DP_Training about doppelganger HOT 15 CLOSED

Could you please show the output of pip list?

from doppelganger.

I run the below command and got the error:
%run binary_class_head.py
AttributeError: module 'tensorflow_estimator.python.estimator.api._v1.estimator' has no attribute 'BinaryClassHead'

from doppelganger.

Package Version Location

absl-py 0.12.0
argon2-cffi 20.1.0
astor 0.8.1
async-generator 1.10
attrs 21.2.0
backcall 0.2.0
bleach 3.3.0
cached-property 1.5.2
certifi 2021.5.30
cffi 1.14.5
colorama 0.4.4
coverage 5.5
cycler 0.10.0
Cython 0.29.23
decorator 5.0.9
defusedxml 0.7.1
dill 0.3.3
dm-tree 0.1.6
entrypoints 0.3
gast 0.2.2
google-pasta 0.1.6
GPUTaskScheduler 0.1.0 c:\downloads\gputaskscheduler-master
grpcio 1.36.1
h5py 3.2.1
importlib-metadata 3.10.0
ipykernel 5.3.4
ipython 7.22.0
ipython-genutils 0.2.0
jedi 0.17.0
Jinja2 3.0.0
jsonschema 3.2.0
jupyter-client 6.1.12
jupyter-core 4.7.1
jupyterlab-pygments 0.1.2
Keras-Applications 1.0.8
Keras-Preprocessing 1.1.2
kiwisolver 1.3.1
Markdown 3.3.4
MarkupSafe 2.0.1
matplotlib 3.4.2
mistune 0.8.4
mkl-fft 1.3.0
mkl-random 1.1.1
mkl-service 2.3.0
mpmath 1.2.1
multiprocess 0.70.11.1
nbclient 0.5.3
nbconvert 6.0.7
nbformat 5.1.3
nest-asyncio 1.5.1
notebook 6.4.0
numpy 1.19.2
packaging 20.9
pandas 1.2.4
pandocfilters 1.4.3
parso 0.8.2
pathos 0.2.7
patsy 0.5.1
pickleshare 0.7.5
Pillow 8.2.0
pip 21.1.2
pox 0.2.9
ppft 1.6.6.3
prometheus-client 0.11.0
prompt-toolkit 3.0.17
protobuf 3.14.0
pycparser 2.20
Pygments 2.9.0
pyparsing 2.4.7
pyreadline 2.1
pyrsistent 0.17.3
python-dateutil 2.8.1
pytz 2021.1
pywin32 227
pywinpty 0.5.7
pyzmq 20.0.0
scipy 1.6.2
Send2Trash 1.5.0
setuptools 52.0.0.post20210125
six 1.15.0
statsmodels 0.12.2
tensorboard 1.14.0
tensorflow 1.14.0
tensorflow-estimator 2.5.0
tensorflow-privacy 0.6.0 c:\downloads\doppelganger-master\privacy
termcolor 1.1.0
terminado 0.9.4
testpath 0.4.4
torch 1.3.1
tornado 6.1
tqdm 4.61.0
traitlets 5.0.5
typing-extensions 3.7.4.3
wcwidth 0.2.5
webencodings 0.5.1
Werkzeug 1.0.1
wheel 0.36.2
wincertstore 0.2
wrapt 1.12.1
zipp 3.4.1

from doppelganger.

I think binary_class_head.py you ran is code inside TensorFlow privacy. So I guess it might be a compatibility problem between TF/TF-privacy/TF-estimator?

Is there any error when you do from tensorflow_privacy.privacy.optimizers import dp_optimizer and from tensorflow_privacy.privacy.analysis.compute_dp_sgd_privacy_lib import compute_dp_sgd_privacy directly in Python command line?

from doppelganger.

The same error:
AttributeError: module 'tensorflow_estimator.python.estimator.api._v1.estimator' has no attribute 'BinaryClassHead'
But if I run 2nd time (both of them), the error is gone.

from doppelganger.

Hey
Now, its running. I changed the code in doppleganger.py. At last the epoch starts. I will let you know if it finished the training successfully.

from doppelganger.

In that case, it should be some issues with TensorFlow Privacy, instead of DoppelGANger, since these two lines are just simply loading TF Privacy. Comparing the libraries you installed to the requirements of TF Privacy, I don't see any problems. And I didn't experience this error before. Maybe it will be more helpful to ask about this error to TF Privacy's authors on TF Privacy's issue page?

from doppelganger.

I am trying to reproduce the Figure-16 from the paper. Could you please let me know the hyperparameters you used for the figure.
Thanks in advance.

from doppelganger.

Hey
Now, its running. I changed the code in doppleganger.py. At last the epoch starts. I will let you know if it finished the training successfully.

Oh I see. What did you change, by loading the libraries twice? Though I am still curious about what caused this error.

from doppelganger.

Yes.

from doppelganger.

Hey
Now, its running. I changed the code in doppleganger.py. At last the epoch starts. I will let you know if it finished the training successfully.

Oh I see. What did you change, by loading the libraries twice? Though I am still curious about what caused this error.

I will open an issue on TF Privacy's page to know more about the error.

from doppelganger.

I am trying to reproduce the Figure-16 from the paper. Could you please let me know the hyperparameters you used for the figure.
Thanks in advance.

By exactly the default configuration in https://github.com/fjxmlzn/DoppelGANger/blob/master/example_dp_training/config.py (so you don't need to change anything). The data for drawing the figure is from the last iteration.

from doppelganger.

Hey,
Thanks for your continuous support. dp_training is working fine. In the paper Section 6.2: Differential Privacy you mentioned that the result showed in the figure 16 is based on epoch 19th. To reproduce the same figure-16 I should run for 20th epoch, right?

This figure is drawn from 14th epoch and epsilon 0.55.
Thanks again for your help.

from doppelganger.

The code is for reproducing the latest arxiv version: https://arxiv.org/pdf/1909.13403.pdf. We lost the DP code that we used for IMC version, so we have re-implemented it and updated the results on arxiv. (Their results look similar and the takeaways are totally the same, but we just wanted to make sure it is reproducible so we updated it.)

from doppelganger.

Thanks for your help.

from doppelganger.