def modelnet_transforms(args):
import kaolin.transforms as tfs
if args.msda_mode in [
'mixup', 'cutmix', 'alt_mixup_fmix', 'alt_mixup_cutmix',
'alt_fmix_cutmix'
]:
raise RuntimeError(
'Mixup and CutMix callbacks not designed for 3D classification.')
fmix_transform = tfs.Compose([
tfs.TriangleMeshToVoxelGrid(args.pointcloud_resolution,
normalize=True),
])
test_transform = tfs.Compose([
tfs.TriangleMeshToPointCloud(num_samples=1000),
tfs.NormalizePointCloud()
])
if args.msda_mode == 'fmix':
transform = fmix_transform
else:
transform = test_transform
return transform, test_transform
help='Number of train epochs.')
parser.add_argument('-lr',
'--learning-rate',
type=float,
default=1e-3,
help='Learning rate.')
parser.add_argument('--batch-size', type=int, default=12, help='Batch size.')
parser.add_argument('--device',
type=str,
default='cuda',
help='Device to use.')
args = parser.parse_args()
transform = tfs.Compose([
tfs.TriangleMeshToPointCloud(num_samples=args.num_points),
tfs.NormalizePointCloud()
])
train_loader = DataLoader(ModelNet(args.modelnet_root,
categories=args.categories,
split='train',
transform=transform,
device=args.device),
batch_size=args.batch_size,
shuffle=True)
val_loader = DataLoader(ModelNet(args.modelnet_root,
categories=args.categories,
split='test',
transform=transform,
def run_feature_selector_algo(args, S, X_train, X_test, T_train, T_test, i,
model_fpsr, model_fnsr, model_msfe, model_mspe,
model_card, model_nme_train, model_nme_test):
log_params = False
file_path_prefix = "./parameters/"
feature_percentage = args.feature_percentage
start_time = time.time()
if args.algo == "RF":
file_path = file_path_prefix + args.data + "/" + args.algo + "-" + str(
i) + ".joblib"
model = RandomForestRegressor(n_estimators=100)
model = create_model(args, file_path, model, X_train, T_train)
importance_vals = model.feature_importances_
# Choose features which has 1% importance according to paper: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6660200/
S_hat = np.argwhere(importance_vals > 0.01).flatten()
if args.data == "MNIST" or args.data == "CIFAR-10": # Take 40% features of MNIST only
file_path = file_path_prefix + args.data + "/" + args.algo + "-" + feature_percentage + "_percent_features-" + str(
i) + ".joblib"
n_sub_feat_size = int(X_train.shape[1] * int(feature_percentage) /
100)
# For RF use this because of the already trained saved model in Sandipan's laptop
# n_sub_feat_size = 315
S_hat = np.argsort(
importance_vals)[::-1][:n_sub_feat_size].flatten(
) #40% features
model = RandomForestRegressor(n_estimators=100)
model = create_model(args, file_path, model, X_train[:, S_hat],
T_train)
X_train = X_train[:, S_hat]
X_test = X_test[:, S_hat]
log_params = True
elif args.algo == "DEEPLIFT":
# Implemented using DeepExplain in SHAP: https://github.com/slundberg/shap
#-------------------------------------------------------------------------#
x_train = X_train
x_test = X_test
X_train = X_train.reshape(X_train.shape[0], 28, 28)
X_test = X_test.reshape(X_test.shape[0], 28, 28)
# Make sure images have shape (28, 28, 1)
X_train = np.expand_dims(X_train, -1)
X_test = np.expand_dims(X_test, -1)
print("X_train shape:", X_train.shape)
print(X_train.shape[0], "train samples")
print(X_test.shape[0], "test samples")
# Model / data parameters
num_classes = 10
input_shape = (28, 28, 1)
"""
## Build the model
"""
model = CNNModel(num_classes, input_shape).create_cnn_model()
model.summary()
file_path = file_path_prefix + args.data + "/" + args.algo + "-" + str(
i) + ".h5"
"""
## Train the model
"""
batch_size = 128
epochs = 15
model.compile(loss="categorical_crossentropy",
optimizer="adam",
metrics=["accuracy"])
model = create_model(args, file_path, model, X_train, T_train)
# Sanity checks
score_train = model.evaluate(X_train, T_train, verbose=0)
score_test = model.evaluate(X_test, T_test, verbose=0)
print("Test loss:", score_test[0])
print("Test accuracy:", score_test[1])
background = X_train[np.random.choice(X_train.shape[0],
100,
replace=False)]
# explain predictions of the model on 10 images
e = shap.DeepExplainer(model, background)
x_test_sample = X_test[np.random.choice(
X_test.shape[0], int(args.deeplift_sample_size), replace=False), :]
shap_values = e.shap_values(x_test_sample)
total_val = np.sum(np.sum(np.abs(shap_values), axis=0),
axis=0).flatten()
S_hat = total_val.argsort()[::-1]
if args.data == "MNIST" or args.data == "CIFAR-10": # Take 40% features of MNIST only
X_train = x_train[:, S_hat]
X_test = x_test[:, S_hat]
X_train = X_train.reshape(X_train.shape[0], 28, 28)
X_test = X_test.reshape(X_test.shape[0], 28, 28)
# Make sure images have shape (28, 28, 1)
X_train = np.expand_dims(X_train, -1)
X_test = np.expand_dims(X_test, -1)
file_path = file_path_prefix + args.data + "/" + args.algo + "-" + feature_percentage + "percent_features-" + str(
i) + ".h5"
n_sub_feat_size = int(X_train.shape[1] * int(feature_percentage) /
100)
S_hat = total_val.argsort()[::-1][:n_sub_feat_size] #40% features
model_new = CNNModel(num_classes, input_shape).create_cnn_model()
model_new.compile(loss="categorical_crossentropy",
optimizer="adam",
metrics=["accuracy"])
model = create_model(args, file_path, model_new, X_train, T_train)
# Just to compare what global features SHAP with DeepLift choose
# X_train_ori = loadmat("./mat_files/MNIST.mat")["train_x"].astype(np.float32)
# show_image([X_train_ori[:,1],X_train_ori[:,20],X_train_ori[:,30]],S_hat[0:len(S)], (args.algo+str(i)))
# show_image(x_train[1,:].flatten(),x_train[20,:].flatten(),x_train[30,:].flatten(),S_hat, (args.algo+str(i)))
log_params = True
elif args.algo == "BART":
# Implemented using XBART: https://github.com/JingyuHe/XBART
#----------------------------------------------------------#
x_train = X_train
x_test = X_test
X_train = pd.DataFrame(X_train)
X_test = pd.DataFrame(X_test)
# Ugly hack otherwise xbart fit does not work
T_train = T_train.flatten()
T_test = T_test.flatten()
file_path = file_path_prefix + args.data + "/" + args.algo + str(
args.tree_size) + "-" + str(i) + ".joblib"
# model = XBART(num_trees = int(args.tree_size), num_sweeps = 20, burnin = 15, verbose = True, parallel = True)
model = XBART(num_trees=int(args.tree_size),
num_sweeps=20,
burnin=15,
verbose=True,
parallel=True)
model = create_model(args, file_path, model, X_train, T_train)
S_hat = sorted(model.importance, key=model.importance.get)[::-1]
imp_vals = np.array(S_hat)
S_hat = imp_vals[imp_vals > 0.01]
if args.data == "MNIST" or args.data == "CIFAR-10": # Take 40% features of MNIST only
file_path = file_path_prefix + args.data + "/" + args.algo + "-" + feature_percentage + "_percent_features-" + str(
i) + ".joblib"
n_sub_feat_size = int(X_train.shape[1] * int(feature_percentage) /
100)
S_hat = sorted(
model.importance,
key=model.importance.get)[::-1][:
n_sub_feat_size] #40% features
model = XBART(num_trees=int(args.tree_size),
num_sweeps=20,
burnin=15,
verbose=True,
parallel=True)
X_train = pd.DataFrame(x_train[:, S_hat])
X_test = pd.DataFrame(x_test[:, S_hat])
model = create_model(args, file_path, model, X_train, T_train)
# Ugly hack otherwise xbart predict does not work
T_train = T_train.reshape(X_train.shape[0], 1)
T_test = T_test.reshape(X_test.shape[0], 1)
log_params = True
elif args.algo == "POINTNET":
import torch
from torch.utils.data import DataLoader
import kaolin as kal
from kaolin import ClassificationEngine
from kaolin.datasets import ModelNet
from kaolin.models.PointNet import PointNetClassifier as PointNet
import kaolin.transforms as tfs
modelnet_path = './mat_files/ModelNet10'
categories = ['chair', 'sofa']
num_points = 1024
device = 'cuda'
transform = tfs.Compose([
tfs.TriangleMeshToPointCloud(num_samples=num_points),
tfs.NormalizePointCloud()
])
train_loader = DataLoader(ModelNet(modelnet_path,
categories=categories,
split='train',
transform=transform,
device=device),
batch_size=12,
shuffle=True)
elif args.algo == "GAM": # Note GAM doesn't work on MNIST properly
file_path = file_path_prefix + args.data + "/" + args.algo + "-" + str(
i) + ".joblib"
thershold = 0.01
gam_fn_form = s(0, n_splines=5)
for feature in range(1, X_train.shape[1]):
gam_fn_form += s(feature, n_splines=5)
# Regression in GAM
# https://pygam.readthedocs.io/en/latest/notebooks/tour_of_pygam.html#Regression
model = GAM(gam_fn_form,
distribution='normal',
link='identity',
max_iter=10,
tol=0.001)
model = create_model(args, file_path, model, X_train, T_train)
feature_vals = np.array(model.statistics_['p_values'])
imp_vals = feature_vals[feature_vals > thershold]
S_hat = np.argsort(imp_vals).flatten()
#S_hat = np.argsort(model.statistics_['p_values'])
log_params = True
elif args.algo == "LASSO":
file_path = file_path_prefix + args.data + "/" + args.algo + "-" + str(
i) + ".joblib"
thershold = 0.01
#T_train = np.argmax(T_train, axis=1)
#T_test = np.argmax(T_test, axis=1)
model = linear_model.Lasso(alpha=0.01, max_iter=5000)
model = create_model(args, file_path, model, X_train, T_train)
imp_vals = model.coef_[model.coef_ > thershold]
S_hat = np.argsort(imp_vals).flatten()
if args.data == "MNIST" or args.data == "CIFAR-10": # Take 40% features of MNIST only
file_path = file_path_prefix + args.data + "/" + args.algo + "-" + feature_percentage + "_percent_features-" + str(
i) + ".joblib"
n_sub_feat_size = int(X_train.shape[1] * int(feature_percentage) /
100)
S_hat = np.argsort(
model.coef_)[::-1][:n_sub_feat_size].flatten() #40% features
model = linear_model.Lasso(alpha=0.01, max_iter=5000)
model = create_model(args, file_path, model, X_train[:, S_hat],
T_train)
X_train = X_train[:, S_hat]
X_test = X_test[:, S_hat]
# Ugly hack otherwise vector norm not calculated
#T_train = T_train.reshape(X_train.shape[0], 1)
#T_test = T_test.reshape(X_test.shape[0], 1)
log_params = True
elif args.algo == "E-NET":
file_path = file_path_prefix + args.data + "/" + args.algo + "-" + str(
i) + ".joblib"
T_train = np.argmax(T_train, axis=1)
T_test = np.argmax(T_test, axis=1)
model = ElasticNet(alpha=0.01, l1_ratio=0.7)
model = create_model(args, file_path, model, X_train, T_train)
S_hat = np.argsort(model.coef_)
log_params = False
elif args.algo == "CORR":
thershold = 0.01
importance_vals = abs(np.dot((X_train.T), T_train).T)[::-1]
S_hat = np.argsort(importance_vals > thershold).flatten()
model_fpsr[0, i] = FPSR(S, S_hat)
model_fnsr[0, i] = FNSR(S, S_hat)
log_params = False
elif args.algo == "SPINN":
# https://github.com/jjfeng/spinn
log_params = False
print("Not yet implemented!")
else:
print("Sorry! No such evaluation exists.")
if log_params:
# Mean squared errors
model_msfe[0, i] = compute_mse_compare(
model.predict(X_train).reshape(T_train.shape), T_train)
model_mspe[0, i] = compute_mse_compare(
model.predict(X_test).reshape(T_test.shape), T_test)
# Selection rate errors
model_fpsr[0, i] = FPSR(S, S_hat)
model_fnsr[0, i] = FNSR(S, S_hat)
# Cardinality of the model
model_card[0, i] = len(S_hat)
# Normalized Error (NME)
model_nme_train[0, i] = compute_nme(
model.predict(X_train).reshape(T_train.shape), T_train)
model_nme_test[0, i] = compute_nme(
model.predict(X_test).reshape(T_test.shape), T_test)
if args.algo == "BART":
val = model.predict(X_train)
normalized = (val - min(val)) / (max(val) - min(val))
accuracy = np.sum([
abs(0.9 * normalized - T_train.flatten()) < 0.2
]) / len(T_train.flatten())
print("**********The train accuracy is: ", accuracy)
else:
print(
"**********The train accuracy is: ",
calculate_accuracy(
model.predict(X_train).reshape(T_train.shape).T,
T_train.T))
if args.algo == "BART":
val = model.predict(X_test)
normalized = (val - min(val)) / (max(val) - min(val))
accuracy = np.sum([abs(0.9 * normalized - T_test.flatten()) < 0.2
]) / len(T_test.flatten())
print("**********The test accuracy is: ", accuracy)
else:
print(
"**********The test accuracy is: ",
calculate_accuracy(
model.predict(X_test).reshape(T_test.shape).T, T_test.T))
print("Time taken for this MC iteration: ", time.time() - start_time)