opencv/modules/dnn/test/test_tflite_importer.cpp

276 lines
10 KiB
C++
Raw Normal View History

// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
/*
Test for TFLite models loading
*/
#include "test_precomp.hpp"
#include "npy_blob.hpp"
#include <opencv2/dnn/layer.details.hpp> // CV_DNN_REGISTER_LAYER_CLASS
#include <opencv2/dnn/utils/debug_utils.hpp>
#include <opencv2/dnn/shape_utils.hpp>
#ifdef OPENCV_TEST_DNN_TFLITE
namespace opencv_test { namespace {
using namespace cv;
using namespace cv::dnn;
class Test_TFLite : public DNNTestLayer {
public:
void testModel(Net& net, const std::string& modelName, const Mat& input, double l1 = 0, double lInf = 0);
void testModel(const std::string& modelName, const Mat& input, double l1 = 0, double lInf = 0);
void testModel(const std::string& modelName, const Size& inpSize, double l1 = 0, double lInf = 0);
void testLayer(const std::string& modelName, double l1 = 0, double lInf = 0);
};
void testInputShapes(const Net& net, const std::vector<Mat>& inps) {
std::vector<MatShape> inLayerShapes;
std::vector<MatShape> outLayerShapes;
net.getLayerShapes(MatShape(), 0, inLayerShapes, outLayerShapes);
ASSERT_EQ(inLayerShapes.size(), inps.size());
for (int i = 0; i < inps.size(); ++i) {
ASSERT_EQ(inLayerShapes[i], shape(inps[i]));
}
}
void Test_TFLite::testModel(Net& net, const std::string& modelName, const Mat& input, double l1, double lInf)
{
l1 = l1 ? l1 : default_l1;
lInf = lInf ? lInf : default_lInf;
net.setPreferableBackend(backend);
net.setPreferableTarget(target);
testInputShapes(net, {input});
net.setInput(input);
std::vector<String> outNames = net.getUnconnectedOutLayersNames();
std::vector<Mat> outs;
net.forward(outs, outNames);
ASSERT_EQ(outs.size(), outNames.size());
for (int i = 0; i < outNames.size(); ++i) {
Mat ref = blobFromNPY(findDataFile(format("dnn/tflite/%s_out_%s.npy", modelName.c_str(), outNames[i].c_str())));
2024-04-06 21:55:17 +08:00
// A workaround solution for the following cases due to inconsistent shape definitions.
// The details please see: https://github.com/opencv/opencv/pull/25297#issuecomment-2039081369
if (modelName == "face_landmark" || modelName == "selfie_segmentation") {
ref = ref.reshape(1, 1);
outs[i] = outs[i].reshape(1, 1);
}
normAssert(ref, outs[i], outNames[i].c_str(), l1, lInf);
}
}
void Test_TFLite::testModel(const std::string& modelName, const Mat& input, double l1, double lInf)
{
Net net = readNet(findDataFile("dnn/tflite/" + modelName + ".tflite", false));
testModel(net, modelName, input, l1, lInf);
}
void Test_TFLite::testModel(const std::string& modelName, const Size& inpSize, double l1, double lInf)
{
Mat input = imread(findDataFile("cv/shared/lena.png"));
input = blobFromImage(input, 1.0 / 255, inpSize, 0, true);
testModel(modelName, input, l1, lInf);
}
void Test_TFLite::testLayer(const std::string& modelName, double l1, double lInf)
{
Mat inp = blobFromNPY(findDataFile("dnn/tflite/" + modelName + "_inp.npy"));
Net net = readNet(findDataFile("dnn/tflite/" + modelName + ".tflite"));
testModel(net, modelName, inp, l1, lInf);
}
// https://google.github.io/mediapipe/solutions/face_mesh
TEST_P(Test_TFLite, face_landmark)
{
if (backend == DNN_BACKEND_CUDA && target == DNN_TARGET_CUDA_FP16)
applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA_FP16);
double l1 = 2e-5, lInf = 2e-4;
if (target == DNN_TARGET_CPU_FP16 || target == DNN_TARGET_CUDA_FP16 || target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD ||
(backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target == DNN_TARGET_OPENCL))
{
l1 = 0.15;
lInf = 0.82;
}
testModel("face_landmark", Size(192, 192), l1, lInf);
}
// https://google.github.io/mediapipe/solutions/face_detection
TEST_P(Test_TFLite, face_detection_short_range)
{
double l1 = 0, lInf = 2e-4;
if (target == DNN_TARGET_CPU_FP16 || target == DNN_TARGET_CUDA_FP16 || target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD ||
(backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target == DNN_TARGET_OPENCL))
{
l1 = 0.04;
lInf = 0.8;
}
testModel("face_detection_short_range", Size(128, 128), l1, lInf);
}
// https://google.github.io/mediapipe/solutions/selfie_segmentation
TEST_P(Test_TFLite, selfie_segmentation)
{
double l1 = 0, lInf = 0;
if (target == DNN_TARGET_CPU_FP16 || target == DNN_TARGET_CUDA_FP16 || target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD ||
(backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target == DNN_TARGET_OPENCL))
{
l1 = 0.01;
lInf = 0.48;
}
testModel("selfie_segmentation", Size(256, 256), l1, lInf);
}
TEST_P(Test_TFLite, max_unpooling)
{
if (backend == DNN_BACKEND_CUDA)
applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA);
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2022010000)
if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_NGRAPH, CV_TEST_TAG_DNN_SKIP_IE_VERSION);
#endif
if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target != DNN_TARGET_CPU) {
if (target == DNN_TARGET_OPENCL_FP16) applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
if (target == DNN_TARGET_OPENCL) applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
if (target == DNN_TARGET_MYRIAD) applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
}
if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
applyTestTag(CV_TEST_TAG_DNN_SKIP_OPENCL_FP16);
// Due Max Unpoling is a numerically unstable operation and small difference between frameworks
// might lead to positional difference of maximal elements in the tensor, this test checks
// behavior of Max Unpooling layer only.
Net net = readNet(findDataFile("dnn/tflite/hair_segmentation.tflite", false));
net.setPreferableBackend(backend);
net.setPreferableTarget(target);
Mat input = imread(findDataFile("cv/shared/lena.png"));
cvtColor(input, input, COLOR_BGR2RGBA);
input = input.mul(Scalar(1, 1, 1, 0));
input = blobFromImage(input, 1.0 / 255);
testInputShapes(net, {input});
net.setInput(input);
std::vector<std::vector<Mat> > outs;
net.forward(outs, {"p_re_lu_1", "max_pooling_with_argmax2d", "conv2d_86", "max_unpooling2d_2"});
ASSERT_EQ(outs.size(), 4);
ASSERT_EQ(outs[0].size(), 1);
ASSERT_EQ(outs[1].size(), 2);
ASSERT_EQ(outs[2].size(), 1);
ASSERT_EQ(outs[3].size(), 1);
Mat poolInp = outs[0][0];
Mat poolOut = outs[1][0];
Mat poolIds = outs[1][1];
Mat unpoolInp = outs[2][0];
Mat unpoolOut = outs[3][0];
ASSERT_EQ(poolInp.size, unpoolOut.size);
ASSERT_EQ(poolOut.size, poolIds.size);
ASSERT_EQ(poolOut.size, unpoolInp.size);
ASSERT_EQ(countNonZero(poolInp), poolInp.total());
for (int c = 0; c < 32; ++c) {
float *poolInpData = poolInp.ptr<float>(0, c);
float *poolOutData = poolOut.ptr<float>(0, c);
float *poolIdsData = poolIds.ptr<float>(0, c);
float *unpoolInpData = unpoolInp.ptr<float>(0, c);
float *unpoolOutData = unpoolOut.ptr<float>(0, c);
for (int y = 0; y < 64; ++y) {
for (int x = 0; x < 64; ++x) {
int maxIdx = (y * 128 + x) * 2;
std::vector<int> indices{maxIdx + 1, maxIdx + 128, maxIdx + 129};
std::string errMsg = format("Channel %d, y: %d, x: %d", c, y, x);
for (int idx : indices) {
if (poolInpData[idx] > poolInpData[maxIdx]) {
EXPECT_EQ(unpoolOutData[maxIdx], 0.0f) << errMsg;
maxIdx = idx;
}
}
EXPECT_EQ(poolInpData[maxIdx], poolOutData[y * 64 + x]) << errMsg;
if (backend != DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) {
EXPECT_EQ(poolIdsData[y * 64 + x], (float)maxIdx) << errMsg;
}
EXPECT_EQ(unpoolOutData[maxIdx], unpoolInpData[y * 64 + x]) << errMsg;
}
}
}
}
TEST_P(Test_TFLite, EfficientDet_int8) {
Merge pull request #23987 from dkurt:openvino_int8_backend OpenVINO backend for INT8 models #23987 ### Pull Request Readiness Checklist TODO: - [x] DetectionOutput layer (https://github.com/opencv/opencv/pull/24069) - [x] Less FP32 fallbacks (i.e. Sigmoid, eltwise sum) - [x] Accuracy, performance tests (https://github.com/opencv/opencv/pull/24039) - [x] Single layer tests (convolution) - [x] ~~Fixes for OpenVINO 2022.1 (https://pullrequest.opencv.org/buildbot/builders/precommit_custom_linux/builds/100334)~~ Performace results for object detection model `coco_efficientdet_lite0_v1_1.0_quant_2021_09_06.tflite`: | backend | performance (median time) | |---|---| | OpenCV | 77.42ms | | OpenVINO 2023.0 | 10.90ms | CPU: `11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz` Serialized model per-layer stats (note that Convolution should use `*_I8` primitives if they are quantized correctly): https://gist.github.com/dkurt/7772bbf1907035441bb5454f19f0feef --- See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-09-28 21:24:43 +08:00
if (target != DNN_TARGET_CPU || (backend != DNN_BACKEND_OPENCV &&
backend != DNN_BACKEND_TIMVX && backend != DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)) {
throw SkipTestException("Only OpenCV, TimVX and OpenVINO targets support INT8 on CPU");
}
Net net = readNet(findDataFile("dnn/tflite/coco_efficientdet_lite0_v1_1.0_quant_2021_09_06.tflite", false));
net.setPreferableBackend(backend);
net.setPreferableTarget(target);
Mat img = imread(findDataFile("dnn/dog416.png"));
Mat blob = blobFromImage(img, 1.0, Size(320, 320));
net.setInput(blob);
Mat out = net.forward();
Mat_<float> ref({3, 7}, {
0, 7, 0.62890625, 0.6014542579650879, 0.13300055265426636, 0.8977657556533813, 0.292389452457428,
0, 17, 0.56640625, 0.15983937680721283, 0.35905322432518005, 0.5155506730079651, 0.9409466981887817,
0, 1, 0.5, 0.14357104897499084, 0.2240825891494751, 0.7183101177215576, 0.9140362739562988
});
normAssertDetections(ref, out, "", 0.5, 0.05, 0.1);
}
TEST_P(Test_TFLite, replicate_by_pack) {
double l1 = 0, lInf = 0;
if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target == DNN_TARGET_OPENCL)
{
l1 = 4e-4;
lInf = 2e-3;
}
testLayer("replicate_by_pack", l1, lInf);
}
TEST_P(Test_TFLite, split) {
testLayer("split");
}
TEST_P(Test_TFLite, fully_connected) {
if (backend == DNN_BACKEND_VKCOM)
applyTestTag(CV_TEST_TAG_DNN_SKIP_VULKAN);
testLayer("fully_connected");
}
Merge pull request #25297 from CNOCycle:tflite/transpose Support Transpose op in TFlite #25297 **Merge with extra**: https://github.com/opencv/opencv_extra/pull/1168 The purpose of this PR is to introduce support for the Transpose op in TFlite format and to add a shape comparison between the output tensors and the references. In some occasional cases, the shape of the output tensor is `[1,4,1,1]`, while the shape of the reference tensor is `[1,4]`. Consequently, the norm check incorrectly reports that the test has passed, as the residual is zero. Below is a Python script for generating testing data. The generated data can be integrated into the repo `opencv_extra`. ```python import numpy as np import tensorflow as tf PREFIX_TFL = '/path/to/opencv_extra/testdata/dnn/tflite/' def generator(input_tensor, model, saved_name): # convert keras model to .tflite format converter = tf.lite.TFLiteConverter.from_keras_model(model) #converter.optimizations = [tf.lite.Optimize.DEFAULT] converter.optimizations = [None] tflite_model = converter.convert() with open(f'{PREFIX_TFL}/{saved_name}.tflite', 'wb') as f: f.write(tflite_model) # save the input tensor to .npy if input_tensor.ndim == 4: opencv_tensor = np.transpose(input_tensor, (0,3,1,2)) else: opencv_tensor = input_tensor opencv_tensor = np.copy(opencv_tensor, order='C').astype(np.float32) np.save(f'{PREFIX_TFL}/{saved_name}_inp.npy', opencv_tensor) # generate output tenosr and save it to .npy mat_out = model(input_tensor).numpy() mat_out = np.copy(mat_out, order='C').astype(np.float32) if mat_out.ndim == 4: mat_out = np.transpose(mat_out, (0,3,1,2)) interpreter = tf.lite.Interpreter(model_content=tflite_model) out_name = interpreter.get_output_details()[0]['name'] np.save(f'{PREFIX_TFL}/{saved_name}_out_{out_name}.npy', mat_out) def build_transpose(): model_name = "keras_permute" mat_in = np.array([[[1,2,3], [4,5,6]]], dtype=np.float32) model = tf.keras.Sequential() model.add(tf.keras.Input(shape=(2,3))) model.add(tf.keras.layers.Permute((2,1))) model.summary() generator(mat_in, model, model_name) if __name__ == '__main__': build_transpose() ``` ### Pull Request Readiness Checklist - [x] I agree to contribute to the project under Apache 2 License. - [X] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [X] The PR is proposed to the proper branch - [ ] There is a reference to the original bug report and related work - [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [X] The feature is well documented and sample code can be built with the project CMake
2024-05-16 01:07:25 +08:00
TEST_P(Test_TFLite, permute) {
testLayer("permutation_3d");
// Temporarily disabled as TFLiteConverter produces a incorrect graph in this case
//testLayer("permutation_4d_0123");
testLayer("permutation_4d_0132");
testLayer("permutation_4d_0213");
testLayer("permutation_4d_0231");
}
TEST_P(Test_TFLite, global_average_pooling_2d) {
testLayer("global_average_pooling_2d");
}
TEST_P(Test_TFLite, global_max_pooling_2d) {
testLayer("global_max_pooling_2d");
}
INSTANTIATE_TEST_CASE_P(/**/, Test_TFLite, dnnBackendsAndTargets());
}} // namespace
#endif // OPENCV_TEST_DNN_TFLITE