Questions about the Projector

[Amanda-Zheng]

We found that the code for the projector part difficult to understand, there seems a lot cumsum and index operations. We might not be familiar for parabolic transformation in the 3d space. Could you provide more references for understand?

[Polkach]

There is no specific theory or paper to read about that. We just realize something like inter-area (cv2-style) interpolation for our transformation. To this end, we estimate which pixels were projected to the current pixel during bending. The new color of the current pixel is a weighted average of colors of its preimages

[misszhucanjie]

Hello, I am going to implement your code recently, but found that our packages are not compatible. Can I trouble you to generate the requirements.txt file so that I can download your operating environment directly?

[Polkach]

@misszhucanjie Hello! Try TensorFlow 1.13.1. If it doesn't help then we will try to find the environment that we used. It isn't easy, so try a specific version of TensorFlow first, please :) It should be enough

[misszhucanjie]

Hello, I have successfully run the code in your paper, but I found that in calculating the cosine similarity, the model you use is R100.pb, I can ask you, what model is the model of this parameter? His input is 3112112, and the output is 1*512. Does the output represent probability?

[fatheral]

Output is the vector (of length 512) representation of the input image (3x112x112).

[misszhucanjie]

Output is the vector (of length 512) representation of the input image (3x112x112).

The output of the R100 model is a 512-dimensional vector, but if we want to test other models, the number of model outputs is wrong. I don't know why in calculating the cosine similarity, why input a photo to the model and predict 512 numbers. After multiplying the output vectors of the two photos, the cosine similarity is obtained.

[Polkach]

@misszhucanjie The common procedure of FaceID model inference is as follows.

First, we feed an input image to the model. The model provides a feature vector for the input image.

This feature vector is some internal representation of the face by the model. Its length can be various for different models. Additionally, it is not a semantic vector, i.e. we don't know what is the meaning of its values. The only thing we know is that if an angle between vectors is small then initial photos for these vectors are photos of the same person.

Note that vectors from different models can't be compared straightforwardly.

[cserezwan]

@misszhucanjie Hello! Try TensorFlow 1.13.1. If it doesn't help then we will try to find the environment that we used. It isn't easy, so try a specific version of TensorFlow first, please :) It should be enough

Which Python version?

[Polkach]

@cserezwan python 3.6

[cserezwan]

@cserezwan python 3.6

Thanks, can you tell me which CUDA version and cuDNN version you have used?
I think I'm getting the following error because of CUDA, cuDNN and Tensorflow mismatch:

tensorflow.python.framework.errors_impl.UnknownError: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.
[[{{node pnet/conv1/Conv2D}}]]
[[{{node pnet/prob1}}]]

[gongmm]

How to slove this error: The GPU Memory is enough
InternalError (see above for traceback): Blas xGEMMBatched launch failed : a.shape=[5,2,3], b.shape=[5,3,810000], m=2, n=810000, k=3, batch_size=5
[[node MatMul (defined at /data/project/advhat/Attack/stn.py:150) ]]
[[node AddN (defined at /data/project/advhat/Attack/stn.py:227) ]]

[misszhucanjie]

你是不是没截取完，我复现完没这个错误，找第一个错误地方

…

---Original--- From: ***@***.***> Date: Sun, Jul 11, 2021 16:19 PM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [papermsucode/advhat] Questions about the Projector (#12) How to slove this error: The GPU Memory is enough InternalError (see above for traceback): Blas xGEMMBatched launch failed : a.shape=[5,2,3], b.shape=[5,3,810000], m=2, n=810000, k=3, batch_size=5 [[node MatMul (defined at /data/project/advhat/Attack/stn.py:150) ]] [[node AddN (defined at /data/project/advhat/Attack/stn.py:227) ]] — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[misszhucanjie]

有个地方输入的数据大小不对，需要改下函数

…

---Original--- From: ***@***.***> Date: Sun, Jul 11, 2021 16:19 PM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [papermsucode/advhat] Questions about the Projector (#12) How to slove this error: The GPU Memory is enough InternalError (see above for traceback): Blas xGEMMBatched launch failed : a.shape=[5,2,3], b.shape=[5,3,810000], m=2, n=810000, k=3, batch_size=5 [[node MatMul (defined at /data/project/advhat/Attack/stn.py:150) ]] [[node AddN (defined at /data/project/advhat/Attack/stn.py:227) ]] — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[cserezwan]

@cserezwan python 3.6

Thanks, can you tell me which CUDA version and cuDNN version you have used? I think I'm getting the following error because of CUDA, cuDNN and Tensorflow mismatch:

tensorflow.python.framework.errors_impl.UnknownError: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above. [[{{node pnet/conv1/Conv2D}}]] [[{{node pnet/prob1}}]]

@Polkach
Please reply to the questions!!
Can you please let us know the specific versions of the important functions that you used for this code?
If possible the requirements.txt file?
Otherwise, it's becoming very difficult to reproduce the results using your code.

[Polkach]

Dear @cserezwan,
Unfortunately, we don't have access to the hardware where we did the research more than 2 years ago. Also, at that time we didn't have an idea to save Conda or Docker environment to simplify the reproduction.
Quick googling of the problem shows that downgrading to Tensorflow==1.8 often helps.
The only thing I can recommend is to use PyTorch instead of Tensorflow to increase the compatibility of your code with future versions of packages)