def __init__(self, filename=None, absolute=False, tolerance=0.1):
try:
if filename is not None:
f = open(filename, 'r')
# Get the scan's pose (stored in cartesian form)
l = f.readline()
scan_loc = reg.findall(l)[0]
self.posx = float(scan_loc[0])
self.posy = float(scan_loc[1])
self.rot = float(scan_loc[2])
self.scan_points = []
if absolute:
self.scan_points.append(polar2origincartesian(self, scan_loc[3], scan_loc[4]))
else:
self.scan_points.append(polar2cartesian(scan_loc[3], scan_loc[4]))
line = f.readline()
# Data is in polar form
while line:
coords = map(float, reg.findall(line)[0][5:7])
if absolute:
self.scan_points.append(polar2origincartesian(self, coords[0], coords[1]))
else:
self.scan_points.append(polar2cartesian(coords[0], coords[1]))
line = f.readline()
f.close()
except ValueError as e:
print "Error in file: ", filename
raise e
self.scan_points = subsample(self.scan_points, tolerance)
def subsample_and_serialize(data_root, in_folder, out_folder):
"""Read MatLab files, optionally attenuate and subsample, and write to TFRecords files
NOTE - place test set files in different folder than training mat files
Arguments:
data_root -- root folder for project's data
in_folder -- folder where training set .mat files are located
out_folder -- folder where *.train and *.valid Protobuf files will be written
"""
raw_folder = os.path.join(data_root, in_folder)
file_names = filter(lambda file_name: file_name.endswith(".mat"),
os.listdir(raw_folder))
preprocessed_dir = os.path.join(data_root, out_folder)
if not os.path.exists(preprocessed_dir):
os.mkdir(preprocessed_dir)
for mat_file_name in file_names:
train_file_name = os.path.join(preprocessed_dir,
mat_file_name.replace(".mat", ".train"))
valid_file_name = os.path.join(preprocessed_dir,
mat_file_name.replace(".mat", ".valid"))
if os.path.exists(train_file_name) and os.path.exists(valid_file_name):
print("Skipping existing file:", train_file_name)
print("Skipping existing file:", valid_file_name)
continue
label = get_label(mat_file_name)
try:
data = mat_to_data(os.path.join(raw_folder, mat_file_name))
except ValueError:
print("Skipping broken file:", mat_file_name)
continue
xs = data["data"]
xs = normalize(xs)
if SUBSAMPLE:
xs = subsample(xs, channels=CHANNELS, rate=SUBSAMPLE_RATE)
num_windows = xs.shape[0] // WINDOW_SIZE
xs = np.reshape(xs, (num_windows, WINDOW_SIZE, CHANNELS))
train_writer = tf.python_io.TFRecordWriter(train_file_name)
valid_writer = tf.python_io.TFRecordWriter(valid_file_name)
print("Writing file:", train_file_name)
print("Writing file:", valid_file_name)
for idx, x in enumerate(xs):
example = to_example_proto(x, label)
if idx % 20 == 0:
valid_writer.write(example.SerializeToString())
else:
train_writer.write(example.SerializeToString())
train_writer.close()
valid_writer.close()
def _sample_from_m_matches(m):
indicators = tf.equal(num_matches, tf.cast(m, tf.float32))
### debug
# hist = tf.bincount(tf.cast(num_matches, tf.int32), minlength=n, maxlength=n)
#indicators = tf.Print(indicators, [m, self._k, num_matches, hist],
# summarize=1000)
####
if self._subsample_hard_examples:
return util.topk_or_pad_inds_with_resampling(
indicators, difficulties, num_samples)
else:
return util.subsample(indicators, num_samples)
def __init__(self, filename, numpoints, samplesize=-1, tolerance=0.1):
self.points = []
self.minX = sys.maxint
self.minY = sys.maxint
self.maxX = -sys.maxint - 1
self.maxY = -sys.maxint - 1
try:
if filename is not None:
with open(filename, 'r') as f:
l = f.readline()
while l:
point = map(float, l.rstrip('\n').split(","))
if point[0] < self.minX:
self.minX = point[0]
elif point[0] > self.maxX:
self.maxX = point[0]
if point[1] < self.minY:
self.minY = point[1]
elif point[1] > self.maxY:
self.maxY = point[1]
self.points.append(point)
l = f.readline()
# Data is in polar form
except ValueError as e:
print "Error in file: ", filename
raise e
self.points = subsample(self.points, tolerance)
dimX = self.maxX - self.minX
dimY = self.maxY - self.minY
Xpoints = int(
math.sqrt(((dimX * numpoints) / dimY) + (math.pow(dimX - dimY, 2) /
(4 * (dimY**2)))) -
((dimX - dimY) / (2 * dimY)))
Ypoints = int(numpoints / (Xpoints))
self.Xstep = dimX / (Xpoints - 1)
self.Ystep = dimY / (Ypoints - 1)
self.grid = []
for x in trange(Xpoints):
row = []
for y in range(Ypoints):
Xrange = (x * self.Xstep, (x + 1) * self.Xstep)
Yrange = (y * self.Xstep, (y + 1) * self.Ystep)
hasPoint = False
for point in self.points:
if inRange(point[0], Xrange) & inRange(point[1], Yrange):
hasPoint = True
break
row.append(1 if hasPoint else 0)
self.grid.append(row)
def __init__(self, filename=None, samplesize=-1, tolerance=0.1):
self.points = []
try:
if filename is not None:
with open(filename, 'r') as f:
l = f.readline()
while l:
point = l.rstrip('\n').split(",")
self.points.append(map(float, point))
l = f.readline()
# Data is in polar form
except ValueError as e:
print "Error in file: ", filename
raise e
self.points = subsample(self.points, tolerance)
def generate_test_segment(data_root, test_folder="test"):
"""
Emit preprocessed segment along with filename
Already chopped up and ready to send windows into feed dict
"""
test_path = os.path.join(data_root, test_folder)
file_names = filter(lambda x: x.endswith(".mat"), os.listdir(test_path))
for file_name in file_names:
file_path = os.path.join(test_path, file_name)
data = mat_to_data(file_path)
segment = data["data"]
segment = normalize(segment)
if SUBSAMPLE:
segment = subsample(segment,
channels=CHANNELS,
rate=SUBSAMPLE_RATE)
num_windows = segment.shape[0] // WINDOW_SIZE
segment = np.reshape(segment, (num_windows, WINDOW_SIZE, CHANNELS))
yield segment, file_name
val_result = np.load('dataset/val_result_mtrx.npy')
annotations_val = '/root/MedleyDB_selected/Annotations/Melody_Annotations/MELODY1/val/'
val_set = PitchEstimationDataSet(annotations_val,
'/root/data/val/',
sr_ratio=2,
audio_type='MIX')
val_pitches = []
for i in range(val_result.shape[0]):
pitch_frame = val_result[i]
max = -1
max_pitch = 0
for j in range(val_result.shape[1]):
if (pitch_frame[j][0] > max):
max = pitch_frame[j][0]
max_pitch = pitch_frame[j][1]
val_pitches.append(max_pitch)
val_pitches = np.asarray(val_pitches)
val_labels = []
for pitches in val_set.pitches:
val_labels += pitches
val_labels = np.asarray(val_labels)
sampled_val_pitches = util.subsample(val_pitches, val_labels)
labels = range(109)
cnf_matrix = confusion_matrix(val_labels, sampled_val_pitches, labels=labels)
plot_confusion_matrix(cnf_matrix, title='cnf matrix')
def main(argv):
print("Start Main")
# Set arguments: Save_Dir Structure Learning_Rate Earling_Stoping Batch_Size Data_Dir
epochs = FLAGS.epochs
data_dir = FLAGS.data_dir
save_dir = FLAGS.save_dir
learning_rate = FLAGS.lr
early_stop = FLAGS.early_stop
batch_size = FLAGS.batch_size
reg_coeff = FLAGS.reg_coeff
split = FLAGS.split
master = FLAGS.master
checkpoint_path = FLAGS.checkpoint_path
input_dir = FLAGS.input_dir
output_dir = FLAGS.output_dir
image_width = FLAGS.image_width
image_height = FLAGS.image_height
num_classes = FLAGS.num_classes
eps = FLAGS.eps
batch_shape = [batch_size, image_height, image_width, 3]
num_ens = FLAGS.num_ens
tf.logging.set_verbosity(tf.logging.INFO)
(x_train, y_train), (x_test, y_test) = mnist.load_data()
if K.image_data_format() == 'channels_first':
x_train = x_train.reshape(x_train.shape[0], 1, image_width,
image_height)
x_test = x_test.reshape(x_test.shape[0], 1, image_width, image_height)
input_shape = (1, image_width, image_height)
else:
x_train = x_train.reshape(x_train.shape[0], image_width, image_height,
1)
x_test = x_test.reshape(x_test.shape[0], image_width, image_height, 1)
input_shape = (image_width, image_height, 1)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
#Our model architecture for MNIST dataset
def model_arch():
model = Sequential()
model.add(
Conv2D(32,
kernel_size=(3, 3),
activation='relu',
input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
return model
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
sess = tf.Session()
keras.backend.set_session(sess)
x_noisy = util.add_gaussian_noise(x_train, 0,
64) #Add gaussian noise to all images
preds_ens = np.zeros(
(x_test.shape[0], 10)
) #variable to store the predictions of each model in the ensemble (10)
max_vote_ens = np.zeros(
x_test.shape[0]
) #variable to store Majority vote from all models in ensemble
for i in range(num_ens):
model = model_arch(
) #Build a new model architecture for every model in the ensemble
sub_imgs, sub_labels = util.subsample(
x_noisy, y_train) #subsample from the entire data, bagging
model.fit(sub_imgs,
sub_labels,
batch_size=batch_size,
epochs=epochs,
verbose=1) #train the model
model.save("models/mnist/" + str(i) + ".h5") #save the model
ans = sess.run(tf.argmax(model.predict(x_test),
axis=1)) #get the predictions of the model
preds_ens[:, i] = ans.reshape(
(x_test.shape[0])
) #store the predictions of this particular model(i) in ith column of pred_ens variable
del model #erase the model
#Now the variable pred_ens consists of the predictions of all test_data for each model in ensemble.
#ith column contains predictions of ith model.
#go through every row
ens_acc = np.zeros(num_ens)
for i in range(num_ens):
for j in range(preds_ens.shape[0]):
b = Counter(
preds_ens[j][0:i + 1]
) #get the entire row which consists of predictions for that particular instance from all models.
max_vote_ens[j] = b.most_common(1)[0][
0] #get the maximum vote i.e which number has more frequency.
ens_acc_i = sess.run(
tf.reduce_mean(
tf.cast(tf.equal(max_vote_ens, tf.argmax(y_test, axis=1)),
tf.float32)))
ens_acc[i] = ens_acc_i #accuracy of ensemble
#TODO print the nonperturbed test accuracy to the output file.
#Build a model for normal training on the entire noisy data.
model = model.model_arch()
model.fit(x_noisy,
y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1)
acc = model.evaluate(x_test, y_test, verbose=0)
acc_noisy_normal = acc[1] #accuracy of normal model on noisy train data
del model
#Build a new model for normal training (without ensemble) on entire train data (with out bagging and noise).
model = model_arch()
model.fit(x_train,
y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1)
acc = model.evaluate(x_test, y_test, verbose=0)
model.save("models/original_model.h5")
#accuracy of normal model
acc_normal = acc[1]
#generate fgsm adversarial examples on test_data
adv_fgsm = util.fgsm_attack(x_test, model, sess)
acc_fgsm = model.evaluate(adv_fgsm, y_test, verbose=0)
acc_fgsm = acc_fgsm[
1] #accuracy of normal model on fgsm adversarial examples
#generate bim adversarial examples on test_data
adv_bim = util.bim_attack(x_test, model, sess)
acc_bim = model.evaluate(adv_bim, y_test, verbose=0)
acc_bim = acc_bim[1] #accuracy of normal model on bim adversarial examples
#generate lbfgs adversarial examples on test_data
# The target is chosen as 6
adv_lbfgs = util.lbfgs_attack(x_test, model, sess, 6)
acc_lbfgs = model.evaluate(adv_lbfgs, y_test, verbose=0)
acc_lbfgs = acc_lbfgs[
1] #accuracy of normal model on lbfgs adversarial examples
preds_ens_fgsm = np.zeros(
(x_test.shape[0], 10)
) #variable to store the predictions of each model in the ensemble (10) for fgsm adversarial examples
max_vote_ens_fgsm = np.zeros(
x_test.shape[0]
) #variable to store Majority vote from all models in ensemble for fgsm adversarial examples
preds_ens_bim = np.zeros(
(x_test.shape[0], 10)
) #variable to store the predictions of each model in the ensemble (10) for bim adversarial examples
max_vote_ens_bim = np.zeros(
x_test.shape[0]
) #variable to store Majority vote from all models in ensemble for bim adversarial examples
preds_ens_lbfgs = np.zeros(
(x_test.shape[0], 10)
) #variable to store the predictions of each model in the ensemble (10) for lbfgs adversarial examples
max_vote_ens_lbfgs = np.zeros(
x_test.shape[0]
) #variable to store Majority vote from all models in ensemble for lbfgs adversarial examples
del model
for i in range(num_ens):
model = load_model("models/" + str(i) + ".h5")
#get predictions of model i for fgsm adversarial examples
ans = sess.run(tf.argmax(model.predict(adv_fgsm), axis=1))
preds_ens_fgsm[:, i] = ans.reshape((adv_fgsm.shape[0]))
#get predictions of model i for bim adversarial examples
ans = sess.run(tf.argmax(model.predict(adv_bim), axis=1))
preds_ens_bim[:, i] = ans.reshape((adv_bim.shape[0]))
#get predictions of model i for lbfgs adversarial examples
ans = sess.run(tf.argmax(model.predict(adv_lbfgs), axis=1))
preds_ens_lbfgs[:, i] = ans.reshape((adv_lbfgs.shape[0]))
del model
#Now the variable pred_ens consists of the predictions of all fgsm adversarial test_data for each model in ensemble.
#ith column contains predictions of ith model.
#go through every row
ens_acc_fgsm = np.zeros(num_ens)
for i in range(num_ens):
for j in range(preds_ens_fgsm.shape[0]):
b = Counter(
preds_ens_fgsm[j][0:i + 1]
) #get the entire row which consists of predictions for that particular instance from all models.
max_vote_ens_fgsm[j] = b.most_common(1)[0][
0] #get the maximum vote i.e which number has more frequency.
#accuracy of ensemble
ens_acc_fgsm_i = sess.run(
tf.reduce_mean(
tf.cast(tf.equal(max_vote_ens_fgsm, tf.argmax(y_test, axis=1)),
tf.float32)))
ens_acc_fgsm[i] = ens_acc_fgsm_i
#Now the variable pred_ens consists of the predictions of all bim adversarial test_data for each model in ensemble.
#ith column contains predictions of ith model.
#go through every row
ens_acc_bim = np.zeros(num_ens)
for i in range(num_ens):
for j in range(preds_ens_bim.shape[0]):
b = Counter(preds_ens_bim[j][0:i + 1])
max_vote_ens_bim[j] = b.most_common(1)[0][0]
#accuracy of ensemble on bim_adv
ens_acc_bim_i = sess.run(
tf.reduce_mean(
tf.cast(tf.equal(max_vote_ens_bim, tf.argmax(y_test, axis=1)),
tf.float32)))
ens_acc_bim[i] = ens_acc_bim_i
#Now the variable pred_ens consists of the predictions of all lbfgs adversarial test_data for each model in ensemble.
#ith column contains predictions of ith model.
#go through every row
ens_acc_lbfgs = np.zeros(num_ens)
for i in range(num_ens):
for i in range(preds_ens_lbfgs.shape[0]):
b = Counter(preds_ens_lbfgs[j][0:i + 1])
max_vote_ens_lbfgs[j] = b.most_common(1)[0][0]
#accuracy of ensemble on lbfgs_adv
ens_acc_lbfgs_i = sess.run(
tf.reduce_mean(
tf.cast(
tf.equal(max_vote_ens_lbfgs, tf.argmax(y_test, axis=1)),
tf.float32)))
ens_acc_lbfgs[i] = ens_acc_lbfgs_i
#-----------------------------------Adversarial Training--------------------------------------------------------------
#first adversarial examples are generated using train_data, then the model is trained on train_data+adv_train_data.
#Then the model is tested on normal test_data, then the model is tested on adversarial_test_data.
#So, we are generating the adversarial examples twice both on train and test data.
model = load_model("models/original_model.h5")
wrap = KerasModelWrapper(model)
#generate adversarial examples on train data.
adv_fgsm_train = util.fgsm_attack(x_train, model, sess)
adv_bim_train = util.bim_attack(x_train, model, sess)
adv_lbfgs_train = util.lbfgs_attack(x_train, model, sess, 6)
train_plus_adv_fgsm = np.concatenate([x_train, adv_fgsm_train])
y_train_plus_adv_fgsm = np.concatenate([y_train, y_train])
train_plus_adv_bim = np.concatenate([x_train, adv_bim_train])
y_train_plus_adv_bim = np.concatenate([y_train, y_train])
train_plus_adv_lbfgs = np.concatenate([x_train, adv_lbfgs_train])
y_train_plus_adv_lbfgs = np.concatenate([y_train, y_train])
del model
#FGSM TRAINING
#build a fresh model for fgsm training
model = model_arch()
wrap = KerasModelWrapper(model)
model.fit(train_plus_adv_fgsm,
y_train_plus_adv_fgsm,
batch_size=batch_size,
epochs=epochs,
verbose=1)
model.save("models/mnist_fgsm_model.h5")
fgsm_acc_train = model.evaluate(x_test, y_test, verbose=0)
fgsm_acc_train[
1] #Accuracy of adversarially trained model on clean examples
#generate adversarial examples for adversarially trained model on test_data
adv_fgsm_test = util.fgsm_attack(x_test, model, sess)
fgsm_adv_acc_train = model.evaluate(adv_fgsm_test, y_test, verbose=0)
fgsm_adv_acc_train[
1] #Accuracy of adversarially trained model on adv_test images
del model
#BIM TRAINING
#build a fresh model for bim training
model = model_arch()
wrap = KerasModelWrapper(model)
model.fit(train_plus_adv_bim,
y_train_plus_adv_bim,
batch_size=batch_size,
epochs=epochs,
verbose=1)
bim_acc_train = model.evaluate(x_test, y_test, verbose=0)
bim_acc_train[
1] #Accuracy of adversarially trained model on clean examples
#generate adversarial examples for adversarially trained model on test_data
adv_bim_test = util.bim_attack(x_test, model, sess)
bim_adv_acc_train = model.evaluate(adv_bim_test, y_test, verbose=0)
bim_adv_acc_train[
1] #Accuracy of adversarially trained model on adv_test images
del model
#LBFGS TRAINING
#build a fresh model for lbfgs training
model = model_arch()
wrap = KerasModelWrapper(model)
model.fit(train_plus_adv_lbfgs,
y_train_plus_adv_lbfgs,
batch_size=batch_size,
epochs=epochs,
verbose=1)
lbfgs_acc_train = model.evaluate(x_test, y_test, verbose=0)
#Accuracy of adversarially trained model on clean examples
lbfgs_acc_train[1]
adv_lbfgs_test = util.lbfgs_attack(x_test, model, sess, 6)
lbfgs_adv_acc_train = model.evaluate(adv_lbfgs_test, y_test, verbose=0)
lbfgs_adv_acc_train[
1] #Accuracy of adversarially trained model on adv_test images
del model
def main(argv):
print("Start Main")
# Set arguments: Save_Dir Structure Learning_Rate Earling_Stoping Batch_Size Data_Dir
data_dir = FLAGS.data_dir
save_dir = FLAGS.save_dir
learning_rate = FLAGS.lr
early_stop = FLAGS.early_stop
batch_size = FLAGS.batch_size
reg_coeff = FLAGS.reg_coeff
split = FLAGS.split
master = FLAGS.master
checkpoint_path = FLAGS.checkpoint_path
input_dir = FLAGS.input_dir
output_dir = FLAGS.output_dir
image_width = FLAGS.image_width
image_height = FLAGS.eps
num_classes = FLAGS.num_classes
eps = FLAGS.eps
batch_shape = [batch_size, image_height, image_width, 3]
input_shape = [image_height, image_width, 3]
tf.logging.set_verbosity(tf.logging.INFO)
def model_arch():
model = Sequential()
model.add(
Conv2D(50,
kernel_size=(5, 5),
activation='relu',
input_shape=input_shape))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(100, (5, 5), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(200, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(400, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(200, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
return model
model = model_arch
#load training data
imgs, labels, names = util.load_training_images('tiny-imagenet-200/train/')
print("Training Images Loaded")
#retrype and resize training data
imgs = imgs[0:100]
labels = labels[0:100]
names = names[0:100]
imgs_large = np.ndarray(shape=[imgs.shape[0], 299, 299, 3])
for i in range(imgs.shape[0]):
imgs_large[i, :, :, :] = util.rescale(imgs[i])
imgs_large = imgs_large.astype('uint8')
imgs_noisy = np.ndarray(shape=imgs_large.shape)
for i in range(imgs_large.shape[0]):
imgs_noisy[i, :, :, :] = util.noisy(1, imgs_large[i])
imgs_noisy = imgs_noisy.astype('uint8')
sub_imgs, sub_labels = util.subsample(imgs_noisy, labels)
batch_shape = [20, 299, 299, 3]
num_classes = 200
def main(argv):
print("Start Main")
# Set arguments: Save_Dir Structure Learning_Rate Earling_Stoping Batch_Size Data_Dir
data_dir = FLAGS.data_dir
save_dir = FLAGS.save_dir
learning_rate = FLAGS.lr
early_stop = FLAGS.early_stop
batch_size = FLAGS.batch_size
epochs = FLAGS.epochs
reg_coeff = FLAGS.reg_coeff
split = FLAGS.split
master = FLAGS.master
checkpoint_path = FLAGS.checkpoint_path
input_dir = FLAGS.input_dir
output_dir = FLAGS.output_dir
image_width = FLAGS.image_width
image_height = FLAGS.eps
num_classes = FLAGS.num_classes
eps = FLAGS.eps
batch_shape = [batch_size, image_height, image_width, 3]
input_shape = [image_height, image_width, 3]
num_ens = FLAGS.num_ens
tf.logging.set_verbosity(tf.logging.INFO)
def model_arch():
model = Sequential()
model.add(Conv2D(50, kernel_size=(5, 5), activation='relu', input_shape=input_shape))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(100, (5, 5), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(200, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(400, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(200, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
return model
model = model_arch
#load training data
x_train,y_train,train_names = util.load_training_images('tiny-imagenet-200/train/')
print("Training Images Loaded")
x_test,y_test,test_names = util.load_training_images('tiny-imagenet-200/test/')
print("Testing Images Loaded")
#retrype and resize training data
x_train = x_train[0:100]
y_train = y_train[0:100]
train_names = train_names[0:100]
x_train_large = np.ndarray(shape= [x_train.shape[0],299,299,3])
for i in range(x_train.shape[0]):
x_train_large[i,:,:,:] = util.rescale(x_train[i])
x_train_large=x_train_large.astype('uint8')
x_train_noisy = np.ndarray(shape= x_train_large.shape)
for i in range(x_train_large.shape[0]):
x_train_noisy[i,:,:,:] = util.noisy(1,x_train_large[i])
x_train_noisy=x_train_noisy.astype('uint8')
x_train_sub,y_train_sub = util.subsample(x_train_noisy,y_train)
batch_shape = [20, 299, 299, 3]
num_classes = 200
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
sess = tf.Session()
keras.backend.set_session(sess)
x_noisy = util.add_gaussian_noise(x_train,0,64) #Add gaussian noise to all images
preds_ens = np.zeros((x_test.shape[0],10)) #variable to store the predictions of each model in the ensemble (10)
max_vote_ens = np.zeros(x_test.shape[0]) #variable to store Majority vote from all models in ensemble
for i in range(num_ens):
model = model_arch() #Build a new model architecture for every model in the ensemble
x_train_sub,y_train_sub = util.subsample(x_train_noisy,y_train) #subsample from the entire data, bagging
model.fit(x_train_sub, y_train_sub, batch_size=batch_size,epochs=epochs,verbose=1) #train the model
model.save("models/imgnet/"+str(i)+".h5") #save the model
ans = sess.run(tf.argmax(model.predict(x_test),axis=1)) #get the predictions of the model
preds_ens[:,i]= ans.reshape((x_test.shape[0])) #store the predictions of this particular model(i) in ith column of pred_ens variable
del model #erase the model
#Now the variable pred_ens consists of the predictions of all test_data for each model in ensemble.
#ith column contains predictions of ith model.
#go through every row
print("Ensemble method Clean")
ens_acc = np.zeros(num_ens)
for i in range(num_ens):
for j in range(preds_ens.shape[0]):
b= Counter(preds_ens[j][0:i+1]) #get the entire row which consists of predictions for that particular instance from all models.
max_vote_ens[j] = b.most_common(1)[0][0] #get the maximum vote i.e which number has more frequency.
ens_acc_i = sess.run(tf.reduce_mean(tf.cast(tf.equal(max_vote_ens, tf.argmax(y_test, axis=1)) , tf.float32)))
ens_acc[i] = ens_acc_i #accuracy of ensemble
#TODO print the nonperturbed test accuracy to the output file.
print("Accuracy : " + str(np.mean(ens_acc)))
#Build a model for normal training on the entire noisy data.
model = model.model_arch()
model.fit(x_train_noisy, y_train, batch_size=batch_size, epochs=epochs, verbose=1)
acc = model.evaluate(x_test, y_test, verbose=0)
acc_noisy_normal = acc[1] #accuracy of normal model on noisy train data
del model
#Build a new model for normal training (without ensemble) on entire train data (with out bagging and noise).
model = model_arch()
model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, verbose=1)
acc = model.evaluate(x_test, y_test, verbose=0)
model.save("models/imgnet/original_model.h5")
#accuracy of normal model
acc_normal = acc[1]
print("accuracy of normal model : " + str(acc_normal))
print("accuracy of normal model on noisy train data : " + str(acc_noisy_normal))
#generate fgsm adversarial examples on test_data
adv_fgsm = util.fgsm_attack(x_test,model,sess)
acc_fgsm = model.evaluate(adv_fgsm, y_test, verbose=0)
acc_fgsm = acc_fgsm[1]
print("accuracy of normal model on fgsm adversarial examples : " + str(acc_fgsm))
#generate bim adversarial examples on test_data
adv_bim = util.bim_attack(x_test,model,sess)
acc_bim = model.evaluate(adv_bim,y_test,verbose=0)
acc_bim = acc_bim[1] #accuracy of normal model on bim adversarial examples
print("accuracy of normal model on bim adversarial examples : " + str(acc_bim))
#generate lbfgs adversarial examples on test_data
# The target is chosen as 6
adv_lbfgs = util.lbfgs_attack(x_test,model,sess,6)
acc_lbfgs = model.evaluate(adv_lbfgs,y_test,verbose=0)
acc_lbfgs = acc_lbfgs[1] #accuracy of normal model on lbfgs adversarial examples
print("accuracy of normal model on lbfgs adversarial examples : " + str(acc_lbfgs))
preds_ens_fgsm = np.zeros((x_test.shape[0],10)) #variable to store the predictions of each model in the ensemble (10) for fgsm adversarial examples
max_vote_ens_fgsm = np.zeros(x_test.shape[0]) #variable to store Majority vote from all models in ensemble for fgsm adversarial examples
preds_ens_bim = np.zeros((x_test.shape[0],10)) #variable to store the predictions of each model in the ensemble (10) for bim adversarial examples
max_vote_ens_bim = np.zeros(x_test.shape[0]) #variable to store Majority vote from all models in ensemble for bim adversarial examples
preds_ens_lbfgs = np.zeros((x_test.shape[0],10)) #variable to store the predictions of each model in the ensemble (10) for lbfgs adversarial examples
max_vote_ens_lbfgs = np.zeros(x_test.shape[0]) #variable to store Majority vote from all models in ensemble for lbfgs adversarial examples
del model
for i in range(num_ens):
model = load_model("models/"+str(i)+".h5")
#get predictions of model i for fgsm adversarial examples
ans = sess.run(tf.argmax(model.predict(adv_fgsm),axis=1))
preds_ens_fgsm[:,i]= ans.reshape((adv_fgsm.shape[0]))
#get predictions of model i for bim adversarial examples
ans = sess.run(tf.argmax(model.predict(adv_bim),axis=1))
preds_ens_bim[:,i]= ans.reshape((adv_bim.shape[0]))
#get predictions of model i for lbfgs adversarial examples
ans = sess.run(tf.argmax(model.predict(adv_lbfgs),axis=1))
preds_ens_lbfgs[:,i]= ans.reshape((adv_lbfgs.shape[0]))
del model
print("Now the variable pred_ens consists of the predictions of all fgsm adversarial test_data for each model in ensemble.")
#ith column contains predictions of ith model.
#go through every row
ens_acc_fgsm = np.zeros(num_ens)
for i in range(num_ens):
for j in range(preds_ens_fgsm.shape[0]):
b= Counter(preds_ens_fgsm[j][0:i+1]) #get the entire row which consists of predictions for that particular instance from all models.
max_vote_ens_fgsm[j] = b.most_common(1)[0][0] #get the maximum vote i.e which number has more frequency.
#accuracy of ensemble
ens_acc_fgsm_i = sess.run(tf.reduce_mean(tf.cast(tf.equal(max_vote_ens_fgsm, tf.argmax(y_test, axis=1)) , tf.float32)))
ens_acc_fgsm[i] = ens_acc_fgsm_i
print(str(np.mean(ens_acc_fgsm)))
print("Now the variable pred_ens consists of the predictions of all bim adversarial test_data for each model in ensemble.")
#ith column contains predictions of ith model.
#go through every row
ens_acc_bim = np.zeros(num_ens)
for i in range(num_ens):
for j in range(preds_ens_bim.shape[0]):
b= Counter(preds_ens_bim[j][0:i+1])
max_vote_ens_bim[j] = b.most_common(1)[0][0]
#accuracy of ensemble on bim_adv
ens_acc_bim_i = sess.run(tf.reduce_mean(tf.cast(tf.equal(max_vote_ens_bim, tf.argmax(y_test, axis=1)) , tf.float32)))
ens_acc_bim[i] = ens_acc_bim_i
print(str(np.mean(ens_acc_bim)))
print("Now the variable pred_ens consists of the predictions of all lbfgs adversarial test_data for each model in ensemble.")
#ith column contains predictions of ith model.
#go through every row
ens_acc_lbfgs = np.zeros(num_ens)
for i in range(num_ens):
for i in range(preds_ens_lbfgs.shape[0]):
b= Counter(preds_ens_lbfgs[j][0:i+1])
max_vote_ens_lbfgs[j] = b.most_common(1)[0][0]
#accuracy of ensemble on lbfgs_adv
ens_acc_lbfgs_i = sess.run(tf.reduce_mean(tf.cast(tf.equal(max_vote_ens_lbfgs, tf.argmax(y_test, axis=1)) , tf.float32)))
ens_acc_lbfgs[i] = ens_acc_lbfgs_i
print(str(np.mean(ens_acc_lbfgs)))
labels = util.make_justpot(labels)
if args.pot_part:
partitions = []
no_pot_imgs = util.filter_no_pot(labels, image_paths)
for color, subjs in util.pot_map.items():
subj_imgs = []
subj_imgs.extend([
x for x in image_paths
if os.path.basename(x).startswith(tuple(subjs))
])
imgs = list(set(no_pot_imgs + subj_imgs))
if args.make_uniform:
imgs = util.make_uniform(imgs, labels)
train, test = util.subsample(imgs, percent=args.samp_percent)
partitions.append((color, train, test))
for p in partitions:
labels = util.read_labels(os.path.join(annot_input_dir, "labels"))
gen_and_write(output_dir,
exp_num,
p[1],
labels,
mode=f"{p[0]}_train",
args=args)
gen_and_write(output_dir,
exp_num,
p[2],
def main(argv):
print("Start Main")
# Set arguments: Save_Dir Structure Learning_Rate Earling_Stoping Batch_Size Data_Dir
data_dir = FLAGS.data_dir
save_dir = FLAGS.save_dir
learning_rate = FLAGS.lr
early_stop = FLAGS.early_stop
batch_size = FLAGS.batch_size
epochs = FLAGS.epochs
reg_coeff = FLAGS.reg_coeff
split = FLAGS.split
master = FLAGS.master
checkpoint_path = FLAGS.checkpoint_path
input_dir = FLAGS.input_dir
output_dir = FLAGS.output_dir
image_width = FLAGS.image_width
image_height = FLAGS.eps
num_classes = FLAGS.num_classes
eps = FLAGS.eps
batch_shape = [batch_size, image_height, image_width, 3]
input_shape = [image_height, image_width, 3]
num_ens = FLAGS.num_ens
tf.logging.set_verbosity(tf.logging.INFO)
def model_arch():
model = Sequential()
model.add(
Conv2D(50,
kernel_size=(5, 5),
activation='relu',
input_shape=input_shape))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(100, (5, 5), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(200, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(400, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(200, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
return model
model = model_arch
#load training data
x_train, y_train, train_names = util.load_training_images(
'tiny-imagenet-200/train/')
print("Training Images Loaded")
x_test, y_test, test_names = util.load_training_images(
'tiny-imagenet-200/test/')
print("Testing Images Loaded")
#retrype and resize training data
x_train = x_train[0:100]
y_train = y_train[0:100]
train_names = train_names[0:100]
x_train_large = np.ndarray(shape=[x_train.shape[0], 299, 299, 3])
for i in range(x_train.shape[0]):
x_train_large[i, :, :, :] = util.rescale(x_train[i])
x_train_large = x_train_large.astype('uint8')
x_train_noisy = np.ndarray(shape=x_train_large.shape)
for i in range(x_train_large.shape[0]):
x_train_noisy[i, :, :, :] = util.noisy(1, x_train_large[i])
x_train_noisy = x_train_noisy.astype('uint8')
x_train_sub, y_train_sub = util.subsample(x_train_noisy, y_train)
batch_shape = [20, 299, 299, 3]
num_classes = 200
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
sess = tf.Session()
keras.backend.set_session(sess)
#-----------------------------------Adversarial Training--------------------------------------------------------------
#first adversarial examples are generated using train_data, then the model is trained on train_data+adv_train_data.
#Then the model is tested on normal test_data, then the model is tested on adversarial_test_data.
#So, we are generating the adversarial examples twice both on train and test data.
model = load_model("models/imgnet/original_model.h5")
wrap = KerasModelWrapper(model)
#generate adversarial examples on train data.
adv_fgsm_train = util.fgsm_attack(x_train, model, sess)
adv_bim_train = util.bim_attack(x_train, model, sess)
adv_lbfgs_train = util.lbfgs_attack(x_train, model, sess, 6)
train_plus_adv_fgsm = np.concatenate([x_train, adv_fgsm_train])
y_train_plus_adv_fgsm = np.concatenate([y_train, y_train])
train_plus_adv_bim = np.concatenate([x_train, adv_bim_train])
y_train_plus_adv_bim = np.concatenate([y_train, y_train])
train_plus_adv_lbfgs = np.concatenate([x_train, adv_lbfgs_train])
y_train_plus_adv_lbfgs = np.concatenate([y_train, y_train])
del model
print("FGSM TRAINING")
#build a fresh model for fgsm training
model = model_arch()
wrap = KerasModelWrapper(model)
model.fit(train_plus_adv_fgsm,
y_train_plus_adv_fgsm,
batch_size=batch_size,
epochs=epochs,
verbose=1)
model.save("models/imgnet/fgsm_model.h5")
fgsm_acc_train = model.evaluate(x_test, y_test, verbose=0)
fgsm_acc_train[
1] #Accuracy of adversarially trained model on clean examples
#generate adversarial examples for adversarially trained model on test_data
adv_fgsm_test = util.fgsm_attack(x_test, model, sess)
fgsm_adv_acc_train = model.evaluate(adv_fgsm_test, y_test, verbose=0)
fgsm_adv_acc_train[
1] #Accuracy of adversarially trained model on adv_test images
del model
print("BIM TRAINING") #BIM TRAINING
#build a fresh model for bim training
model = model_arch()
wrap = KerasModelWrapper(model)
model.fit(train_plus_adv_bim,
y_train_plus_adv_bim,
batch_size=batch_size,
epochs=epochs,
verbose=1)
bim_acc_train = model.evaluate(x_test, y_test, verbose=0)
print("Accuracy of adversarially trained model on clean examples\n" +
str(bim_acc_train[1]))
#generate adversarial examples for adversarially trained model on test_data
adv_bim_test = util.bim_attack(x_test, model, sess)
bim_adv_acc_train = model.evaluate(adv_bim_test, y_test, verbose=0)
print("Accuracy of adversarially trained model on adv_test images\n" +
str(bim_adv_acc_train[1]))
del model
print("LBFGS TRAINING")
#build a fresh model for lbfgs training
model = model_arch()
wrap = KerasModelWrapper(model)
model.fit(train_plus_adv_lbfgs,
y_train_plus_adv_lbfgs,
batch_size=batch_size,
epochs=epochs,
verbose=1)
print("Accuracy of adversarially trained model on clean examples")
lbfgs_acc_train = model.evaluate(x_test, y_test, verbose=0)
print(str(lbfgs_acc_train[1]))
print("Accuracy of adversarially trained model on lbfgs examples")
lbfgs_acc_train[1]
adv_lbfgs_test = util.lbfgs_attack(x_test, model, sess, 6)
lbfgs_adv_acc_train = model.evaluate(adv_lbfgs_test, y_test, verbose=0)
print(str(lbfgs_adv_acc_train[1])
) #Accuracy of adversarially trained model on adv_test images
del model
