def training():
config = Configuration()
net = Network()
optimizer = optim.Adam(net.parameters(), lr=0.01)
train_loader = get_train_loader()
test_loader = get_test_loader()
for epoch in range(config.nn_training_epochs):
train(net, optimizer, train_loader, epoch)
test(net, test_loader)
net.save_model()
def main():
config = Configuration()
image_id = 1362
bnn = BNNWrapper()
bnn.load_model()
loss_fn = pyro.infer.Trace_ELBO(
num_particles=config.bnn_adversary_samples
).differentiable_loss
nn = Network()
nn.load_model()
test_loader = get_test_loader(1, shuffle=False)
x, y = test_loader.dataset[image_id]
y = torch.tensor([y])
bnn_d, bnn_imgs, bnn_pertubation_imgs = bnn_adversary.run_attack(
bnn, loss_fn, x, y, config.epsilons, image_id
)
nn_d, nn_imgs, nn_pertubation_imgs = nn_adversary.run_attack(
nn, x, y, config.epsilons, 3
)
ids = [2, 5]
utils.img_show(
x,
f"Image, BNN Prediction: {bnn_d.iloc[0]['y_']}, NN Prediction: {nn_d.iloc[0]['y_']}",
)
for id in ids:
utils.img_two_show(
bnn_pertubation_imgs[id].cpu(),
f"BNN Noise (epsilon: {bnn_d.iloc[id]['epsilon']})",
nn_pertubation_imgs[id].cpu(),
f"NN Noise (epsilon: {bnn_d.iloc[id]['epsilon']})",
)
utils.img_two_show(
bnn_imgs[id].cpu(),
f"Noise added to image, BNN Prediction: {bnn_d.iloc[id]['y_']}",
nn_imgs[id].cpu(),
f"Noise added to image, NN Prediction: {nn_d.iloc[id]['y_']}",
)
def main():
config = Configuration()
image_id = 1362
nn = Network()
nn.load_model()
test_loader = get_test_loader(1, shuffle=False)
x, y = test_loader.dataset[image_id]
y = torch.tensor([y])
nn_d, nn_imgs, nn_pertubation_imgs = nn_adversary.run_attack(
nn, x, y, config.epsilons, 3)
id = 2
utils.img_show(x, f"Image, NN Prediction: {nn_d.iloc[0]['y_']}")
utils.img_show(nn_pertubation_imgs[id].cpu(),
f"Noise (epsilon: {nn_d.iloc[id]['epsilon']})")
utils.img_show(nn_imgs[id].cpu(),
f"Noise added to image, Prediction: {nn_d.iloc[id]['y_']}")
tmp_dict["epsilon"].append(epsilon)
tmp_dict["y"].append(y.item())
tmp_dict["y_"].append(y_)
return pandas.DataFrame.from_dict(tmp_dict), pertubed_images, pertubation
if __name__ == "__main__":
config = Configuration()
net = Network()
net.load_model()
net.eval()
test_loader = get_test_loader(batch_size=1, shuffle=False)
result = []
for batch_id, (x, y) in enumerate(test_loader):
df, _ = run_attack(net, x, y, config.epsilons, batch_id)
result.append(df)
if batch_id % 100 == 0:
print(f"Step {batch_id}/{len(test_loader.dataset)}")
result_df = pandas.concat(result) # type: pandas.DataFrame
result_df.reset_index(inplace=True, drop=True)
result_path = Path("data/")
result_path.mkdir(exist_ok=True, parents=False)
result_df.to_csv(result_path.joinpath(f"{config.id:02}_nn_result.csv"))
def plot_images(config: Configuration):
bnn = BNNWrapper()
bnn.load_model()
loss_fn = pyro.infer.Trace_ELBO(
num_particles=config.bnn_adversary_samples
).differentiable_loss
nn = Network()
nn.load_model()
# sample images from the MNIST test data set
imgs = [3, 4, 1362, 6930]
fig, axes = plt.subplots(12, 2 * len(imgs), figsize=(36, 26))
for j, img in enumerate(imgs):
test_loader = get_test_loader(1, shuffle=False)
x, y = test_loader.dataset[img]
y = torch.tensor([y])
bnn_d, bnn_imgs, _ = bnn_adversary.run_attack(
bnn, loss_fn, x, y, config.epsilons, img
)
# font sizes
title = 22
body = 20
index = 2 * j
axes[0][index].set_title("Bayesian Neural Network", fontsize=title)
for i in range(12):
axes[i][index].imshow(
bnn_imgs[i][0].cpu().detach(), cmap="gray", vmin=0, vmax=1
)
axes[i][index].set_xlabel(
f"Label: {bnn_d['y'][i]}, Prediction: {bnn_d['y_'][i]}", fontsize=body
)
axes[i][index].set_ylabel(f"Eps: {bnn_d['epsilon'][i]}", fontsize=body)
axes[i][index].set_yticklabels([])
axes[i][index].set_xticklabels([])
nn_d, nn_imgs, _ = nn_adversary.run_attack(nn, x, y, config.epsilons, 3)
index = 2 * j + 1
axes[0][index].set_title("Deep Neural Network", fontsize=title)
for i in range(12):
axes[i][index].imshow(
nn_imgs[i][0].cpu().detach(), cmap="gray", vmin=0, vmax=1
)
axes[i][index].set_xlabel(
f"Label: {nn_d['y'][i]}, Prediction: {nn_d['y_'][i]}", fontsize=body
)
axes[i][index].set_ylabel(f"Eps: {bnn_d['epsilon'][i]}", fontsize=body)
axes[i][index].set_yticklabels([])
axes[i][index].set_xticklabels([])
fig.tight_layout()
fig.subplots_adjust(hspace=0.5)
plot_path = Path(f"data/{config.id:02}_example_imgs.svg")
plt.savefig(plot_path)
plt.show()
plt.close()