def main(num_layer, gamma):
x_train, y_train = du.load_mnist_data('training')
x_test, y_test = du.load_mnist_data('testing')
input_dim = 784
latent_dim = 50
encoder_dim = np.ones([num_layer], np.int32) * 200
decoder_dim = np.ones([num_layer], np.int32) * 200
sample_num = 1
batch_size = 20
stage_num = 8
# model = VaeSimple(input_dim, latent_dim, encoder_dim, decoder_dim, 5000)
model = VaeScale(input_dim,
latent_dim,
encoder_dim,
decoder_dim,
5000,
gamma=gamma)
with tf.Session() as sess:
saver = tf.train.Saver()
model.restore_from(saver, sess, 'model/stage' + str(stage_num - 1))
nll = test_model(model, x_test, batch_size, sess)
print('NLL = {0}.'.format(nll))
def main():
model_dimension = [784, 500, 500, 2000, 10]
model = VAEClustering(100, model_dimension, 10, 'sigmoid')
train_data, train_label = du.load_mnist_data('training')
test_data, test_label = du.load_mnist_data('testing')
# train_model(model, train_data, 100)
test_model(model, train_data, train_label)
test_model(model, test_data, test_label)
def main(exploration_ratio):
model = VAEClustering(100, [784, 500, 500, 2000, 10], 10, 0.01, 'sigmoid')
model.build_network()
optimizer = VAEOptimizer(model)
x_train, y_train = du.load_mnist_data('training')
x_test, y_test = du.load_mnist_data('testing')
train_model(model, optimizer, x_train, y_train, 1000, exploration_ratio)
test_model(model, x_train, y_train)
test_model(model, x_test, y_test)
def main(num_layer, gamma):
x_train, y_train = du.load_mnist_data('training')
x_test, y_test = du.load_mnist_data('testing')
input_dim = 784
latent_dim = 50
encoder_dim = np.ones([num_layer], np.int32) * 200
decoder_dim = np.ones([num_layer], np.int32) * 200
sample_num = 1
batch_size = 20
stage_num = 8
# model = VaeSimple(input_dim, latent_dim, encoder_dim, decoder_dim, sample_num)
model = VaeScale(input_dim,
latent_dim,
encoder_dim,
decoder_dim,
sample_num,
gamma=gamma)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter('graph', sess.graph)
saver = tf.train.Saver()
for stage in range(stage_num):
lr = 0.001 * 10**(-float(stage) / 7.0)
num_epoch = 3**stage
print('Stage = {0}, lr = {1}, num_epoch = {2}.'.format(
stage, lr, num_epoch))
train_model(model, x_train, lr, num_epoch, batch_size, sess,
writer)
model.save_to(saver, sess, 'model/stage' + str(stage))
sess.close()
tf.reset_default_graph()
# model = VaeSimple(input_dim, latent_dim, encoder_dim, decoder_dim, 5000)
model = VaeScale(input_dim,
latent_dim,
encoder_dim,
decoder_dim,
5000,
gamma=gamma)
with tf.Session() as sess:
saver = tf.train.Saver()
model.restore_from(saver, sess, 'model/stage' + str(stage_num - 1))
nll = test_model(model, x_test, batch_size, sess)
print('NLL = {0}.'.format(nll))
def main():
images_train, labels_train = du.load_mnist_data('training')
# images_test, labels_test = load_mnist_data('testing')
model = SaeNet(DIM_LAYER, 100, 0.0001, 'sigmoid')
model.build_graph()
train_model(model, images_train, labels_train)
def main():
# exp info
save_dir = os.path.join('experiments', 'save')
if not os.path.exists(save_dir):
os.mkdir(save_dir)
exp_id = args.exp_id if args.exp_id != '' else str(
len(os.listdir(save_dir)))
save_path = os.path.join(save_dir, exp_id)
if not os.path.exists(save_path):
os.mkdir(save_path)
summary_dir = os.path.join('experiments', 'summary')
if not os.path.exists(summary_dir):
os.mkdir(summary_dir)
summary_path = os.path.join(summary_dir, exp_id)
if not os.path.exists(summary_path):
os.mkdir(summary_path)
# prepare data
x_train, y_train = load_mnist_data('training')
x_test, y_test = load_mnist_data('testing')
print(np.shape(x_train))
print(np.shape(x_test))
# model
model = VaeModel(args.num_hidden, 1, args.shortcut, args.dense,
args.continuous)
optimizer = VaeOptimizer(model)
# train
if not args.val:
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter(summary_path, sess.graph)
saver = tf.train.Saver()
for stage in range(args.num_stage):
lr = 0.001 * math.pow(0.1,
float(stage) / float(args.num_stage - 1))
epochs = math.pow(3, stage)
kl_coef = 0.001 * math.pow(
1000,
float(stage) /
float(args.num_stage - 1)) if args.warmup else 1.0
print('Stage = {}, lr = {:.6f}, kl_coef = {:.4f} epochs = {}.'.
format(stage, lr, kl_coef, epochs))
train(optimizer, sess, x_train, epochs, lr, kl_coef, writer)
saver.save(sess, os.path.join(save_path, str(stage)))
tf.reset_default_graph()
model = VaeModel(args.num_hidden, 5000, args.shortcut, args.dense,
args.continuous)
with tf.Session() as sess:
saver = tf.train.Saver()
saver.restore(sess, os.path.join(save_path, str(args.num_stage - 1)))
nll = test_nll(model, sess, x_test)
print('NLL = {:.4f}.'.format(nll))
wo = model.wo.eval(sess)
wo_norm = np.sum(np.square(wo), -1)
wo_norm = -np.sort(-wo_norm)
fid = open(os.path.join(save_path, 'wo.txt'), 'wt')
for w in wo_norm:
fid.write('{:.4f}\n'.format(w))
fid.close()
print('Num active: {}.'.format(
np.sum(wo_norm > 0.05 * np.max(wo_norm))))