Error Loading Tensorflow Frozen Inference Graph to OpenCV DNN

Question

I have trained an object detection model using Tensorflow API, following an example based on this Google Colaboratory notebook by Roboflow. https://colab.research.google.com/drive/1wTMIrJhYsQdq_u7ROOkf0Lu_fsX5Mu8a

So far so good and i have successfully extracted my trained model as an Inference graph, again following the same notebook:

import re
import numpy as np

output_directory = './fine_tuned_model'

lst = os.listdir(model_dir)
lst = [l for l in lst if 'model.ckpt-' in l and '.meta' in l]
steps=np.array([int(re.findall('\d+', l)[0]) for l in lst])
last_model = lst[steps.argmax()].replace('.meta', '')

last_model_path = os.path.join(model_dir, last_model)
print(last_model_path)
!python /content/models/research/object_detection/export_inference_graph.py \
    --input_type=image_tensor \
    --pipeline_config_path={pipeline_fname} \
    --output_directory={output_directory} \
    --trained_checkpoint_prefix={last_model_path}

That gives me a frozen_inference_graph.pb file that i can use to make my object detection program in OpenCV DNN. Also following this example https://stackoverflow.com/a/57055266/9914815 i prepared a.pbtxt file of the model and pipeline config as the second argument for the cv2.dnn.readNetFromTensorflow function. Here is the code just enough to reproduce the error i'm having:

model = cv2.dnn.readNetFromTensorflow('models/trained/frozen_inference_graph.pb', 
                                      'models/trained/output.pbtxt')

This code works successfully when i used the pretrained SSD MobileNet V2 COCO model, ssd_mobilenet_v2_coco_2018_03_29.pbtxt

however using my trained.pbtxt file, it will throw this error:

C:\Users\Satria\Desktop\ExploreOpencvDnn-master>python trainedmodel_video.py -i test1.mp4 -o test1result.mp4
Traceback (most recent call last):                                                                                                                            
File "trainedmodel_video.py", line 48, in <module> 'models/trained/output.pbtxt') cv2.error:
OpenCV(4.1.1) C:\projects\opencv-python\opencv\modules\dnn\src\tensorflow\tf_importer.cpp:544:error:
(-2:Unspecified error) Input layer not found: FeatureExtractor/MobilenetV2/Conv/weights in function
'cv::dnn::dnn4_v20190621::`anonymous-namespace'::TFImporter::connect' 

It says that Input Layer is not found. Why does this happen? Also Notice the error message points out to a directory:

C:\projects\opencv-python\opencv\modules\dnn\src\tensorflow\tf_importer.cpp

which is incredibly strange, because i do not have that directory at all in my computer. I tried diffchecking the pbtxt and config files of my and the sample SSD mobilenet model and i cannot find any instance of that particular directory used in anywhere, nor even they have a directory path inside.

Is this caused by training using Google Colab? Is there any correct way i can use Colab-trained Tensorflow models in OpenCV DNN?

Thanks in advance!

Answer 1

Solved after adding an additional input node in my own generated pbtxt file

Someone suggested that OpenCV Version 4.11 which i was using is outdated. I updated to 4.30, still not working, however it now lets me to use FusedBatchNormV3 which is very important in the future.

Now, after taking a close look at the diffcheck in the sample and the generated pbtxt,

In the sample.pbtxt file ssd_mobilenet_v2_coco_2018_03_29.pbtxt , line 30 onward

node {  
  name: "Preprocessor/mul"  
  op: "Mul" 
  input: "image_tensor" 
  input: "Preprocessor/mul/x"   
}   
node {  
  name: "Preprocessor/sub"  
  op: "Sub" 
  input: "Preprocessor/mul" 
  input: "Preprocessor/sub/y"   
}   
node {  
  name: "FeatureExtractor/MobilenetV2/Conv/Conv2D"  
  op: "Conv2D"  
  input: "Preprocessor/sub" 
  input: "FeatureExtractor/MobilenetV2/Conv/weights"    

It has an additional Input nodes which uses Preprocessor , not only FeatureExtractor/MobilenetV2/Conv/Conv2D

meanwhile on the generated pbtxt it only has this

node {
  name: "FeatureExtractor/MobilenetV2/Conv/Conv2D"
  op: "Conv2D"
  input: "FeatureExtractor/MobilenetV2/Conv/weights"

I copied the Input nodes of the sample.pbtxt and into my own generated.pbtxt and it worked!!!