def train_and_eval_models(train_x, test_x, train_y, test_y):
train_x, dev_x, train_y, dev_y = train_test_split(train_x,
train_y,
test_size=0.1,
random_state=42,
stratify=train_y)
train_x, scaler = data_utils.normalize(train_x)
dev_x, _ = data_utils.normalize(dev_x, scaler=scaler)
test_x, _ = data_utils.normalize(test_x, scaler=scaler)
dump(scaler, 'scaler_{}.joblib'.format(MODALITY))
cprint('Normalized data (zero-score standardization)', 'yellow')
print('Dataset shapes: ', train_x.shape, dev_x.shape, test_x.shape)
print('Class distributions for train/dev/test:')
print(np.unique(train_y, return_counts=True))
print(np.unique(dev_y, return_counts=True))
print(np.unique(test_y, return_counts=True))
# code for SVM training
# svm = SVC(C=C, kernel=KERNEL, degree=DEGREE, gamma='auto')
# svm.fit(train_x, train_y)
# dump(svm, 'svm_{}_{}-kernel_C{}.joblib'.format(MODALITY, KERNEL, C))
# code for MLP training
mlp = MLPClassifier(hidden_layer_sizes=HIDDEN_LAYERS,
alpha=ALPHA,
early_stopping=True,
max_iter=200)
mlp.fit(train_x, train_y)
print(f'mlp.n_iter_: {mlp.n_iter_}')
dump(mlp, 'mlp_{}.joblib'.format(MODALITY))
# Decision Tree
# dt = tree.DecisionTreeClassifier(max_depth=None)
# dt.fit(train_x, train_y)
# print('DT information\nDepth: {}\nNumber_leaves: {}'.format(
# dt.get_depth(),
# dt.get_n_leaves()
# ))
# dump(dt, 'dt_{}.joblib'.format(MODALITY))
# print('DT feature importances:\n{}'.format(dt.feature_importances_))
# print(dt.feature_importances_.shape, np.min(dt.feature_importances_),
# np.max(dt.feature_importances_), sum(dt.feature_importances_))
# predictions
# cprint('\nSVM PERFORMANCE', 'green')
# cprint(f'SVM parameters:\n {svm}\n', 'yellow')
# svm_uars = score(svm, train_x, train_y, dev_x, dev_y, test_x, test_y)
cprint('\nMLP PERFORMANCE', 'green')
cprint(f'MLP parameters:\n {mlp}\n', 'yellow')
mlp_uars = score(mlp, train_x, train_y, dev_x, dev_y, test_x, test_y)
# cprint('\nDT PERFORMANCE', 'green')
# cprint(f'DT parameters:\n {dt}\n', 'yellow')
# dt_uars = score(dt, train_x, train_y, dev_x, dev_y, test_x, test_y)
# return svm_uars, mlp_uars, dt_uars
return mlp_uars
def combined_fill_between(axes, base_lines, x, ymin, ymax, *args, **kwargs):
axes[0].fill_between(x, ymin, ymax, *args, **kwargs)
base_lines = handle_base_line_exceptions(base_lines, axes)
for ax, base_line in zip(axes[1:], base_lines):
if base_line is not None:
comb = data_utils.normalize(base_line, x, ymin, yerr=ymax)
ax.fill_between(comb[0], comb[1], comb[2], *args, **kwargs)
def combined_plot(axes, base_lines, x, y, *args, **kwargs):
axes[0].plot(x, y, *args, **kwargs)
base_lines = handle_base_line_exceptions(base_lines, axes)
for ax, base_line in zip(axes[1:], base_lines):
if base_line is not None:
comb = data_utils.normalize(base_line, x, y)
ax.plot(comb[0], comb[1], *args, **kwargs)
def learn(filepath, save_figures=False):
X, Y = data_utils.get_full_set(filepath)
X = data_utils.normalize(X)
Y = data_utils.aggregate_wine_labels(Y)
score_fn = accuracy_score
learning.pet_learning_rate(X, Y, score_fn, save=save_figures)
plt.show()
def adv_attack_template_S(img_tensor, GAN, target_sz=(127, 127)):
'''adversarial attack to template'''
'''input: pytorch tensor(0,255) ---> output: pytorch tensor(0,255)'''
'''step1: Normalization'''
img_tensor = normalize(img_tensor)
'''step2: pass to G'''
with torch.no_grad():
img_adv = GAN.transform(img_tensor, target_sz)
return img_adv
def combined_errorbar(axes, base_lines, x, y, yerr=None, **kwargs):
axes[0].errorbar(x, y, yerr=yerr, **kwargs)
base_lines = handle_base_line_exceptions(base_lines, axes)
for ax, base_line in zip(axes[1:], base_lines):
comb = data_utils.normalize(base_line, x, y, yerr=yerr)
if len(comb[0]) != len(comb[1]):
raise Exception(
"normalize gave incoherent data: len(x) != len(y): " +
str(len(comb[0])) + " != " + str(len(comb[1])))
ax.errorbar(comb[0], comb[1], yerr=comb[2], **kwargs)
def adv_attack_template(img_tensor, GAN):
'''adversarial attack to template'''
'''input: pytorch tensor(0,255) ---> output: pytorch tensor(0,255)'''
'''step1: Normalization'''
img_tensor = normalize(img_tensor)
'''step2: pass to G'''
with torch.no_grad():
GAN.template_clean1 = img_tensor
GAN.forward()
img_adv = GAN.template_adv255
return img_adv
def forward(self, x):
block1 = self.block1(normalize(x))
block2 = self.block2(block1)
block3 = self.block3(block2)
block4 = self.block4(block3)
block5 = self.block5(block4)
block6 = self.block6(block5)
block7 = self.block7(block6)
block8 = self.block8(block1 + block7)
return (torch.tanh(block8) + 1) / 2
def main(
model_dir: str,
vc_src: str,
vc_tgt: str,
adv_tgt: str,
output: str,
eps: float,
n_iters: int,
attack_type: str,
):
assert attack_type == "emb" or vc_src is not None
model, config, attr, device = load_model(model_dir)
vc_tgt = file2mel(vc_tgt, **config["preprocess"])
adv_tgt = file2mel(adv_tgt, **config["preprocess"])
vc_tgt = normalize(vc_tgt, attr)
adv_tgt = normalize(adv_tgt, attr)
vc_tgt = torch.from_numpy(vc_tgt).T.unsqueeze(0).to(device)
adv_tgt = torch.from_numpy(adv_tgt).T.unsqueeze(0).to(device)
if attack_type != "emb":
vc_src = file2mel(vc_src, **config["preprocess"])
vc_src = normalize(vc_src, attr)
vc_src = torch.from_numpy(vc_src).T.unsqueeze(0).to(device)
if attack_type == "e2e":
adv_inp = e2e_attack(model, vc_src, vc_tgt, adv_tgt, eps, n_iters)
elif attack_type == "emb":
adv_inp = emb_attack(model, vc_tgt, adv_tgt, eps, n_iters)
elif attack_type == "fb":
adv_inp = fb_attack(model, vc_src, vc_tgt, adv_tgt, eps, n_iters)
else:
raise NotImplementedError()
adv_inp = adv_inp.squeeze(0).T
adv_inp = denormalize(adv_inp.data.cpu().numpy(), attr)
adv_inp = mel2wav(adv_inp, **config["preprocess"])
sf.write(output, adv_inp, config["preprocess"]["sample_rate"])
def set_input(self, input):
"""Unpack input data from the dataloader and perform necessary pre-processing steps.
Parameters:
input (dict): include the data itself and its metadata information.
The option 'direction' can be used to swap images in domain A and domain B.
"""
self.template_clean255 = input[0].squeeze(0).cuda() # pytorch tensor, shape=(1,3,127,127) [0,255]
self.template_clean1 = normalize(self.template_clean255)
# print('clean image shape:',self.init_frame_clean.size())
self.X_crops = input[1].squeeze(0).cuda() # pytorch tensor, shape=(N,3,255,255)
def adv_attack_search_new(img_tensor, GAN, search_sz=(255, 255)):
'''adversarial attack to search region'''
'''input: pytorch tensor(0,255) ---> output: pytorch tensor(0,255)'''
'''step1: Normalization'''
img_tensor = normalize(img_tensor)
'''step2: pass to G'''
with torch.no_grad():
GAN.tensor_clean1 = img_tensor
GAN.num_search = img_tensor.size(0)
GAN.forward(search_sz)
img_adv = GAN.tensor_adv255
return img_adv
def test_one(self):
test_array = np.array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10])
normalized = normalize(test_array)
assert normalized[0] == 0.0 and normalized[10] == 1.0
def parse_method(args, dataset, n, c, d, device):
if args.method == 'link':
model = LINK(n, c).to(device)
elif args.method == 'gcn':
if args.dataset == 'ogbn-proteins':
# Pre-compute GCN normalization.
dataset.graph['edge_index'] = normalize(
dataset.graph['edge_index'])
model = GCN(in_channels=d,
hidden_channels=args.hidden_channels,
out_channels=c,
dropout=args.dropout,
save_mem=True,
use_bn=not args.no_bn).to(device)
else:
model = GCN(in_channels=d,
hidden_channels=args.hidden_channels,
out_channels=c,
num_layers=args.num_layers,
dropout=args.dropout,
use_bn=not args.no_bn).to(device)
elif args.method == 'mlp' or args.method == 'cs':
model = MLP(in_channels=d,
hidden_channels=args.hidden_channels,
out_channels=c,
num_layers=args.num_layers,
dropout=args.dropout).to(device)
elif args.method == 'sgc':
if args.cached:
model = SGC(in_channels=d, out_channels=c,
hops=args.hops).to(device)
else:
model = SGCMem(in_channels=d, out_channels=c,
hops=args.hops).to(device)
elif args.method == 'gprgnn':
model = GPRGNN(d, args.hidden_channels, c,
alpha=args.gpr_alpha).to(device)
elif args.method == 'appnp':
model = APPNP_Net(d, args.hidden_channels, c,
alpha=args.gpr_alpha).to(device)
elif args.method == 'gat':
model = GAT(d,
args.hidden_channels,
c,
num_layers=args.num_layers,
dropout=args.dropout,
heads=args.gat_heads).to(device)
elif args.method == 'lp':
mult_bin = args.dataset == 'ogbn-proteins'
model = MultiLP(c, args.lp_alpha, args.hops, mult_bin=mult_bin)
elif args.method == 'mixhop':
model = MixHop(d,
args.hidden_channels,
c,
num_layers=args.num_layers,
dropout=args.dropout,
hops=args.hops).to(device)
elif args.method == 'gcnjk':
model = GCNJK(d,
args.hidden_channels,
c,
num_layers=args.num_layers,
dropout=args.dropout,
jk_type=args.jk_type).to(device)
elif args.method == 'gatjk':
model = GATJK(d,
args.hidden_channels,
c,
num_layers=args.num_layers,
dropout=args.dropout,
heads=args.gat_heads,
jk_type=args.jk_type).to(device)
elif args.method == 'h2gcn':
model = H2GCN(d,
args.hidden_channels,
c,
dataset.graph['edge_index'],
dataset.graph['num_nodes'],
num_layers=args.num_layers,
dropout=args.dropout,
num_mlp_layers=args.num_mlp_layers).to(device)
else:
raise ValueError('Invalid method')
return model
def forward(self, x):
xn = normalize(x)
x1 = self.block1(xn)
x2 = self.block2(xn)
x3 = x1 + x2
return denormalize(x3)
def forward(self, x):
x1 = self.block1(normalize(x))
re = self.residual(x1)
x2 = self.block2(re)
x3 = self.block3(x1 + x2)
return denormalize(x3)
def test_three(self):
binned = normalize(bin_spectra(self.data, 10))
assert binned[280] == 0.3
shuffle=True)
import sys
# get a record
next_text, next_label = next(iter(amz_train))
try:
print "Next record shape: {}".format(next_text.shape)
except AttributeError as e:
print "(No shape) Text: '{}'".format(next_text)
# batch training, testing sets
amz_train_batch = batch_data(amz_train,
normalizer_fun=lambda x: data_utils.normalize(x,
max_length=300,
truncate_left=True,
encoding=None),
transformer_fun=None)
amz_test_batch = batch_data(amz_test,
normalizer_fun=None,transformer_fun=None)
# Spit out some sample data
next_batch = amz_train_batch.next()
data, label = next_batch
np.set_printoptions(threshold=np.nan)
print "Batch properties:"
print "Shape (data): {}".format(data.shape)
print "Shape (label): {}".format(label.shape)
print "Type: {}".format(type(data))
print
print "First record of first batch:"
def test_two(self):
binned = normalize(bin_spectra(self.data, 10))
assert binned[1] == 1.0 and binned[10] == 0.5
def test_one(self):
binned = normalize(bin_spectra(self.data, 1))
assert binned[1] == 0.5 and binned[500] == 1.0
def test_two(self):
test_array = np.array([0, 0, 0])
normalized = normalize(test_array)
assert np.all(normalized == test_array)
def tune(filepath, save_figures=False):
### Getting and processing the dataset
test_X, test_Y, train_X, train_Y = data_utils.get_separate_sets(filepath)
# Normalize continuous features, and aggregate categories of pets
train_X['AgeuponOutcome'] = data_utils.normalize(train_X['AgeuponOutcome'])
train_Y = aggregate_pet_labels(train_Y)
test_X['AgeuponOutcome'] = data_utils.normalize(test_X['AgeuponOutcome'])
test_Y = aggregate_pet_labels(test_Y)
### Running the tuning
score_fn = accuracy_score
score_fn_name = 'accuracy'
print("Running KNN")
if save_figures:
save_path = '../graphs/pet/tuning_knn.png'
else:
save_path = None
best_param, tuning_scores, model = hyperparameter_tuning(
'knn', score_fn_name, train_X, train_Y, save_path=save_path)
test_and_print(model, score_fn, train_X, train_Y, test_X, test_Y,
'K-Nearest Neighbour', best_param, tuning_scores)
print("Running SVM (linear)")
if save_figures:
save_path = '../graphs/pet/tuning_svm_linear.png'
else:
save_path = None
best_param, tuning_scores, model = hyperparameter_tuning(
'svm_linear', score_fn_name, train_X, train_Y, save_path=save_path)
test_and_print(model, score_fn, train_X, train_Y, test_X, test_Y,
'Support Vector Machines (linear)', best_param,
tuning_scores)
print("Running SVM (poly)")
if save_figures:
save_path = '../graphs/pet/tuning_svm_poly.png'
else:
save_path = None
best_param, tuning_scores, model = hyperparameter_tuning(
'svm_poly', score_fn_name, train_X, train_Y, save_path=save_path)
test_and_print(model, score_fn, train_X, train_Y, test_X, test_Y,
'Support Vector Machines (poly)', best_param, tuning_scores)
print("Running decision trees")
if save_figures:
save_path = '../graphs/pet/tuning_dt.png'
else:
save_path = None
best_param, tuning_scores, model = hyperparameter_tuning(
'dec_tree', score_fn_name, train_X, train_Y, save_path=save_path)
test_and_print(model, score_fn, train_X, train_Y, test_X, test_Y,
'Decision Trees', best_param, tuning_scores)
print("Running boosting")
if save_figures:
save_path = '../graphs/pet/tuning_boosting.png'
else:
save_path = None
best_param, tuning_scores, model = hyperparameter_tuning(
'boosting', score_fn_name, train_X, train_Y, save_path=save_path)
test_and_print(model, score_fn, train_X, train_Y, test_X, test_Y,
'Adaboost', best_param, tuning_scores)
print("Running neural networks")
if save_figures:
save_path = '../graphs/pet/tuning_neuraln.png'
else:
save_path = None
best_param, tuning_scores, model = hyperparameter_tuning(
'neural_n', score_fn_name, train_X, train_Y, save_path=save_path)
test_and_print(model, score_fn, train_X, train_Y, test_X, test_Y,
'Neural Nets', best_param, tuning_scores)
plt.show()
args = parser.parse_args()
# get training and testing sets, and their sizes for amazon.
# this HDF5 file uses an 80/20 train/test split and lives at /data/pcallier/amazon
(amtr, amte), (amntr, amnte) = datasets, sizes = batch_data.split_data(
None,
h5_path=args.h5_path,
overwrite_previous=False,
in_memory=False,
shuffle=True)
import sys
# batch training, testing sets
am_train_batch = batch_data.batch_data(amtr,
normalizer_fun=lambda x: data_utils.normalize(x[0],
max_length=300,
truncate_left=True),
transformer_fun=None)
am_test_batch = batch_data.batch_data(amte,
normalizer_fun=None,transformer_fun=None)
# Spit out some sample data
next_batch = am_train_batch.next()
data, label = next_batch
np.set_printoptions(threshold=np.nan)
print "Batch properties:"
print "Length: {}".format(len(data))
print "Type: {}".format(type(data))
print
print "First record of first batch:"
print "Type (1 level in): {}".format(type(data[0]))