faster_rcnn predictions not consistent

[shortcipher3]

Repeated predictions give different results, I modified the example demo.py by replacing:

    img = utils.read_image(args.image, color=True)
    bboxes, labels, scores = model.predict([img])
    bbox, label, score = bboxes[0], labels[0], scores[0]

with

    img = utils.read_image(args.image, color=True)
    bboxes, labels, scores = model.predict([img])
    bbox, label, score = bboxes[0], labels[0], scores[0]
    print(label)
    print(score)

    img = utils.read_image(args.image, color=True)
    bboxes, labels, scores = model.predict([img])
    bbox, label, score = bboxes[0], labels[0], scores[0]
    print(label)
    print(score)

Running on an image prints two different results. [ 6 9 11 12 14 14 14] [0.9994035 0.7133383 0.9320137 0.9976526 0.9986652 0.98459226 0.7558696 ] [ 6 11 12 14 14] [0.99914455 0.81991583 0.96614665 0.9929738 0.9864108 ]

Any idea as to why that would be?

[yuyu2172]

Hmm. I tried to reproduce your problem, but the model outputs the same consistently with images I have. Is it possible to have your image?

[Hakuyume]

Some operations of cuDNN do not have deterministic behaviour. https://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/#reproducibility Perhaps, this is related to your problem.

[shortcipher3]

Hmm.

I'm using the projects reference image: https://cloud.githubusercontent.com/assets/2062128/26187667/9cb236da-3bd5-11e7-8bcf-7dbd4302e2dc.jpg

@Hakuyume I'm not using cuDNN, I'm using a CPU based model.

I dockerized my setup and also uploaded the modified demo.py. Had to change extensions to upload, should be demo-mod.py,

Dockerfile.txt demo-mod.txt

Are you running on a GPU @yuyu2172 ?

Thanks for your help!

[yuyu2172]

I tested on GPU/CPU on my machine. I ran your script, and I get consistent results. I am building the docker container now.

$ python demo-mod.py  9cb236da-3bd5-11e7-8bcf-7dbd4302e2dc.jpg  
[ 6 11 12 14 14]
[0.99953234 0.9766117  0.997649   0.9980629  0.99296147]
[ 6 11 12 14 14]
[0.99953234 0.9766117  0.997649   0.9980629  0.99296147]
[ 6 11 12 14 14]
[0.99953234 0.9766117  0.997649   0.9980629  0.99296147]

[shortcipher3]

My results are looking more consistent today:

[ 6 11 12 14 14]
[0.9991875  0.96595764 0.99568295 0.99677104 0.9895921 ]
[ 6 11 12 14 14]
[0.9994771 0.9520375 0.9942531 0.9980769 0.9920569]
[ 6 11 12 14 14]
[0.9995468  0.9600438  0.99564916 0.99760795 0.9974911 ]

The scores are still slightly different. Thanks for looking into it!

[knorth55]

Can i close this issue? @yuyu2172