darknet.py get different results with the same test image on command line

[Digital2Slave]

I trained an objects detect model based on resnet50 in Ubuntu 16.04 LTS by refer 432 . But, when i test the model with darknet.py on command line. I got different results with the same *.cfg, *.weights, *.data and test image.

For example: By use python darknet.py with the *.cfg, *.weights, *.data and test image. I got the result label 5. But, by use python darknet.py with the same *.cfg, *.weights, *.data and test image many times, I finally got the result label 1.

This is weird. All the command lines run in normal user. I don't know why, anyone could help me out? Thanks a lot. @AlexeyAB

[AlexeyAB]

@JohnTian Hi,

• Did you change anything in darknet.py?
• How do you test image many times?
• Can you show screenshots of different results?

[Digital2Slave]

@AlexeyAB Thanks for your reply.

• I only change the way to pass .cfg, .weights, *.data and test image in main function

  if __name__ == "__main__":
  
  cfgPath = sys.argv[1]
  weightsPath = sys.argv[2]
  dataPath = sys.argv[3]
  imagePath = sys.argv[4]
  
  res = performDetect(
  imagePath=imagePath,
  thresh=0.25,
  configPath=cfgPath,
  weightPath=weightsPath,
  metaPath=dataPath,
  showImage=False)
  classes = set([r[0] for r in res])
  print(classes)

• Actually, I didn't count the number of times. This number may be between 20 and 30.
• The different reults as following:

➜  epbox_tiny_mobile git:(master) python tf_models/pddvsm/pddvsm.py \  
tf_models/pddvsm/pddvsm_resnet50_yolo_recall.cfg \
tf_models/pddvsm/pddvsm_resnet50_yolo_final.weights \   
tf_models/pddvsm/pddvsm.data \  
/mnt/images/epbox/92902000094/2019-01-18/M9290200009490118020/M9290200009490118020_0_crop.jpg  
5  
➜  epbox_tiny_mobile git:(master) python tf_models/pddvsm/pddvsm.py \  
tf_models/pddvsm/pddvsm_resnet50_yolo_recall.cfg \
tf_models/pddvsm/pddvsm_resnet50_yolo_final.weights \   
tf_models/pddvsm/pddvsm.data \  
/mnt/images/epbox/92902000094/2019-01-18/M9290200009490118020/M9290200009490118020_0_crop.jpg  
1  

Finally, after retrain the model by adding new labelled images. This weird phenomena disappeared! By now, for the same test image, this model always produce '5'.

➜  epbox_tiny_mobile git:(master) python tf_models/pddvsm/pddvsm.py \  
tf_models/pddvsm/pddvsm_resnet50_yolo_recall.cfg \
tf_models/pddvsm/pddvsm_resnet50_yolo_final.weights \   
tf_models/pddvsm/pddvsm.data \  
/mnt/images/epbox/92902000094/2019-01-18/M9290200009490118020/M9290200009490118020_0_crop.jpg   
5  

[OndoyManing]

@JohnTian Hello! Would it be possible for you to share your darknet.py? I just want to test it out because the one included in the zip file does not work for me.

[Digital2Slave]

@OndoyManing darknet.py Be sure, your darknet.so is generated correct.

[benjaminabruzzo]

I had a similar issue, though I posted it on the pjreddie page: https://github.com/pjreddie/darknet/issues/1547