Which version of Theano in this experiment is runnable?

[dljgs1]

I cannot run code using newest Theano, it seems incompatible with lasagne in python2

[csong27]

Hi,

Are you using Theano 1.0.0? Could you post the log file of error?

[dljgs1]

Thanks for reply. This is error log:

Traceback (most recent call last): File "distributed_sgd.py", line 4, in import lasagne File "/usr/lib64/python2.7/site-packages/lasagne/init.py", line 19, in from . import layers File "/usr/lib64/python2.7/site-packages/lasagne/layers/init.py", line 7, in from .pool import * File "/usr/lib64/python2.7/site-packages/lasagne/layers/pool.py", line 6, in from theano.tensor.signal import downsample ImportError: cannot import name downsample

pip install theano==1.0.0 Requirement already satisfied: theano==1.0.0 in /usr/lib64/python2.7/site-packages (1.0.0)

It seems that Theano does not support downsample while lasagne still use it in python2.

[csong27]

Could you try upgrade lasagne? I think their latest version doesn't import downsample. See https://github.com/Lasagne/Lasagne/blob/master/lasagne/layers/pool.py

[dljgs1]

Solved! Thanks a lot.

[dljgs1]

Hi Song, I‘ve read the paper <Exploiting Unintended Feature Leakage in Collaborative Learning> and I found that you have a GAN attack experiment on gender classification model. But when I try to do this work, I cannot reappear the result(I use the same network and datasets but cannot recover train data). Could you tell me where to find the runable code or any trick in this experiment ? Tanks a lot.

[csong27]

Hi,

Do you mean Figure 1 in the paper? It is not actually our attack but a re-implementation from previous attack and we show we cannot recover training data as well. See this paper for detail: https://arxiv.org/abs/1702.07464