Experimental code for Structural Coding Paper.
Applying structural coding on a typical CNN with near zero effort:
# linearcode/autonomy.py
from torchvision.models import resnet50
from linearcode.protection import apply_sc_automatically
model = resnet50(pretrained=True)
print(model)
model = apply_sc_automatically(model, n=256, k=32)
print(model)Running the resnet50 network without protection with no fault:
$ docker run --env CONSTRAINTS="{'dataset': 'imagenet', 'model': 'resnet50', 'sampler': 'tiny', 'flips': 0, 'protection': 'none'}" --env PRINT_STAT=1 dsn2022paper165/sc python map.py
Downloading: "https://download.pytorch.org/models/resnet50-19c8e357.pth" to /root/.cache/torch/hub/checkpoints/resnet50-19c8e357.pth
100.0%
Done with batch 0 after injection
{'injection': 0, 'model': 'resnet50', 'quantization': False, 'sampler': 'tiny', 'dataset': 'imagenet', 'flips': 0, 'protection': 'none'}
accuracy 0.75 correct 12 all 16
loss 0.060732901096343994
Running the resnet50 network without protection with row fault model:
$ docker run --env CONSTRAINTS="{'dataset': 'imagenet', 'model': 'resnet50', 'sampler': 'tiny', 'flips': 'row', 'protection': 'none'}" --env PRINT_STAT=1 dsn2022paper165/sc python map.py
Downloading: "https://download.pytorch.org/models/resnet50-19c8e357.pth" to /root/.cache/torch/hub/checkpoints/resnet50-19c8e357.pth
100.0%
Done with batch 0 after injection
{'injection': 0, 'model': 'resnet50', 'quantization': False, 'sampler': 'tiny', 'dataset': 'imagenet', 'flips': 'row', 'protection': 'none'}
Injecting 1193 faults at granularity 16
accuracy 0.0 correct 0 all 16
loss 4708924.5
Running the resnet50 network with sc (Structural Coding) protection with row fault model:
$ docker run --env CONSTRAINTS="{'dataset': 'imagenet', 'model': 'resnet50', 'sampler': 'tiny', 'flips': 'row', 'protection': 'sc'}" --env PRINT_STAT=1 dsn2022paper165/sc python map.py
Downloading: "https://download.pytorch.org/models/resnet50-19c8e357.pth" to /root/.cache/torch/hub/checkpoints/resnet50-19c8e357.pth
100.0%
Done with batch 0 after injection
{'injection': 0, 'model': 'resnet50', 'quantization': False, 'sampler': 'tiny', 'dataset': 'imagenet', 'flips': 'row', 'protection': 'sc'}
Injecting 1193 faults at granularity 16
accuracy 0.75 correct 12 all 16
loss 0.061373598873615265
Running the resnet50 network with milr (MILR) protection with row fault model:
$ docker run --env CONSTRAINTS="{'dataset': 'imagenet', 'model': 'resnet50', 'sampler': 'tiny', 'flips': 'row', 'protection': 'milr'}" --env PRINT_STAT=1 dsn2022paper165/sc python map.py
Downloading: "https://download.pytorch.org/models/resnet50-19c8e357.pth" to /root/.cache/torch/hub/checkpoints/resnet50-19c8e357.pth
100.0%[W NNPACK.cpp:80] Could not initialize NNPACK! Reason: Unsupported hardware.
Done with batch 0 after injection
{'injection': 0, 'model': 'resnet50', 'quantization': False, 'sampler': 'tiny', 'dataset': 'imagenet', 'flips': 'row', 'protection': 'milr'}
Injecting 1193 faults at granularity 16
accuracy 0.75 correct 12 all 16
loss 0.06021653115749359
[Assuming ubuntu linux]
Create a virtual environment and activate it:
python3.8 -m venv venv
source venv/bin/activate
pip install --upgrade pip
Within the root directory of the code, install the project requirements:
pip install -r requirements.txt
Set the PYTHONPATH environment variable:
export PYTHONPATH=`pwd`
Navigate to the experiment subdirectory:
cd linearcode
You can run the example experiment commands:
$ export CONSTRAINTS="{'dataset': 'imagenet', 'model': 'resnet50', 'sampler': 'tiny', 'flips': 'row', 'protection': 'milr'}"; export PRINT_STAT=1; python map.py
{'injection': 0, 'model': 'resnet50', 'quantization': False, 'sampler': 'tiny', 'dataset': 'imagenet', 'flips': 'row', 'protection': 'milr'}
[W NNPACK.cpp:80] Could not initialize NNPACK! Reason: Unsupported hardware.
Injecting 1193 faults at granularity 16
Done with batch 0 after injection
accuracy 0.75 correct 12 all 16
loss 0.06021653115749359