def run_cnn(dataset_dir):
# utilities.random_seed_np_tf(FLAGS.seed)
utilities.random_seed_np_tf(-1)
# common parameters
cifar_dir = dataset_dir
num_epochs = 3
batch_size = 64
n_classes = 10
# parameters for cnn
name_cnn = 'cnn'
original_shape_cnn = '32,32,3'
layers_cnn = 'conv2d-5-5-32-1,maxpool-2,conv2d-5-5-64-1,maxpool-2,full-1024,softmax'
loss_func_cnn = 'softmax_cross_entropy'
opt_cnn = 'adam'
learning_rate_cnn = 1e-4
momentum_cnn = 0.5 # not used
dropout_cnn = 0.5
batch_norm = True
# prepare data
trX, trY, teX, teY = datasets.load_cifar10_dataset(cifar_dir, mode='supervised')
# due to the memory limit, cannot use the whole training set
trY_non_one_hot = trY
trY = np.array(utilities.to_one_hot(trY))
teY = np.array(teY)
teY_non_one_hot = teY[5000:]
teY = np.array(utilities.to_one_hot(teY))
# first half test set is validation set
vlX = teX[:5000]
vlY = teY[:5000]
teX = teX[5000:]
teY = teY[5000:]
# define Convolutional Network
cnn = conv_net.ConvolutionalNetwork(
original_shape=[int(i) for i in original_shape_cnn.split(',')],
layers=layers_cnn, name=name_cnn, loss_func=loss_func_cnn,
num_epochs=num_epochs, batch_size=batch_size, opt=opt_cnn,
learning_rate=learning_rate_cnn, momentum=momentum_cnn, dropout=dropout_cnn,
batch_norm = batch_norm
)
print('Start Convolutional Network training...')
cnn.fit(trX, trY, vlX, vlY) # supervised learning
if FLAGS.dataset == 'mnist':
# ################# #
# MNIST Dataset #
# ################# #
trX, trY, vlX, vlY, teX, teY = datasets.load_mnist_dataset(
mode='supervised')
elif FLAGS.dataset == 'cifar10':
# ################### #
# Cifar10 Dataset #
# ################### #
trX, trY, teX, teY = datasets.load_cifar10_dataset(FLAGS.cifar_dir,
mode='supervised')
vlX = teX[:5000] # Validation set is the first half of the test set
vlY = teY[:5000]
elif FLAGS.dataset == 'custom':
# ################## #
# Custom Dataset #
# ################## #
def load_from_np(dataset_path):
if dataset_path != '':
return np.load(dataset_path)
else:
return None
# ################# #
# MNIST Dataset #
# ################# #
trX, vlX, teX = datasets.load_mnist_dataset(mode='unsupervised')
trRef = trX
vlRef = vlX
teRef = teX
elif FLAGS.dataset == 'cifar10':
# ################### #
# Cifar10 Dataset #
# ################### #
trX, teX = datasets.load_cifar10_dataset(FLAGS.cifar_dir,
mode='unsupervised')
# Validation set is the first half of the test set
vlX = teX[:5000]
trRef = trX
vlRef = vlX
teRef = teX
elif FLAGS.dataset == 'custom':
# ################## #
# Custom Dataset #
# ################## #
def load_from_np(dataset_path):
if dataset_path != '':
return np.load(dataset_path)
if FLAGS.dataset == 'mnist':
# ################# #
# MNIST Dataset #
# ################# #
trX, vlX, teX = datasets.load_mnist_dataset(mode='unsupervised')
width, height = 28, 28
elif FLAGS.dataset == 'cifar10':
# ################### #
# Cifar10 Dataset #
# ################### #
trX, teX = datasets.load_cifar10_dataset(FLAGS.cifar_dir, mode='unsupervised')
vlX = teX[:5000] # Validation set is the first half of the test set
width, height = 32, 32
elif FLAGS.dataset == 'custom':
# ################## #
# Custom Dataset #
# ################## #
def load_from_np(dataset_path):
if dataset_path != '':
return np.load(dataset_path)
else:
return None
if FLAGS.dataset == 'mnist':
# ################# #
# MNIST Dataset #
# ################# #
trX, trY, vlX, vlY, teX, teY = datasets.load_mnist_dataset(mode='supervised')
elif FLAGS.dataset == 'cifar10':
# ################### #
# Cifar10 Dataset #
# ################### #
trX, trY, teX, teY = datasets.load_cifar10_dataset(FLAGS.cifar_dir, mode='supervised')
vlX = teX[:5000] # Validation set is the first half of the test set
vlY = teY[:5000]
elif FLAGS.dataset == 'custom':
# ################## #
# Custom Dataset #
# ################## #
def load_from_np(dataset_path):
if dataset_path != '':
return np.load(dataset_path)
else:
return None
def generate_feature_sets(dataset_dir, fs_filename, tr_size):
# utilities.random_seed_np_tf(FLAGS.seed)
utilities.random_seed_np_tf(-1)
# common parameters
cifar_dir = dataset_dir
num_epochs = 3
batch_size = 64
n_classes = 10
# parameters for dae
name_dae = 'dae'
n_components_dae = 1024
enc_act_func_dae = tf.nn.sigmoid
dec_act_func_dae = tf.nn.sigmoid
corr_type_dae = 'masking'
corr_frac_dae = 0.5
loss_func_dae = 'cross_entropy'
opt_dae = 'momentum'
regcoef_dae = 5e-4
learning_rate_dae = 0.05
momentum_dae = 0.9
# parameters for cnn
name_cnn = 'cnn'
original_shape_cnn = '32,32,3'
layers_cnn = 'conv2d-5-5-32-1,maxpool-2,conv2d-5-5-64-1,maxpool-2,full-1024,softmax'
loss_func_cnn = 'softmax_cross_entropy'
opt_cnn = 'adam'
learning_rate_cnn = 1e-4
momentum_cnn = 0.5 # not used
dropout_cnn = 0.5
# loading data
trX, trY, teX, teY = datasets.load_cifar10_dataset(cifar_dir,
mode='supervised')
# due to the memory limit, cannot use the whole training set
trX = trX[:tr_size]
trY = trY[:tr_size]
trY_non_one_hot = trY
trY = np.array(utilities.to_one_hot(trY))
teY = np.array(teY)
teY_non_one_hot = teY[5000:]
teY = np.array(utilities.to_one_hot(teY))
# first half test set is validation set
vlX = teX[:5000]
vlY = teY[:5000]
teX = teX[5000:]
teY = teY[5000:]
fs_file = open(fs_filename, 'wb')
pickle.dump(trY_non_one_hot, fs_file)
pickle.dump(teY_non_one_hot, fs_file)
# define Denoising Autoencoder
dae = denoising_autoencoder.DenoisingAutoencoder(
name=name_dae,
n_components=n_components_dae,
enc_act_func=enc_act_func_dae,
dec_act_func=dec_act_func_dae,
corr_type=corr_type_dae,
corr_frac=corr_frac_dae,
loss_func=loss_func_dae,
opt=opt_dae,
regcoef=regcoef_dae,
learning_rate=learning_rate_dae,
momentum=momentum_dae,
num_epochs=num_epochs,
batch_size=batch_size)
print('Start Denoising Autoencoder training...')
dae.fit(trX, trX, vlX, vlX) # unsupervised learning
feature_train_set_1 = dae.extract_features(trX)
pickle.dump(feature_train_set_1, fs_file)
feature_test_set_1 = dae.extract_features(teX)
pickle.dump(feature_test_set_1, fs_file)
# define Convolutional Network
cnn = conv_net.ConvolutionalNetwork(
original_shape=[int(i) for i in original_shape_cnn.split(',')],
layers=layers_cnn,
name=name_cnn,
loss_func=loss_func_cnn,
num_epochs=num_epochs,
batch_size=batch_size,
opt=opt_cnn,
learning_rate=learning_rate_cnn,
momentum=momentum_cnn,
dropout=dropout_cnn)
print('Start Convolutional Network training...')
cnn.fit(trX, trY, vlX, vlY) # supervised learning
feature_train_set_2 = cnn.extract_features(trX)
pickle.dump(feature_train_set_2, fs_file)
feature_test_set_2 = cnn.extract_features(teX)
pickle.dump(feature_test_set_2, fs_file)
fs_file.close()
