Keras to tflite version of tests (#15042)
* 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>
This commit is contained in:
Ruslan Nugmanov
GitHub
parent
9fe4db56fe
commit
4601b31bd0
|
|
@ -77,9 +77,10 @@ class CommonLayerTest:
|
|||
# assert flag, '\n'.join(resp)
|
||||
|
||||
config = None
|
||||
# GPU default execution precision is FP16, so if we want to check FP32 inference we need to set explicit precision hint
|
||||
# GPU default execution precision is FP16, so if we want to check FP32 inference
|
||||
# we need to set explicit precision hint
|
||||
if ie_device == 'GPU' and precision == 'FP32':
|
||||
config = {'INFERENCE_PRECISION_HINT' : 'f32'}
|
||||
config = {'INFERENCE_PRECISION_HINT': 'f32'}
|
||||
|
||||
if self.use_old_api:
|
||||
ie_engine = IEInfer(model=path_to_xml,
|
||||
|
|
|
|||
|
|
@ -4,6 +4,7 @@
|
|||
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):
|
||||
|
|
@ -19,9 +20,21 @@ class CommonTF2LayerTest(CommonLayerTest):
|
|||
input_model_key = "saved_model_dir"
|
||||
|
||||
def produce_model_path(self, framework_model, save_path):
|
||||
return save_to_tf2_savedmodel(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
|
||||
|
||||
|
|
|
|||
|
|
@ -1,59 +1,11 @@
|
|||
# Copyright (C) 2018-2022 Intel Corporation
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
import os
|
||||
|
||||
from common.layer_test_class import CommonLayerTest
|
||||
from common.utils.tf_utils import summarize_graph
|
||||
|
||||
|
||||
def transpose_nchw_to_nhwc(data, use_new_frontend, use_old_api):
|
||||
if use_new_frontend or not use_old_api:
|
||||
return data
|
||||
|
||||
if len(data.shape) == 4: # reshaping for 4D tensors
|
||||
return data.transpose(0, 2, 3, 1)
|
||||
elif len(data.shape) == 5: # reshaping for 5D tensors
|
||||
return data.transpose(0, 2, 3, 4, 1)
|
||||
else:
|
||||
return data
|
||||
|
||||
|
||||
def transpose_nhwc_to_nchw(data, use_new_frontend, use_old_api):
|
||||
if use_new_frontend or not use_old_api:
|
||||
return data
|
||||
|
||||
if len(data.shape) == 4: # reshaping for 4D tensors
|
||||
return data.transpose(0, 3, 1, 2) # 2, 0, 1
|
||||
elif len(data.shape) == 5: # reshaping for 5D tensors
|
||||
return data.transpose(0, 4, 1, 2, 3) # 3, 0, 1, 2
|
||||
else:
|
||||
return data
|
||||
|
||||
|
||||
def save_to_pb(tf_model, path_to_saved_tf_model):
|
||||
import tensorflow as tf
|
||||
tf.io.write_graph(tf_model, path_to_saved_tf_model, 'model.pb', False)
|
||||
assert os.path.isfile(os.path.join(path_to_saved_tf_model, 'model.pb')), "model.pb haven't been saved " \
|
||||
"here: {}".format(path_to_saved_tf_model)
|
||||
return os.path.join(path_to_saved_tf_model, 'model.pb')
|
||||
|
||||
|
||||
def save_pb_to_tflite(pb_model):
|
||||
import tensorflow as tf
|
||||
|
||||
graph_summary = summarize_graph(pb_model)
|
||||
inputs = [k for k in graph_summary['inputs'].keys()]
|
||||
outputs = graph_summary['outputs']
|
||||
|
||||
converter = tf.compat.v1.lite.TFLiteConverter.from_frozen_graph(pb_model, inputs, outputs)
|
||||
tflite_model = converter.convert()
|
||||
|
||||
tflite_model_path = os.path.join(os.path.dirname(pb_model), 'model.tflite')
|
||||
with tf.io.gfile.GFile(tflite_model_path, 'wb') as f:
|
||||
f.write(tflite_model)
|
||||
|
||||
return tflite_model_path
|
||||
from common.utils.tflite_utils import get_tflite_results, save_pb_to_tflite
|
||||
from common.utils.tf_utils import save_to_pb, transpose_nhwc_to_nchw, transpose_nchw_to_nhwc
|
||||
|
||||
|
||||
class CommonTFLayerTest(CommonLayerTest):
|
||||
|
|
@ -99,33 +51,6 @@ class CommonTFLayerTest(CommonLayerTest):
|
|||
self.use_old_api)
|
||||
return result
|
||||
|
||||
def get_tflite_results(self, inputs_dict, model_path):
|
||||
import tensorflow as tf
|
||||
interpreter = tf.compat.v1.lite.Interpreter(model_path=model_path)
|
||||
interpreter.allocate_tensors()
|
||||
input_details = interpreter.get_input_details()
|
||||
output_details = interpreter.get_output_details()
|
||||
|
||||
input_name_to_id_mapping = {input['name']: input['index'] for input in input_details}
|
||||
|
||||
for layer, data in inputs_dict.items():
|
||||
tensor_index = input_name_to_id_mapping[layer]
|
||||
tensor_id = next(i for i, tensor in enumerate(input_details) if tensor['index'] == tensor_index)
|
||||
interpreter.set_tensor(input_details[tensor_id]['index'], data)
|
||||
|
||||
interpreter.invoke()
|
||||
tf_result = dict()
|
||||
for output in output_details:
|
||||
tf_result[output['name']] = interpreter.get_tensor(output['index'])
|
||||
|
||||
result = dict()
|
||||
for out in tf_result.keys():
|
||||
_tf_res = tf_result[out]
|
||||
result[out] = transpose_nhwc_to_nchw(_tf_res, self.use_new_frontend,
|
||||
self.use_old_api)
|
||||
|
||||
return tf_result
|
||||
|
||||
def get_framework_results(self, inputs_dict, model_path):
|
||||
if not getattr(self, 'tflite', False):
|
||||
# prepare inputs
|
||||
|
|
@ -134,4 +59,4 @@ class CommonTFLayerTest(CommonLayerTest):
|
|||
return self.get_tf_results(inputs_dict, model_path)
|
||||
else:
|
||||
# get results from tflite
|
||||
return self.get_tflite_results(inputs_dict, model_path)
|
||||
return get_tflite_results(self.use_new_frontend, self.use_old_api, inputs_dict, model_path)
|
||||
|
|
|
|||
|
|
@ -9,6 +9,7 @@ import tensorflow as tf
|
|||
|
||||
from openvino.tools.mo.ops.op import PermuteAttrs
|
||||
|
||||
|
||||
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
|
||||
|
||||
|
||||
|
|
@ -135,3 +136,34 @@ def permute_nchw_to_nhwc(shape, use_new_frontend=False):
|
|||
|
||||
def permute_axis(axis, permutation_inv):
|
||||
return permutation_inv[axis]
|
||||
|
||||
|
||||
def transpose_nchw_to_nhwc(data, use_new_frontend, use_old_api):
|
||||
if use_new_frontend or not use_old_api:
|
||||
return data
|
||||
|
||||
if len(data.shape) == 4: # reshaping for 4D tensors
|
||||
return data.transpose(0, 2, 3, 1)
|
||||
elif len(data.shape) == 5: # reshaping for 5D tensors
|
||||
return data.transpose(0, 2, 3, 4, 1)
|
||||
else:
|
||||
return data
|
||||
|
||||
|
||||
def transpose_nhwc_to_nchw(data, use_new_frontend, use_old_api):
|
||||
if use_new_frontend or not use_old_api:
|
||||
return data
|
||||
|
||||
if len(data.shape) == 4: # reshaping for 4D tensors
|
||||
return data.transpose(0, 3, 1, 2) # 2, 0, 1
|
||||
elif len(data.shape) == 5: # reshaping for 5D tensors
|
||||
return data.transpose(0, 4, 1, 2, 3) # 3, 0, 1, 2
|
||||
else:
|
||||
return data
|
||||
|
||||
|
||||
def save_to_pb(tf_model, path_to_saved_tf_model):
|
||||
tf.io.write_graph(tf_model, path_to_saved_tf_model, 'model.pb', False)
|
||||
assert os.path.isfile(os.path.join(path_to_saved_tf_model, 'model.pb')), "model.pb haven't been saved " \
|
||||
"here: {}".format(path_to_saved_tf_model)
|
||||
return os.path.join(path_to_saved_tf_model, 'model.pb')
|
||||
|
|
|
|||
|
|
@ -0,0 +1,60 @@
|
|||
import os
|
||||
|
||||
import tensorflow as tf
|
||||
from common.utils.tf_utils import summarize_graph, transpose_nhwc_to_nchw
|
||||
|
||||
|
||||
def save_pb_to_tflite(pb_model):
|
||||
graph_summary = summarize_graph(pb_model)
|
||||
inputs = [k for k in graph_summary['inputs'].keys()]
|
||||
outputs = graph_summary['outputs']
|
||||
|
||||
converter = tf.compat.v1.lite.TFLiteConverter.from_frozen_graph(pb_model, inputs, outputs)
|
||||
tflite_model = converter.convert()
|
||||
|
||||
tflite_model_path = os.path.join(os.path.dirname(pb_model), 'model.tflite')
|
||||
with tf.io.gfile.GFile(tflite_model_path, 'wb') as f:
|
||||
f.write(tflite_model)
|
||||
|
||||
return tflite_model_path
|
||||
|
||||
|
||||
def get_tflite_results(use_new_frontend, use_old_api, inputs_dict, model_path):
|
||||
interpreter = tf.compat.v1.lite.Interpreter(model_path=model_path)
|
||||
interpreter.allocate_tensors()
|
||||
input_details = interpreter.get_input_details()
|
||||
output_details = interpreter.get_output_details()
|
||||
|
||||
input_name_to_id_mapping = {input['name']: input['index'] for input in input_details}
|
||||
|
||||
for layer, data in inputs_dict.items():
|
||||
tensor_index = input_name_to_id_mapping[layer]
|
||||
tensor_id = next(i for i, tensor in enumerate(input_details) if tensor['index'] == tensor_index)
|
||||
interpreter.set_tensor(input_details[tensor_id]['index'], data)
|
||||
|
||||
interpreter.invoke()
|
||||
tf_lite_result = dict()
|
||||
for output in output_details:
|
||||
tf_lite_result[output['name']] = interpreter.get_tensor(output['index'])
|
||||
|
||||
result = dict()
|
||||
for out in tf_lite_result.keys():
|
||||
_tf_res = tf_lite_result[out]
|
||||
result[out] = transpose_nhwc_to_nchw(_tf_res, use_new_frontend,
|
||||
use_old_api)
|
||||
|
||||
return tf_lite_result
|
||||
|
||||
|
||||
def save_tf2_saved_model_to_tflite(savedmodel):
|
||||
import tensorflow as tf
|
||||
|
||||
converter = tf.compat.v1.lite.TFLiteConverter.from_saved_model(savedmodel)
|
||||
tflite_model = converter.convert()
|
||||
|
||||
tflite_model_path = os.path.join(os.path.dirname(savedmodel), 'model.tflite')
|
||||
with tf.io.gfile.GFile(tflite_model_path, 'wb') as f:
|
||||
f.write(tflite_model)
|
||||
|
||||
return tflite_model_path
|
||||
|
||||
|
|
@ -12,6 +12,7 @@ from common.utils.common_utils import copy_files_by_pattern
|
|||
def pytest_generate_tests(metafunc):
|
||||
test_gen_attrs_names = list(inspect.signature(get_params).parameters)
|
||||
params = get_params()
|
||||
setattr(metafunc.cls, 'tflite', metafunc.config.getoption('tflite'))
|
||||
metafunc.parametrize(test_gen_attrs_names, params, scope="function")
|
||||
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue