def main():
rubrics = [
'Мир', 'Россия', 'Политика', 'Экономика', 'Наука и техника', 'Украина',
'Госэкономика', 'Спорт', 'Общество', 'Бывший СССР', 'Культура',
'Медиа', 'Футбол', 'Музыка', 'Наука'
]
lenta = Dataset(use_title=False,
rubrics=rubrics,
random_state=SEED,
subsample=0.3)
X, y = lenta.get_data()
models = [
LogisticRegression(random_state=SEED),
GaussianNB(),
SVC(),
RandomForestClassifier(),
XGBClassifier()
]
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=SEED)
transformations = ['elmo']
with tf.device("/gpu:0"):
# Setup operations
for trans in transformations:
for model in models:
clf = model
pipeline = lenta.get_transform_pipeline(clf,
trans,
standardize=False)
scores = cross_val_score(pipeline,
X,
y,
verbose=10,
cv=cv,
n_jobs=1)
mean = np.mean(scores)
std = np.std(scores)
res = {'mean': round(mean, 3), 'std': round(std, 3)}
msg = 'Transformation: {}, Model: {}, accuracy {:.3f}(+- {:.3f})'
print(msg.format(trans, model, mean, std))
save_result(type(model).__name__ + '_' + trans, res)
n_test = 480
batch_size = 480
im_shape = (32, 32)
elif data_name == 'MNIST':
n_train = 10000
n_test = 5000
batch_size = 1000
im_shape = (28, 28)
d_componentsss = [2]
train_RBM = False # only true if the RBM are not trained yet
train_NN = False
show_plot = True
noisy = False # if data should be noisy
epochs = 50
X, y, X_train, y_train, X_test, y_test = dataset.get_data(name=data_name,
n_train=n_train,
n_test=n_test)
# hyperparameters
thetas = [0.9] #[0.1, 0.3, 0.5, 0.7, 0.9, 0.99]
c = 0.1
# training RBM
for d_components in d_componentsss:
file_path_weights = 'Models/weightsRBM/' + data_name + '/TEST' + data_name + str(
n_train) + 'dim' + str(d_components)
if data_name == 'MNIST':
layer_dim = [784, 500, 500, 2000, d_components]
elif data_name == 'COIL20':
layer_dim = [1024, 500, 500, 2000, d_components]
# initialization dataset/model
seed = 0
dataset = Dataset(seed)
model_type = 'reg' # par/reg/auto/PCA/kernel
data_name = 'COIL20' # MNIST/ COIL20
if data_name == 'COIL20':
n_train = 960
n_test = 480
batch_size = 480
elif data_name == 'MNIST':
n_train = 10000
n_test = 5000
batch_size = 1000
d_componentss = [2, 10, 20]
epochs = 30
X, labels, X_train, labels_train, X_test, labels_test = dataset.get_data(
name=data_name, n_train=n_train, n_test=n_test)
nsplits = 3
kfold = KFold(n_splits=nsplits)
for d_components in d_componentss:
RBM_file = 'Models/weightsRBM/' + data_name + '/' + data_name + str(
n_train) + 'dim' + str(d_components)
layer_dim = [784, 500, 500, 2000, d_components]
lambdas = [0.01, 0.1, 0.3, 0.5, 0.7, 0.9, 0.99]
accuracies = np.zeros(len(lambdas))
losses = []
for i, labda in enumerate(lambdas):
print('Now training for lambda = %.2f ' % (labda))
accuracy = 0
print("Gradient descent took %.4f seconds" % (time() - t0))
return Y, cost, grad_value
if __name__ == '__main__':
'''
main to implement tSNE on the datasets
'''
seed = 0
dataset = Dataset(seed)
d_components = [2]
data_name = 'COIL20'
grad_method = 'gains'
n_train = 960
X, y, X_train, y_train, X_test, y_test = dataset.get_data(
data_name, n_train, 10000)
#X, y, X_train, y_train, X_test, y_test = dataset.get_coil20_data()
for d in d_components:
if grad_method == 'ADAM':
model = tsne(random_state=0,
initialization='PCA',
initial_dims=30,
grad_method='ADAM',
perplexity=40,
max_iter=1000,
d_components=d,
learning_rate=0.1)
elif grad_method == 'gains':
model = tsne(random_state=0,
initialization='PCA',