This is an implementation of Ladder Network in Keras. Ladder network is a model for semi-supervised learning. Refer to the paper titled Semi-Supervised Learning with Ladder Networks by A Rasmus, H Valpola, M Honkala,M Berglund, and T Raiko
This implementation was used in the official code of our paper Unsupervised Clustering using Pseudo-semi-supervised Learning . The code can be found here and the blog post can be found here
The model achives 98% test accuracy on MNIST with just 100 labeled examples.
The code only works with Tensorflow backend.
- Python 2.7+/3.6+
- Tensorflow (1.4.0)
- numpy
- keras (2.1.4)
Note that other versions of tensorflow/keras should also work.
Load the dataset
from keras.datasets import mnist
import keras
import random
# get the dataset
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train.reshape(60000, 28*28).astype('float32')/255.0
x_test = x_test.reshape(10000, 28*28).astype('float32')/255.0
y_train = keras.utils.to_categorical( y_train )
y_test = keras.utils.to_categorical( y_test )
# only select 100 training samples
idxs_annot = range( x_train.shape[0])
random.seed(0)
random.shuffle( idxs_annot )
idxs_annot = idxs_annot[ :100 ]
x_train_unlabeled = x_train
x_train_labeled = x_train[ idxs_annot ]
y_train_labeled = y_train[ idxs_annot ]Repeat the labeled dataset to match the shapes
n_rep = x_train_unlabeled.shape[0] / x_train_labeled.shape[0]
x_train_labeled_rep = np.concatenate([x_train_labeled]*n_rep)
y_train_labeled_rep = np.concatenate([y_train_labeled]*n_rep)Initialize the model
from ladder_net import get_ladder_network_fc
inp_size = 28*28 # size of mnist dataset
n_classes = 10
model = get_ladder_network_fc( layer_sizes = [ inp_size , 1000, 500, 250, 250, 250, n_classes ] )Train the model
model.fit([ x_train_labeled_rep , x_train_unlabeled ] , y_train_labeled_rep , epochs=100)Get the test accuracy
from sklearn.metrics import accuracy_score
y_test_pr = model.test_model.predict(x_test , batch_size=100 )
print "test accuracy" , accuracy_score(y_test.argmax(-1) , y_test_pr.argmax(-1) )