model.compile(
optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
# Change to True if you want to train from scratch
train = False
if (train):
# Save the untrained weights for future training with modified dataset
model.save_weights('h5/cnnc-untrained.h5')
model.fit(train_images,
train_labels,
batch_size=100,
epochs=5,
validation_data=(test_images, test_labels))
model.save_weights('h5/cnnc-trained.h5')
else:
model.load_weights('h5/cnnc-trained.h5')
test_loss, test_acc = model.evaluate(test_images, test_labels, verbose=2)
print("Accuracy before faults:", test_acc)
tfi.inject(model=model, confFile="confFiles/sample.yaml")
test_loss, test_acc = model.evaluate(test_images, test_labels, verbose=2)
print("Accuracy after faults:", test_acc)
totsdc = 0.
start = time.time()
for i in range(numFaults):
train_labels1 = copy.deepcopy(train_labels);train_images1=copy.deepcopy(train_images)
model.load_weights('h5/lenet-trained.h5')
ind = []
init = random.sample(range(num), numInjections+offset)
for i in init:
test_loss, test_acc = model.evaluate(test_images[i:i+1], test_labels[i:i+1], verbose=0)
if(test_acc == 1.):
ind.append(i)
ind = ind[:numInjections]
model.load_weights('h5/lenet-untrained.h5')
train_images_,train_labels_ = tfi.inject(x_test=train_images1,y_test=train_labels1, confFile=conf)
model.fit(train_images_, train_labels_, batch_size=100, epochs=15,
validation_data=(test_images, test_labels), verbose=0)
sdc = 0.
for i in ind:
test_loss, test_acc = model.evaluate(test_images[i:i+1], test_labels[i:i+1], verbose=0)
if(test_acc == 0.):
sdc = sdc + 1.
f.write(str(sdc/numInjections))
f.write("\n")
totsdc = totsdc + sdc
f.write("\n")
f.write(str(totsdc/(numFaults*numInjections)))
f.write("\n")
start = time.time()
conf = sys.argv[1]
filePath = sys.argv[2]
filePath = os.path.join(filePath, "res.csv")
f = open(filePath, "w")
numFaults = int(sys.argv[3])
for k in range(numFaults):
for i in range(43):
count = 0.
tesXi = tesX[i]
for j in range(30):
res = tfi.inject(model=model,
x_test=X_test[tesXi[j:j + 1]],
confFile=conf)
if (res == i):
count = count + 1.
countX[i] = countX[i] + count
for i in range(43):
countX[i] = countX[i] / numFaults
f.write(str(countX))
f.write("\n")
f.write("Time for %d injections: %f seconds" %
(numFaults, time.time() - start))
f.close()
'''
conf = sys.argv[1]
filePath = sys.argv[2]
filePath = os.path.join(filePath, "res.csv")
f = open(filePath, "w")
numFaults = int(sys.argv[3])
numInjections = 10
#numInjections = int(sys.argv[4])
totsdc = 0.0
start = time.time()
for i in range(numFaults):
model.load_weights('h5/resnet-trained.h5')
tfi.inject(model=model, confFile=conf)
sdc = 0.
for i in ind:
image = load_img(images[i], target_size=(224, 224))
image = img_to_array(image)
image = image.reshape(
(1, image.shape[0], image.shape[1], image.shape[2]))
image = preprocess_input(image)
out = model.predict(image)
label = decode_predictions(out)
label = label[0][0]
if (label[1] != img_labels[i]):
sdc = sdc + 1.
f.write(str(sdc / numInjections))
f.write("\n")
totsdc = totsdc + sdc
numInjections = 10
#numInjections = int(sys.argv[4])
totsdc = 0.0
start = time.time()
for j in range(numFaults):
model.load_weights('h5/sqznet-trained.h5')
#tfi.inject(model=model, confFile=conf)
sdc = 0.
for i in ind:
image = load_img(images[i], target_size=(227, 227))
image = img_to_array(image)
image = image.reshape(
(1, image.shape[0], image.shape[1], image.shape[2]))
image = preprocess_input(image)
res = tfi.inject(model=model, x_test=image, confFile=conf)
label_ = decode_predictions(res[0])
label_ = label_[0][0]
if (label_[1] != img_labels[i]):
sdc = sdc + 1.
f.write(str(sdc / numInjections))
f.write("\n")
totsdc = totsdc + sdc
f.write("\n")
f.write(str(totsdc / (numFaults * numInjections)))
f.write("\n")
f.write("Time for %d injections: %f seconds" %
(numFaults * numInjections, time.time() - start))
f.close()