def main(index_exp=0):
dirName = '%s_%s_data%d_%s' % (args.ext_model, args.rgr_model,
args.data_cate, args.append_name)
fileName = '%s_exp%d' % (dirName, index_exp)
# Create folder for results of this model
if not os.path.exists('./results/%s' % (dirName)):
os.makedirs('./results/%s' % (dirName))
print('Extraction model: %s' % (args.ext_model))
print('Regression model: %s' % (args.rgr_model))
if args.ext_model == 'vgg16':
net = tv_models.vgg16(pretrained=True).to(device=device)
set_parameter_requires_grad(net, True)
net.classifier[6] = Identity()
elif args.ext_model == 'resnet50':
net = tv_models.resnet50(pretrained=True).to(device=device)
set_parameter_requires_grad(net, True)
net.fc = Identity()
# Get dataset
batchSize = 64
input_size = 224
# Load Data
data_transforms = {
'train': transforms.Compose([ndl.Rescale(input_size),
ndl.ToTensor()]),
'test': transforms.Compose([ndl.Rescale(input_size),
ndl.ToTensor()])
}
print("Initializing Datasets and Dataloaders...")
# Create training and testing datasets
image_datasets = {
x: ndl.TopoplotLoader(args.image_folder,
x,
transform=data_transforms[x],
index_exp=index_exp)
for x in ['train', 'test']
}
# Create training and testing dataloaders
dataloaders_dict = {
'train':
Data.DataLoader(image_datasets['train'],
batch_size=batchSize,
shuffle=False,
num_workers=4),
'test':
Data.DataLoader(image_datasets['test'],
batch_size=batchSize,
shuffle=False,
num_workers=4)
}
# Extract features by VGG16
net.eval() # Disable batchnorm, dropout
X_train, Y_train = extract_layer(dataloaders_dict['train'], net)
X_test, Y_test = extract_layer(dataloaders_dict['test'], net)
# Standardize data before PCA
if args.scale:
X_train, X_test = preprocessing.scale(X_train, X_test, mode=args.scale)
# Apply PCA to reduce dimension
if args.n_components > 1:
args.n_components = int(args.n_components)
pca = PCA(n_components=args.n_components, svd_solver='full')
pca.fit(X_train)
X_train = pca.transform(X_train)
X_test = pca.transform(X_test)
print('(X) Number of features after PCA: %d' % (X_train.shape[1]))
print('(X) Explained variance ratio: %.3f' %
(np.sum(pca.explained_variance_ratio_)))
# Add conditional entropy
if args.add_CE and args.data_cate == 2:
print('Add conditional entropy as additional features...')
with open(
'./raw_data/CE_sub%d_channel21_exp%d_train.data' %
(args.subject_ID, index_exp), 'rb') as fp:
CE_train = pickle.load(fp)
with open(
'./raw_data/CE_sub%d_channel21_exp%d_test.data' %
(args.subject_ID, index_exp), 'rb') as fp:
CE_test = pickle.load(fp)
# Scale CE
CE_train, CE_test = preprocessing.scale(CE_train, CE_test)
# Apply PCA
pca = PCA(n_components=30, svd_solver='full')
pca.fit(CE_train)
CE_train = pca.transform(CE_train)
CE_test = pca.transform(CE_test)
print('(CE) Number of features after PCA: %d' % (CE_train.shape[1]))
print('(CE) Explained variance ratio: %.3f' %
(np.sum(pca.explained_variance_ratio_)))
# Concatentate with X
X_train = np.concatenate((X_train, CE_train), axis=1)
X_test = np.concatenate((X_test, CE_test), axis=1)
# Regression to predict solution latency
X_train_Reg = X_train
X_test_Reg = X_test
if args.rgr_model == 'LR':
rgr = linear_model.LinearRegression()
elif args.rgr_model == 'Ridge':
rgr = linear_model.Ridge(alpha=1)
elif args.rgr_model == 'GPR':
kernel = RBF(10, (1e-2, 1e2)) + ConstantKernel(10, (1e-2, 1e2))
rgr = GaussianProcessRegressor(kernel=kernel, random_state=0)
elif args.rgr_model == 'ELMK':
rgr = elm.ELMKernel()
elif args.rgr_model == 'ELMR':
params = ["sigmoid", 1, 500, False]
rgr = elm.ELMRandom(params)
if args.rgr_model not in ['ELMK', 'ELMR']:
rgr.fit(X_train_Reg, Y_train)
pred_train = rgr.predict(X_train_Reg)
pred_test = rgr.predict(X_test_Reg)
else:
# Scale target into -1~1
if args.scale_target == 2:
scaler = TargetScaler(num_step=10)
scaler.fit(Y_train)
Y_train, Y_test = scaler.transform(Y_train), scaler.transform(
Y_test)
elif args.scale_target == 1:
Y_train, Y_test = (Y_train - 30) / 30, (Y_test - 30) / 30
# Concatenate data for extreme learning machine
train_data = np.concatenate((Y_train[:, np.newaxis], X_train), axis=1)
test_data = np.concatenate((Y_test[:, np.newaxis], X_test), axis=1)
rgr.search_param(train_data, cv="kfold", of="rmse", eval=10)
pred_train = rgr.train(train_data).predicted_targets
pred_test = rgr.test(test_data).predicted_targets
# Scale target back to 0~60
if args.scale_target == 2:
[Y_train, Y_test, pred_train, pred_test] = [scaler.transform(x, mode='inverse') for x in \
[Y_train, Y_test, pred_train, pred_test]]
elif args.scale_target == 1:
[Y_train, Y_test, pred_train, pred_test] = [x*30+30 for x in \
[Y_train, Y_test, pred_train, pred_test]]
evaluate_result.plot_scatter(Y_test, pred_test, dirName, fileName)
print('Train std: %.3f' % (mean_squared_error(Y_train, pred_train)**0.5))
print('Test std: %.3f' % (mean_squared_error(Y_test, pred_test)**0.5))
# Save targets and predictions
dict_target = {}
dict_target['target'], dict_target['pred'] = Y_test, pred_test
with open('./results/%s/%s.data' % (dirName, fileName), 'wb') as fp:
pickle.dump(dict_target, fp)
return
