(input_for_model) == 1:
|
|
input_for_model = input_for_model[0]
|
|
|
|
# infer by original framework and complete a dictionary with reference results
|
|
tf_res_list = loaded_model(input_for_model)
|
|
if tf.is_tensor(tf_res_list):
|
|
tf_res_list = [tf_res_list]
|
|
else:
|
|
# in this case tf_res_list is a list of the single tuple of outputs
|
|
tf_res_list = tf_res_list[0]
|
|
|
|
result = dict()
|
|
for ind, tf_res in enumerate(tf_res_list):
|
|
if ind == 0:
|
|
output = "Identity"
|
|
else:
|
|
output = "Identity_{}".format(ind)
|
|
|
|
tf_res = tf_res.numpy()
|
|
tf_res_rank = len(tf_res.shape)
|
|
if self.use_old_api and tf_res_rank > 3:
|
|
# create axis order for NCHW layout
|
|
axis_order = np.arange(tf_res_rank)
|
|
axis_order = np.insert(axis_order, 1, axis_order[-1])
|
|
axis_order = np.delete(axis_order, [-1])
|
|
result[output] = tf_res.transpose(axis_order)
|
|
else:
|
|
result[output] = tf_res
|
|
return result
|