Opencv: c++ opencv4.01 called tensorflow inception-v4 model ,readNetFromTensorflow ERROR!

[JaosonMa]

train a tf flowers model with inception-v4,
after got the mode.ckpy , use the code to convert it:
echo "create model.pb over"
cd ..
CUDA_VISIBLE_DEVICES=1 python3 -u export_inference_graph.py \
--model_name=inception_v4 \
--output_file=./data/flowers_5/models/infer/flowers.pb \
--dataset_name=flowers \
--dataset_dir=./data/flowers_5

echo "start create frzee pb"

CUDA_VISIBLE_DEVICES=1 python3 -u /usr/local/lib/python3.5/dist-packages/tensorflow/python/tools/freeze_graph.py \
--input_graph=./data/flowers_5/models/infer/flowers.pb \
--input_checkpoint=./data/flowers_5/models/model.ckpt-10027 \
--output_graph=./data/flowers_5/models/infer/my_freeze.pb \
--input_binary=True \
--output_node_name=InceptionV4/Logits/Predictions

i test the flowers.pb with classify.sh (https://github.com/isiosia/models/tree/lession/slim)
it works well

id:[0] name:[daisy] (score = 0.80188)
id:[1] name:[dandelion] (score = 0.11308)
id:[3] name:[sunflowers] (score = 0.05036)
id:[4] name:[tulips] (score = 0.02128)
id:[2] name:[roses] (score = 0.01340)

then i use it for vs2015+opencv4.01, with the code

then got the error

load models
OpenCV: terminate handler is called! The last OpenCV error is:
OpenCV(4.0.1) Error: Assertion failed (scaleMat.type() == CV_32FC1) in cv::dnn::dnn4_v20181221::`anonymous-namespace'::TFImporter::populateNet, file E:\product\opencv-4.0.1\modules\dnn\src\tensorflow\tf_importer.cpp, line 1377

so i dont know where was wrong?
how can i make it work in c++ opoencv4.01?

@dkurt

[alalek]

Usage questions should go to Users OpenCV Q/A forum: http://answers.opencv.org

If you think this is a bug which should be fixed - provide complete reproducer including models.

[JaosonMa]

Usage questions should go to Users OpenCV Q/A forum: http://answers.opencv.org

If you think this is a bug which should be fixed - provide complete reproducer including models.

i have done this, but there don have the A,

[alalek]

We really don't have spare time to download datasets and try to train models using outdated instructions (you point on commits which are modified 2+ years ago).

Reproducer should have:

• OpenCV C++ code
• data/images processed by OpenCV code
• no more external dependencies.

Ask for help on http://answers.opencv.org if you are not able to provide this.

---

Could you try add this before mentioned assertion?
.diff +std::cout << typeToString(scaleMat.type()) << std::endl; +scaleMat.convertTo(scaleMat, CV_32FC1); CV_Assert(scaleMat.type() == CV_32FC1);

[JaosonMa]

std::cout << typeToString(scaleMat.type()) << std::endl;
+scaleMat.convertTo(scaleMat, CV_32FC1);

i tried the code ,

here is the result before change and after change with debug

before change:

after change:

here is my pb. baidu ，i don 't know can you download or not ,
链接：https://pan.baidu.com/s/1d9nWjKV_NCLFuzReOP4WKQ
提取码：94sx

c++ code

include "pch.h"

include

include

include

include

include

include

include

using namespace cv;
using namespace cv::dnn;
using namespace std;
//自己新建一个txt文件，写入分类的标签（一行写一个标签，例如二分类，第一行写good，第二行bad）
string labels_txt_file = R"(D:\downloads\googledownloads\labels.txt)";
string tf_pb_file = R"(D:\downloads\googledownloads\my_freeze.pb)";
vector readClassNames();
void main()
{
Mat src = imread(R"(D:\downloads\googledownloads\437859108_173fb33c98.jpg)");
if (src.empty()) {
cout << "error:no img" << endl;
}
vector labels = readClassNames();
Mat rgb;
int w = 299;
int h = 299;
resize(src, src, Size(w, h));
cvtColor(src, rgb, COLOR_BGR2RGB);
cout << "load models" << endl;
Net net = readNetFromTensorflow(tf_pb_file);
cout << "load over" << endl;
DWORD timestart = GetTickCount();
if (net.empty())
{
cout << "error:no model" << endl;
}
Mat inputBlob = blobFromImage(src, 0.00390625f, Size(w, h), Scalar(), true, false);
//inputBlob -= 117.0;
//执行图像分类
Mat prob;
net.setInput(inputBlob, "input");
prob = net.forward("output");
cout << prob << endl;
//prob=net.forward("softmax2");
//得到最大分类概率
Mat probMat = prob.reshape(1, 1);
Point classNumber;
double classProb;
minMaxLoc(probMat, NULL, &classProb, NULL, &classNumber);
DWORD timeend = GetTickCount();
int classidx = classNumber.x;
printf("\n current image classification : %s, possible : %.2f\n", labels.at(classidx).c_str(), classProb);
cout << "用时(毫秒):" << timeend - timestart << endl;
// 显示文本
putText(src, labels.at(classidx), Point(20, 20), FONT_HERSHEY_SIMPLEX, 1.0, Scalar(0, 0, 255), 2, 8);
imshow("Image Classfication", src);
waitKey(0);

}
vector readClassNames()
{
vector classNames;
fstream fp(labels_txt_file);
if (!fp.is_open())
{
cout << "does not open" << endl;
exit(-1);
}
string name;
while (!fp.eof())
{
getline(fp, name);
if (name.length())
classNames.push_back(name);
}
fp.close();
return classNames;
}

[JaosonMa]

here is the layers in my_freeze.pb

input

DecodePng

eval_image/convert_image/Cast

eval_image/convert_image/y

eval_image/convert_image

eval_image/central_crop/Shape_1

eval_image/central_crop/strided_slice/stack

eval_image/central_crop/strided_slice/stack_1

eval_image/central_crop/strided_slice/stack_2

eval_image/central_crop/strided_slice

eval_image/central_crop/Shape_2

eval_image/central_crop/strided_slice_1/stack

eval_image/central_crop/strided_slice_1/stack_1

eval_image/central_crop/strided_slice_1/stack_2

eval_image/central_crop/strided_slice_1

eval_image/central_crop/ToDouble

eval_image/central_crop/mul/y

eval_image/central_crop/mul

eval_image/central_crop/sub

eval_image/central_crop/truediv/y

eval_image/central_crop/truediv

eval_image/central_crop/ToInt32

eval_image/central_crop/ToDouble_1

eval_image/central_crop/mul_1/y

eval_image/central_crop/mul_1

eval_image/central_crop/sub_1

eval_image/central_crop/truediv_1/y

eval_image/central_crop/truediv_1

eval_image/central_crop/ToInt32_1

eval_image/central_crop/mul_2/y

eval_image/central_crop/mul_2

eval_image/central_crop/sub_2

eval_image/central_crop/mul_3/y

eval_image/central_crop/mul_3

eval_image/central_crop/sub_3

eval_image/central_crop/stack/2

eval_image/central_crop/stack

eval_image/central_crop/stack_1/2

eval_image/central_crop/stack_1

eval_image/central_crop/Slice

eval_image/ExpandDims/dim

eval_image/ExpandDims

eval_image/ResizeBilinear/size

eval_image/ResizeBilinear

eval_image/Squeeze

eval_image/Sub/y

eval_image/Sub

eval_image/Mul/y

eval_image/Mul

ExpandDims/dim

ExpandDims

InceptionV4/Conv2d_1a_3x3/weights

InceptionV4/Conv2d_1a_3x3/weights/read

InceptionV4/InceptionV4/Conv2d_1a_3x3/Conv2D

InceptionV4/InceptionV4/Conv2d_1a_3x3/BatchNorm/Const

InceptionV4/Conv2d_1a_3x3/BatchNorm/beta

InceptionV4/Conv2d_1a_3x3/BatchNorm/beta/read

InceptionV4/Conv2d_1a_3x3/BatchNorm/moving_mean

InceptionV4/Conv2d_1a_3x3/BatchNorm/moving_mean/read

InceptionV4/Conv2d_1a_3x3/BatchNorm/moving_variance

InceptionV4/Conv2d_1a_3x3/BatchNorm/moving_variance/read

InceptionV4/InceptionV4/Conv2d_1a_3x3/BatchNorm/FusedBatchNorm

InceptionV4/InceptionV4/Conv2d_1a_3x3/Relu

InceptionV4/Conv2d_2a_3x3/weights

InceptionV4/Conv2d_2a_3x3/weights/read

InceptionV4/InceptionV4/Conv2d_2a_3x3/Conv2D

InceptionV4/InceptionV4/Conv2d_2a_3x3/BatchNorm/Const

InceptionV4/Conv2d_2a_3x3/BatchNorm/beta

InceptionV4/Conv2d_2a_3x3/BatchNorm/beta/read

InceptionV4/Conv2d_2a_3x3/BatchNorm/moving_mean

InceptionV4/Conv2d_2a_3x3/BatchNorm/mo
...
InceptionV4/Logits/PreLogitsFlatten/flatten/strided_slice/stack

InceptionV4/Logits/PreLogitsFlatten/flatten/strided_slice/stack_1

InceptionV4/Logits/PreLogitsFlatten/flatten/strided_slice/stack_2

InceptionV4/Logits/PreLogitsFlatten/flatten/strided_slice

InceptionV4/Logits/PreLogitsFlatten/flatten/Reshape/shape/1

InceptionV4/Logits/PreLogitsFlatten/flatten/Reshape/shape

InceptionV4/Logits/PreLogitsFlatten/flatten/Reshape

InceptionV4/Logits/Logits/weights

InceptionV4/Logits/Logits/weights/read

InceptionV4/Logits/Logits/biases

InceptionV4/Logits/Logits/biases/read

InceptionV4/Logits/Logits/MatMul

InceptionV4/Logits/Logits/BiasAdd

InceptionV4/Logits/Predictions
@alalek @dkurt

[JaosonMa]

here is my train.sh with tensorflow's slim API:
cd ..
CUDA_VISIBLE_DEVICES=1 python3 train_image_classifier.py \
--dataset_name=flowers \
--dataset_dir=./data/flowers_5 \
--model_name=inception_v4 \
--train_dir=./data/flowers_5/models \
--learning_rate=0.001 \
--learning_rate_decay_factor=0.76 \
--num_epochs_per_decay=50 \
--moving_average_decay=0.9999 \
--optimizer=adam \
--ignore_missing_vars=True \
--batch_size=32

[mshabunin]

i have done this, but there don have the A,

I can not find your question on answers.opencv.org? Please provide a link.

[JaosonMa]

i have done this, but there don have the A,

I can not find your question on answers.opencv.org? Please provide a link.

sorry , i mean i don't find the A on answers.opencv.org? and i can not sing in or create an username,
so i have not ask this Q on answers.opecn.org , so sorry

[JaosonMa]

OpenCV: terminate handler is called! The last OpenCV error is:
OpenCV(4.0.1) Error: Unspecified error (Input [eval_image/central_crop/stack] for node [eval_image/central_crop/Slice] not found) in cv::dnn::dnn4_v20181221::`anonymous-namespace'::TFImporter::getConstBlob, file E:\product\opencv-4.0.1\modules\dnn\src\tensorflow\tf_importer.cpp, line 535

[dkurt]

@JaosonMa, Please follow issue's template and make it reproducible. Provide a reference to model so we can test it locally.

[JaosonMa]

System information (version)

• OpenCV => :4.01:
• Operating System / Platform => win10 64bite
• Compiler => :vs2017 x86:

Detailed description

i train a model with a tensorflow slim, inception-v4 with code like this
CUDA_VISIBLE_DEVICES=1 python3 train_image_classifier.py \ --dataset_name=flowers \ --dataset_dir=./data/flowers_5 \ --model_name=inception_v4 \ --train_dir=./data/flowers_5/models \ --learning_rate=0.001 \ --learning_rate_decay_factor=0.76 \ --num_epochs_per_decay=50 \ --moving_average_decay=0.9999 \ --optimizer=adam \ --ignore_missing_vars=True \ --batch_size=32
1> i got a tf-model save as model.ckpt-xxx
2> freeze the model to .pb file with the code like this
`echo "create model.pb start"
CUDA_VISIBLE_DEVICES=1 python3 -u export_inference_graph.py \
--model_name=inception_v4 \
--output_file=./data/flowers_5/models/infer_isiosia/flowers.pb \
--dataset_name=flowers \
--dataset_dir=./data/flowers_5

echo "start create frzee pb"

CUDA_VISIBLE_DEVICES=1 python3 -u /usr/local/lib/python3.5/dist-packages/tensorflow/python/tools/freeze_graph.py \
--input_graph=./data/flowers_5/models/infer_isiosia/flowers.pb \
--input_checkpoint=./data/flowers_5/models/model.ckpt-10027 \
--output_graph=./data/flowers_5/models/infer_isiosia/my_freeze.pb \
--input_binary=True \
--output_node_name=InceptionV4/Logits/Predictions
`
3> use the my_freeze.pb to opencv4.01 + vs2017 with code like this

first i got the error , like this

then add some code in openv-dnn module, as @alalek say
https://github.com/opencv/opencv/issues/14149#issuecomment-476475899
then got the error like this

OpenCV: terminate handler is called! The last OpenCV error is:
OpenCV(4.0.1) Error: Unspecified error (Input [eval_image/central_crop/stack] for node [eval_image/central_crop/Slice] not found) in cv::dnn::dnn4_v20181221::`anonymous-namespace'::TFImporter::getConstBlob, file E:\product\opencv-4.0.1\modules\dnn\src\tensorflow\tf_importer.cpp, line 535

Steps to reproduce

'
using namespace cv;
using namespace cv::dnn;
using namespace std;
string labels_txt_file = R"(D:\downloads\googledownloads\labels.txt)";
string tf_pb_file = R"(D:\downloads\googledownloads\my_freeze.pb)";
vector readClassNames();
void main()
{
Mat src = imread(R"(D:\downloads\googledownloads\437859108_173fb33c98.jpg)");
if (src.empty()) {
cout << "error:no img" << endl;
}
vector labels = readClassNames();
Mat rgb;
int w = 299;
int h = 299;
resize(src, src, Size(w, h));
cvtColor(src, rgb, COLOR_BGR2RGB);
cout << "load models" << endl;
Net net = readNetFromTensorflow(tf_pb_file);
cout << "load over" << endl;
DWORD timestart = GetTickCount();
if (net.empty())
...
}
'
pb.file
link: (https://pan.baidu.com/s/14nIYO6yeHaPCDrRCXRsuGQ)
pwd:imzb

@dkurt
forgive me for my bad english !!

[dkurt]

@JaosonMa, please do not use Baidu for sharing. We just cannot download anything from there because it downloads .exe files! Please provide a direct link.

[JaosonMa]

@JaosonMa, please do not use Baidu for sharing. We just cannot download anything from there because it downloads .exe files! Please provide a direct link.

sorry, i can not google , and i can not find some good ways ,that can just provide a direct link , can you give me a email or some other ways... so sorry for thie

[JaosonMa]

@JaosonMa, please do not use Baidu for sharing. We just cannot download anything from there because it downloads .exe files! Please provide a direct link.

sorry, i can not google , and i can not find some good ways ,that can just provide a direct link , can you give me a email or some other ways... so sorry for thie

can you tell me , my steps was right or not,
1> use tensorflow - slim train a model ( https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/preprocessing_factory.py )saved as model.ckpt
2>, use export_inference_graph.py export the model.pb file
3>. use freeze_graph.py ceate the freeze.pb
does opencv 4.01 can load the the freeze.pb direct or not ?
did i miss some steps?

[dkurt]

@JaosonMa, please try Dropbox at least.

[JaosonMa]

@JaosonMa, please try Dropbox at least.
ok ,i will give a shoot.
i also tried this mobilenet_v2_1.0_224

does the mobilenet_v2_1.4_224_frozen.pb can used by opencv4.01 direct?

[JaosonMa]

@JaosonMa, please try Dropbox at least.

i faild , my internet can not use DropBox

[dkurt]

@JaosonMa, Thanks for your issue! Please test the changes from PR https://github.com/opencv/opencv/pull/14166. It'll be merged into 3.4 and then ported to master branch.

[JaosonMa]

@JaosonMa, Thanks for your issue! Please test the changes from PR #14166. It'll be merged into 3.4 and then ported to master branch.

i change these codes in my opencv4.01 source code

then rebuild the opencvdnn modules . try two times with diffient codes:

and mobilenet_v2 SUCC!

But inceptionv-v4 Failed!

i closed this , i think i must do something wrong with the inception-v4, i will keep trying....

[JaosonMa]

@dkurt @alalek
I tried many many times ,but always failed! ,so please help me again!
i ran the mobilenet_v2_1.0_224 sucess!
so i train the data again with mobilenet_v2_1.0_224 on my data,
after training, got the ckpt,
`cd ..

Where the pre-trained MobilenetV1 checkpoint is saved to.

PRETRAINED_CHECKPOINT_DIR=./data/flowers_2/models/all

Where the training (fine-tuned) checkpoint and logs will be saved to.

TRAIN_DIR=/tmp/flowers-models/mobilenet_v1_1.0_224

Where the dataset is saved to.

DATASET_DIR=./data/flowers_2
DATASET_NAME=flowers
INFER_DIR=./data/flowers_2/models/all/infer
MODEL_NAME=mobilenet_v2_140

echo "create model.pb start"
CUDA_VISIBLE_DEVICES=0 python3 -u export_inference_graph.py \
--model_name=${MODEL_NAME} \
--output_file=${INFER_DIR}/flowers.pb \
--dataset_name=${DATASET_NAME} \
--dataset_dir=${DATASET_DIR}

echo "start create frzee pb"

CUDA_VISIBLE_DEVICES=0 python3 -u /usr/local/lib/python3.5/dist-packages/tensorflow/python/tools/freeze_graph.py \
--input_graph=${INFER_DIR}/flowers.pb \
--input_checkpoint=${PRETRAINED_CHECKPOINT_DIR}/model.ckpt-44885 \
--output_graph=${INFER_DIR}/my_freeze.pb \
--input_binary=True \
--output_node_name=MobilenetV2/Predictions/Reshape_1
`
export_inference_graph.pu

then i got the flowers.pb
and the my_freeze.pb

the use the same code ,the got error :

it seems like that some layers was not found in opencv4.01,so i check the two pbfile(my_freeze.pb and mobilenet_v2_1.0_224_frozen.pb) with code like this:

`mobile_net = '../data/mobilenet_v2_1.4_224/mobilenet_v2_1.4_224_frozen.pb'
my_mobile_net = '../data/flowers_2/models/all/infer/my_freeze.pb'
def create_graph():
with tf.gfile.FastGFile(my_mobile_net, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
tf.import_graph_def(graph_def, name='')
create_graph()
tensor_name_list_my = [tensor.name for tensor in tf.get_default_graph().as_graph_def().node]

print(len(tensor_name_list))
print(len(tensor_name_list_my))
for idx,tensor_name in enumerate(tensor_name_list_my):
if(idx print(idx+1,"-->",tensor_name,idx+1,"-->",tensor_name_list[idx],'\n')
else:
print(idx+1,"-->",tensor_name,'\n')
`
the length is not same ,

...

...

so i think some thing was wrong with my export_inference_graph.py, can you tell me how can i got the pb just as same as the tf's mobilenet_v1_1.0_224_.pb,

[dkurt]

@JaosonMa, before the neural network this graph has a lot of support nodes (jpeg decoding, cropping, resizing). It's might be done out of the network (i.e. by blobFromImage). So you can remove these nodes:

before:

import tensorflow as tf
from tensorflow.tools.graph_transforms import TransformGraph

with tf.gfile.FastGFile('my_freeze.pb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())

graph_def = TransformGraph(graph_def, ["ExpandDims"], ["MobilenetV2/Predictions/Reshape_1"], ["strip_unused_nodes"])
with tf.gfile.FastGFile("graph.pb", 'wb') as f:
    f.write(graph_def.SerializeToString())

after:

[JaosonMa]

first thank you a lot!!!
i try the code you give me, error like this :


so i change from
with tf.gfile.FastGFile(my_mobile_net) as f: to with tf.gfile.FastGFile(my_mobile_net,'rb') as f:
then it run succ
got the frozen_graph.pb

use frozen_graph.pb in opencv4.01 ,got error:

so i print the nodes , like this:

so which is the input?
` Net net = readNetFromTensorflow(tf_pb_file);
Mat src = imread(R"(D:\downloads\googledownloads\dc165458a6955e67399f2186f90cccc9.jpg)");

Mat rgb;
int w = 224;
int h = 224;
resize(src, src, Size(w, h));
Mat inputBlob = blobFromImage(src, 0.00390625f, Size(w, h), Scalar(), true, false);
Mat prob;
net.setInput(inputBlob, "input");  // what should i put in there?
prob = net.forward("MobilenetV2/Predictions/Reshape_1");`

net.setInput(inputBlob, "input"); what should i put in there?

@dkurt

[JaosonMa]

Mat prob; net.setInput(inputBlob, "ExpandDims"); prob = net.forward("MobilenetV2/Predictions/Reshape_1");

i put ExpandDims in ，it run succ!
i think it is right!

[JaosonMa]

@dkurt So sorry to disturb you again.
I used the pb with input:ExpanDims run succ in opencv4.01,
but the acc is lower than the my_freeze.pb in tensorflow code ,
the acc when before TransformGraph:

my data only have two classes, They are somewhat similar. so acc=0.9 It's almost what I expected.
but used the pb after TranfromGraph , the acc is only 0.65, i don't know why, i check three graphs in tensorboard ,1> tensorflow's slim mobilenet_v2_1.4_224.pb 2> my_freeze.pb 3>frozen_graph.pb
the mobilenetV2 part is all the same

but there input layer was diffient ,
1> mobilenet_v2_1.4_224.pb

2>my_freeze.pb

3>froze_graph.pb

can i just got a pb , just like the mobilenet_v2_1.4_224.pb with the input "input",
Is it possible that this input caused the reduction?

[dkurt]

@JaosonMa, Please provide two code snippets: how you run the model with TensorFlow and with OpenCV. Please make it as minimal as possible: remove dataset reading. Just a single image.

[JaosonMa]

@dkurt
tensorflow:
classify code: classify.py
my_freeze.pb:
class.sh:
labels.txt:

c++:
classfi.cpp:
frozen_graph.pb:
labels.txt: same as tenflow labels.txt

test_img:

this image is one of the val data , its label is 'N'
in my computer :
tensorflow output is

C++ outPut is

it is diffient!

[JaosonMa]

@dkurt Is this enough for you to debug?What else do you need?I will certainly give it to you.

[JaosonMa]

@dkurt please help me , My project has been delayed for 10 days.

[dkurt]

@JaosonMa, please stop it. We already figured out that the problem might be fixed with current version of OpenCV. For now the question is to align TensorFlow and OpenCV usage. Actually, the problem in your code. So be patient please.

[JaosonMa]

got you , so sorry for what i have done , i will just wait .

[dkurt]

@JaosonMa, There are two nodes which impact on different predictions from OpenCV and TensorFlow: DecodeJpeg and ResizeBilinear. The thing is that precision of Jpeg decoding is not similar even between different versions of decoding libraries. ResizeBilinear's interpolation strategy differs from OpenCV's one.

So to achieve similar outputs you need to pass similar inputs to the node after these ones:

import numpy as np
import cv2 as cv

import tensorflow as tf
from tensorflow.tools.graph_transforms import TransformGraph

image = cv.imread("2_1_25_3_1.jpg")
image = image[:,:,[2,1,0]]  # BGB 2 RGB
image = cv.resize(image, (224, 224))
paddedImage = cv.copyMakeBorder(image, 15, 15, 15, 15, cv.BORDER_CONSTANT)

#
# TensorFlow
#
with tf.gfile.FastGFile("my_freeze.pb", 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())

with tf.Session() as sess:
    sess.graph.as_default()
    tf.import_graph_def(graph_def, name='')

    tfOut = sess.run(sess.graph.get_tensor_by_name('MobilenetV2/Predictions/Reshape_1:0'),
                     feed_dict={'eval_image/convert_image/Cast:0': paddedImage})
    print tfOut

#
# OpenCV
#
net = cv.dnn.readNet("graph.pb")
net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)

blob = cv.dnn.blobFromImage(image, 1.0/127.5, (224, 224), [127.5, 127.5, 127.5])
net.setInput(blob)
cvOut = net.forward()

print cvOut

Central crop looks also a bit confusing (it crops by a 0.875 ratio). So I resized an image out of the graph and added zero paddings so both Central Crop and ResizeBilinear don't affect an accuracy:

TF
[[1.7407431e-01 8.2585084e-01 3.5822773e-05 3.9013106e-05]]
CV
[[1.7408027e-01 8.2584488e-01 3.5821628e-05 3.9013088e-05]]

To achieve better similarity, you man try to define ExpandDims layer as a custom one (read a tutorial https://docs.opencv.org/master/dc/db1/tutorial_dnn_custom_layers.html) so ResizeBilinear will be inside a graph.

Please use a forum for future usage questions: http://answers.opencv.org

[bnascimento]

Hey @dkurt, can you clarify the reason to the 15 pixel padding?

[dkurt]

@bnascimento, I guess the same padding (which depends on image size) is done inside the TensorFlow graph (please visualize it with TensorBoard and show the nodes from the beginning up to eval_image/convert_image/Cast:.

[memeda2232]

@JaosonMa hi brother, i also met this problem,Have you already solve it?