Cannot make this autoencoder network function properly (with convolutional and maxpool layers)

Question

Autoencoder networks seems to be way trickier than normal classifier MLP networks. After several attempts using Lasagne all what I get in the reconstructed output is something that resembles at its best a blurry averaging of all the images of the MNIST database without distinction on what the input digit actually is.

The networks structure I chose are the following cascade layers:

1. input layer (28x28)
2. 2D convolutional layer, filter size 7x7
3. Max Pooling layer, size 3x3, stride 2x2
4. Dense (fully connected) flattening layer, 10 units (this is the bottleneck)
5. Dense (fully connected) layer, 121 units
6. Reshaping layer to 11x11
7. 2D convolutional layer, filter size 3x3
8. 2D Upscaling layer factor 2
9. 2D convolutional layer, filter size 3x3
10. 2D Upscaling layer factor 2
11. 2D convolutional layer, filter size 5x5
12. Feature max pooling (from 31x28x28 to 28x28)

All the 2D convolutional layers have the biases untied, sigmoid activations and 31 filters.

All the fully connected layers have sigmoid activations.

The loss function used is squared error, the updating function is adagrad. The length of the chunk for the learning is 100 samples, multiplied for 1000 epochs.

The following is an illustration of the problem: the upper row are some samples set as inputs of the network, the lower row is the reconstruction:

Just for completeness, the following is the code I used:

import theano.tensor as T
import theano
import sys
sys.path.insert(0,'./Lasagne') # local checkout of Lasagne
import lasagne
from theano import pp
from theano import function
import gzip
import numpy as np
from sklearn.preprocessing import OneHotEncoder
import matplotlib.pyplot as plt
def load_mnist():

    def load_mnist_images(filename):
        with gzip.open(filename, 'rb') as f:
            data = np.frombuffer(f.read(), np.uint8, offset=16)
        # The inputs are vectors now, we reshape them to monochrome 2D images,
        # following the shape convention: (examples, channels, rows, columns)
        data = data.reshape(-1, 1, 28, 28)
        # The inputs come as bytes, we convert them to float32 in range [0,1].
        # (Actually to range [0, 255/256], for compatibility to the version
        # provided at http://deeplearning.net/data/mnist/mnist.pkl.gz.)
        return data / np.float32(256)

    def load_mnist_labels(filename):
        # Read the labels in Yann LeCun's binary format.
        with gzip.open(filename, 'rb') as f:
            data = np.frombuffer(f.read(), np.uint8, offset=8)
        # The labels are vectors of integers now, that's exactly what we want.
        return data

    X_train = load_mnist_images('train-images-idx3-ubyte.gz')
    y_train = load_mnist_labels('train-labels-idx1-ubyte.gz')
    X_test = load_mnist_images('t10k-images-idx3-ubyte.gz')
    y_test = load_mnist_labels('t10k-labels-idx1-ubyte.gz')
    return X_train, y_train, X_test, y_test

def plot_filters(conv_layer):
    W = conv_layer.get_params()[0]
    W_fn = theano.function([],W)
    params = W_fn()
    ks = np.squeeze(params)
    kstack = np.vstack(ks)
    plt.imshow(kstack,interpolation='none')
    plt.show()

def main():

    #theano.config.exception_verbosity="high"
    #theano.config.optimizer='None'

    X_train, y_train, X_test, y_test = load_mnist()
    ohe = OneHotEncoder()

    y_train = ohe.fit_transform(np.expand_dims(y_train,1)).toarray()
    chunk_len = 100
    visamount = 10
    num_epochs = 1000
    num_filters=31
    dropout_p=.0
    print "X_train.shape",X_train.shape,"y_train.shape",y_train.shape
    input_var = T.tensor4('X')
    output_var = T.tensor4('X')
    conv_nonlinearity = lasagne.nonlinearities.sigmoid
    net = lasagne.layers.InputLayer((chunk_len,1,28,28), input_var)
    conv1 = net = lasagne.layers.Conv2DLayer(net,num_filters,(7,7),nonlinearity=conv_nonlinearity,untie_biases=True)
    net = lasagne.layers.MaxPool2DLayer(net,(3,3),stride=(2,2))
    net = lasagne.layers.DropoutLayer(net,p=dropout_p)
    #conv2_layer = lasagne.layers.Conv2DLayer(dropout_layer,num_filters,(3,3),nonlinearity=conv_nonlinearity)
    #pool2_layer = lasagne.layers.MaxPool2DLayer(conv2_layer,(3,3),stride=(2,2))
    net = lasagne.layers.DenseLayer(net,10,nonlinearity=lasagne.nonlinearities.sigmoid)

    #augment_layer1 = lasagne.layers.DenseLayer(reduction_layer,33,nonlinearity=lasagne.nonlinearities.sigmoid)
    net = lasagne.layers.DenseLayer(net,121,nonlinearity=lasagne.nonlinearities.sigmoid)

    net = lasagne.layers.ReshapeLayer(net,(chunk_len,1,11,11))

    net = lasagne.layers.Conv2DLayer(net,num_filters,(3,3),nonlinearity=conv_nonlinearity,untie_biases=True)
    net = lasagne.layers.Upscale2DLayer(net,2)

    net = lasagne.layers.Conv2DLayer(net,num_filters,(3,3),nonlinearity=conv_nonlinearity,untie_biases=True)
    #pool_after0 = lasagne.layers.MaxPool2DLayer(conv_after1,(3,3),stride=(2,2))
    net = lasagne.layers.Upscale2DLayer(net,2)

    net = lasagne.layers.DropoutLayer(net,p=dropout_p)

    #conv_after2 = lasagne.layers.Conv2DLayer(upscale_layer1,num_filters,(3,3),nonlinearity=conv_nonlinearity,untie_biases=True)
    #pool_after1 = lasagne.layers.MaxPool2DLayer(conv_after2,(3,3),stride=(1,1))
    #upscale_layer2 = lasagne.layers.Upscale2DLayer(pool_after1,4)

    net = lasagne.layers.Conv2DLayer(net,num_filters,(5,5),nonlinearity=conv_nonlinearity,untie_biases=True)
    net = lasagne.layers.FeaturePoolLayer(net,num_filters,pool_function=theano.tensor.max)
    print "output_shape:",lasagne.layers.get_output_shape(net)
    params = lasagne.layers.get_all_params(net, trainable=True)
    prediction = lasagne.layers.get_output(net)
    loss = lasagne.objectives.squared_error(prediction, output_var)
    #loss = lasagne.objectives.binary_crossentropy(prediction, output_var)
    aggregated_loss = lasagne.objectives.aggregate(loss)
    updates = lasagne.updates.adagrad(aggregated_loss,params)
    train_fn = theano.function([input_var, output_var], loss, updates=updates)

    test_prediction = lasagne.layers.get_output(net, deterministic=True)
    predict_fn = theano.function([input_var], test_prediction)

    print "starting training..."
    for epoch in range(num_epochs):
        selected = list(set(np.random.random_integers(0,59999,chunk_len*4)))[:chunk_len]
        X_train_sub = X_train[selected,:]
        _loss = train_fn(X_train_sub, X_train_sub)
        print("Epoch %d: Loss %g" % (epoch + 1, np.sum(_loss) / len(X_train)))
        """
        chunk = X_train[0:chunk_len,:,:,:]
        result = predict_fn(chunk)
        vis1 = np.hstack([chunk[j,0,:,:] for j in range(visamount)])
        vis2 = np.hstack([result[j,0,:,:] for j in range(visamount)])
        plt.imshow(np.vstack([vis1,vis2]))
        plt.show()
        """
    print "done."

    chunk = X_train[0:chunk_len,:,:,:]
    result = predict_fn(chunk)
    print "chunk.shape",chunk.shape
    print "result.shape",result.shape
    plot_filters(conv1)
    for i in range(chunk_len/visamount):
        vis1 = np.hstack([chunk[i*visamount+j,0,:,:] for j in range(visamount)])
        vis2 = np.hstack([result[i*visamount+j,0,:,:] for j in range(visamount)])
        plt.imshow(np.vstack([vis1,vis2]))
        plt.show()
    import ipdb; ipdb.set_trace()

if __name__ == "__main__":
    main()

Any ideas on how to improve this network to get a reasonably functioning autoencoder?

Problem solved!

With an implementation that is quite different, using a leaky rectifier instead of a sigmoid function in the convolutional layers, only 2 (!!) nodes in the bottleneck layer and convolutions with 1x1 kernels at the very end.

Here is the result of some reconstruction:

Code:

import theano.tensor as T
import theano
import sys
sys.path.insert(0,'./Lasagne') # local checkout of Lasagne
import lasagne
from theano import pp
from theano import function
import theano.tensor.nnet
import gzip
import numpy as np
from sklearn.preprocessing import OneHotEncoder
import matplotlib.pyplot as plt
def load_mnist():

    def load_mnist_images(filename):
        with gzip.open(filename, 'rb') as f:
            data = np.frombuffer(f.read(), np.uint8, offset=16)
        # The inputs are vectors now, we reshape them to monochrome 2D images,
        # following the shape convention: (examples, channels, rows, columns)
        data = data.reshape(-1, 1, 28, 28)
        # The inputs come as bytes, we convert them to float32 in range [0,1].
        # (Actually to range [0, 255/256], for compatibility to the version
        # provided at http://deeplearning.net/data/mnist/mnist.pkl.gz.)
        return data / np.float32(256)

    def load_mnist_labels(filename):
        # Read the labels in Yann LeCun's binary format.
        with gzip.open(filename, 'rb') as f:
            data = np.frombuffer(f.read(), np.uint8, offset=8)
        # The labels are vectors of integers now, that's exactly what we want.
        return data

    X_train = load_mnist_images('train-images-idx3-ubyte.gz')
    y_train = load_mnist_labels('train-labels-idx1-ubyte.gz')
    X_test = load_mnist_images('t10k-images-idx3-ubyte.gz')
    y_test = load_mnist_labels('t10k-labels-idx1-ubyte.gz')
    return X_train, y_train, X_test, y_test

def main():

    X_train, y_train, X_test, y_test = load_mnist()
    ohe = OneHotEncoder()

    y_train = ohe.fit_transform(np.expand_dims(y_train,1)).toarray()
    chunk_len = 100
    num_epochs = 10000
    num_filters=7
    input_var = T.tensor4('X')
    output_var = T.tensor4('X')
    #conv_nonlinearity = lasagne.nonlinearities.sigmoid
    #conv_nonlinearity = lasagne.nonlinearities.rectify
    conv_nonlinearity = lasagne.nonlinearities.LeakyRectify(.1)
    softplus = theano.tensor.nnet.softplus
    #conv_nonlinearity = theano.tensor.nnet.softplus
    net = lasagne.layers.InputLayer((chunk_len,1,28,28), input_var)
    conv1 = net = lasagne.layers.Conv2DLayer(net,num_filters,(7,7),nonlinearity=conv_nonlinearity,untie_biases=True)
    net = lasagne.layers.MaxPool2DLayer(net,(3,3),stride=(2,2))
    net = lasagne.layers.DenseLayer(net,2,nonlinearity=lasagne.nonlinearities.sigmoid)
    net = lasagne.layers.DenseLayer(net,49,nonlinearity=lasagne.nonlinearities.sigmoid)
    net = lasagne.layers.ReshapeLayer(net,(chunk_len,1,7,7))
    net = lasagne.layers.Conv2DLayer(net,num_filters,(3,3),nonlinearity=conv_nonlinearity,untie_biases=True)
    net = lasagne.layers.MaxPool2DLayer(net,(3,3),stride=(1,1))
    net = lasagne.layers.Upscale2DLayer(net,4)
    net = lasagne.layers.Conv2DLayer(net,num_filters,(3,3),nonlinearity=conv_nonlinearity,untie_biases=True)
    net = lasagne.layers.MaxPool2DLayer(net,(3,3),stride=(1,1))
    net = lasagne.layers.Upscale2DLayer(net,4)
    net = lasagne.layers.Conv2DLayer(net,num_filters,(5,5),nonlinearity=conv_nonlinearity,untie_biases=True)
    net = lasagne.layers.Conv2DLayer(net,num_filters,(1,1),nonlinearity=conv_nonlinearity,untie_biases=True)
    net = lasagne.layers.FeaturePoolLayer(net,num_filters,pool_function=theano.tensor.max)
    net = lasagne.layers.Conv2DLayer(net,1,(1,1),nonlinearity=conv_nonlinearity,untie_biases=True)
    print "output shape:",net.output_shape
    params = lasagne.layers.get_all_params(net, trainable=True)
    prediction = lasagne.layers.get_output(net)
    loss = lasagne.objectives.squared_error(prediction, output_var)
    #loss = lasagne.objectives.binary_hinge_loss(prediction, output_var)
    aggregated_loss = lasagne.objectives.aggregate(loss)
    #updates = lasagne.updates.adagrad(aggregated_loss,params)
    updates = lasagne.updates.nesterov_momentum(aggregated_loss,params,0.5)#.005
    train_fn = theano.function([input_var, output_var], loss, updates=updates)

    test_prediction = lasagne.layers.get_output(net, deterministic=True)
    predict_fn = theano.function([input_var], test_prediction)

    print "starting training..."
    for epoch in range(num_epochs):
        selected = list(set(np.random.random_integers(0,59999,chunk_len*4)))[:chunk_len]
        X_train_sub = X_train[selected,:]
        _loss = train_fn(X_train_sub, X_train_sub)
        print("Epoch %d: Loss %g" % (epoch + 1, np.sum(_loss) / len(X_train)))
    print "done."

    chunk = X_train[0:chunk_len,:,:,:]
    result = predict_fn(chunk)
    print "chunk.shape",chunk.shape
    print "result.shape",result.shape
    visamount = 10
    for i in range(10):
        vis1 = np.hstack([chunk[i*visamount+j,0,:,:] for j in range(visamount)])
        vis2 = np.hstack([result[i*visamount+j,0,:,:] for j in range(visamount)])
        plt.imshow(np.vstack([vis1,vis2]))
        plt.show()

    import ipdb; ipdb.set_trace()
if __name__ == "__main__":
    main()

Answer 1

You might gain more insight by visualizing the weights instead of just the reconstructions. I had a similar problem when my biases were misconfigured. Everything below is written based on my experiences writing my own learning library. You can see the code here on Github http://github.com/josephcatrambone/aij.

Here is a screenshot of my program when there are no biases. This is after only maybe ten epochs since I'm in a hurry to finish this writeup:

The weight update is done by these operations:

weights.add_i(positiveProduct.subtract(negativeProduct).elementMultiply(learningRate / (float) batchSize));
//visibleBias.add_i(batch.subtract(negativeVisibleProbabilities).meanRow().elementMultiply(learningRate));
//hiddenBias.add_i(positiveHiddenProbabilities.subtract(negativeHiddenProbabilities).meanRow().elementMultiply(learningRate));

If I uncomment the visible bias code, I get this result:

If I screw up the sign of the visible bias code (subtracting instead of adding):

visibleBias.subtract_i(batch.subtract(negativeVisibleProbabilities).meanRow().elementMultiply(learningRate));

I get this image:

Which snowballs and eventually reaches something like what you have above. Check the signage of your error functions.