626_privacy/tensorflow_privacy
2
0
Fork 1
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions

Compare commits

merge into: 626_privacy:master
Branches Tags
626_privacy:master
626_privacy:akemi/lira_2021_fix
...
pull from: 626_privacy:akemi/lira_2021_fix
Branches Tags
626_privacy:akemi/lira_2021_fix
626_privacy:master

5 commits
master ... akemi/lira

Author SHA1 Message Date
Akemi Izuko
72fcf6732a
Lira2021: 4x multigpu 2024-11-05 11:53:25 -07:00
Akemi Izuko
1a61c2a36a
Lira2021: fix readme pip install 2024-11-05 11:53:25 -07:00
Akemi Izuko
9a0611fb37
Lira2021: fix readme install instructions 2024-11-05 11:53:25 -07:00
Akemi Izuko
9c67e0abef
Lira2021: increase training script verbosity 2024-11-05 11:53:25 -07:00
Akemi Izuko
f3af0a3a5c
Lira2021: fix training for cuda 12 2024-11-05 11:53:25 -07:00
4 changed files with 62 additions and 31 deletions
Show stats Expand all files Collapse all files

17
research/mi_lira_2021/README.md
View file

@ -8,14 +8,19 @@ by Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and F
### INSTALLING
You will need to install fairly standard dependencies
You will need to install fairly standard dependencies and python 3.11 minimum.
`pip install scipy, sklearn, numpy, matplotlib`
```
pip install scipy scikit-learn numpy matplotlib tensorflow tensorflow_datasets
and also some machine learning framework to train models. We train our models
with JAX + ObJAX so you will need to follow build instructions for that
https://github.com/google/objax
https://objax.readthedocs.io/en/latest/installation_setup.html
# This needs to be separate
pip install objax
RELEASE_URL="https://storage.googleapis.com/jax-releases/jax_cuda_releases.html"
JAX_VERSION=`python3 -c 'import jax; print(jax.__version__)'`
pip uninstall -y jaxlib
pip install -f $RELEASE_URL jax[cuda]==$JAX_VERSION
```
### RUNNING THE CODE

53
research/mi_lira_2021/scripts/train_demo.sh
View file

@ -11,20 +11,41 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
mkdir -p logs
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 0 --logdir exp/cifar10 &> logs/log_0
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 1 --logdir exp/cifar10 &> logs/log_1
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 2 --logdir exp/cifar10 &> logs/log_2
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 3 --logdir exp/cifar10 &> logs/log_3
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 4 --logdir exp/cifar10 &> logs/log_4
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 5 --logdir exp/cifar10 &> logs/log_5
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 6 --logdir exp/cifar10 &> logs/log_6
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 7 --logdir exp/cifar10 &> logs/log_7
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 8 --logdir exp/cifar10 &> logs/log_8
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 9 --logdir exp/cifar10 &> logs/log_9
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 10 --logdir exp/cifar10 &> logs/log_10
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 11 --logdir exp/cifar10 &> logs/log_11
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 12 --logdir exp/cifar10 &> logs/log_12
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 13 --logdir exp/cifar10 &> logs/log_13
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 14 --logdir exp/cifar10 &> logs/log_14
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 15 --logdir exp/cifar10 &> logs/log_15
SECONDS=0
echo '======== 1 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 0 --logdir exp/cifar10 2>&1 | tee logs/log_0
echo '======== 2 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 1 --logdir exp/cifar10 2>&1 | tee logs/log_1
echo '======== 3 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 2 --logdir exp/cifar10 2>&1 | tee logs/log_2
echo '======== 4 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 3 --logdir exp/cifar10 2>&1 | tee logs/log_3
echo '======== 5 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 4 --logdir exp/cifar10 2>&1 | tee logs/log_4
echo '======== 6 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 5 --logdir exp/cifar10 2>&1 | tee logs/log_5
echo '======== 7 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 6 --logdir exp/cifar10 2>&1 | tee logs/log_6
echo '======== 8 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 7 --logdir exp/cifar10 2>&1 | tee logs/log_7
echo '======== 9 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 8 --logdir exp/cifar10 2>&1 | tee logs/log_8
echo '======== 10 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 9 --logdir exp/cifar10 2>&1 | tee logs/log_9
echo '======== 11 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 10 --logdir exp/cifar10 2>&1 | tee logs/log_10
echo '======== 12 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 11 --logdir exp/cifar10 2>&1 | tee logs/log_11
echo '======== 13 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 12 --logdir exp/cifar10 2>&1 | tee logs/log_12
echo '======== 14 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 13 --logdir exp/cifar10 2>&1 | tee logs/log_13
echo '======== 15 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 14 --logdir exp/cifar10 2>&1 | tee logs/log_14
echo '======== 16 of 16 ========'
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 15 --logdir exp/cifar10 2>&1 | tee logs/log_15
echo "COMPLETE: Took ${SECONDS} seconds"

19
research/mi_lira_2021/scripts/train_demo_multigpu.sh
View file

@ -11,22 +11,25 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
mkdir -p logs
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 0 --logdir exp/cifar10 &> logs/log_0 &
CUDA_VISIBLE_DEVICES='1' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 1 --logdir exp/cifar10 &> logs/log_1 &
CUDA_VISIBLE_DEVICES='2' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 2 --logdir exp/cifar10 &> logs/log_2 &
CUDA_VISIBLE_DEVICES='3' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 3 --logdir exp/cifar10 &> logs/log_3 &
CUDA_VISIBLE_DEVICES='4' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 4 --logdir exp/cifar10 &> logs/log_4 &
CUDA_VISIBLE_DEVICES='5' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 5 --logdir exp/cifar10 &> logs/log_5 &
CUDA_VISIBLE_DEVICES='6' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 6 --logdir exp/cifar10 &> logs/log_6 &
CUDA_VISIBLE_DEVICES='7' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 7 --logdir exp/cifar10 &> logs/log_7 &
wait;
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 4 --logdir exp/cifar10 &> logs/log_4 &
CUDA_VISIBLE_DEVICES='1' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 5 --logdir exp/cifar10 &> logs/log_5 &
CUDA_VISIBLE_DEVICES='2' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 6 --logdir exp/cifar10 &> logs/log_6 &
CUDA_VISIBLE_DEVICES='3' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 7 --logdir exp/cifar10 &> logs/log_7 &
wait;
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 8 --logdir exp/cifar10 &> logs/log_8 &
CUDA_VISIBLE_DEVICES='1' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 9 --logdir exp/cifar10 &> logs/log_9 &
CUDA_VISIBLE_DEVICES='2' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 10 --logdir exp/cifar10 &> logs/log_10 &
CUDA_VISIBLE_DEVICES='3' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 11 --logdir exp/cifar10 &> logs/log_11 &
CUDA_VISIBLE_DEVICES='4' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 12 --logdir exp/cifar10 &> logs/log_12 &
CUDA_VISIBLE_DEVICES='5' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 13 --logdir exp/cifar10 &> logs/log_13 &
CUDA_VISIBLE_DEVICES='6' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 14 --logdir exp/cifar10 &> logs/log_14 &
CUDA_VISIBLE_DEVICES='7' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 15 --logdir exp/cifar10 &> logs/log_15 &
wait;
CUDA_VISIBLE_DEVICES='0' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 12 --logdir exp/cifar10 &> logs/log_12 &
CUDA_VISIBLE_DEVICES='1' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 13 --logdir exp/cifar10 &> logs/log_13 &
CUDA_VISIBLE_DEVICES='2' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 14 --logdir exp/cifar10 &> logs/log_14 &
CUDA_VISIBLE_DEVICES='3' python3 -u train.py --dataset=cifar10 --epochs=100 --save_steps=20 --arch wrn28-2 --num_experiments 16 --expid 15 --logdir exp/cifar10 &> logs/log_15 &
wait;

4
research/mi_lira_2021/train.py
View file

@ -66,7 +66,9 @@ class TrainLoop(objax.Module):
for k, v in kv.items():
if jn.isnan(v):
raise ValueError('NaN, try reducing learning rate', k)
if summary is not None:
if summary is not None and v.ndim == 1:
summary.scalar(k, float(v[0]))
elif summary is not None:
summary.scalar(k, float(v))
def train(self, num_train_epochs: int, train_size: int, train: DataSet, test: DataSet, logdir: str, save_steps=100, patience=None):