sparsityParam = 0.01 # desired average activation of the hidden units.
lambdaParam = 0.0001 # weight decay parameter
betaParam = 3 # weight of sparsity penalty term
def sparseAutoencoderCostCallback(x):
return sparse_autoencoder.cost(x, visibleSize, hiddenSize, lambdaParam,
sparsityParam, betaParam, patches)
patches = getPatches(numPatches=10000, patchSize=patchSize)
thetaParam = sparse_autoencoder.initializeParameters(hiddenSize, visibleSize)
options = {
'maxiter': 400,
'disp': True,
}
result = optimize.minimize(sparseAutoencoderCostCallback,
thetaParam,
method='L-BFGS-B',
jac=True,
options=options)
W1 = result.x[0:hiddenSize * visibleSize].reshape(hiddenSize, visibleSize)
image_filename = 'images.png'
print 'Saving learned features to %s' % image_filename
visualizeNetwork(W1.T, image_filename)
from visualize_network import visualizeNetwork
import sparse_autoencoder
patchSize=8
visibleSize = patchSize*patchSize # number of input units
hiddenSize = 25 # number of hidden units
sparsityParam = 0.01 # desired average activation of the hidden units.
lambdaParam = 0.0001 # weight decay parameter
betaParam = 3 # weight of sparsity penalty term
def sparseAutoencoderCostCallback(x):
return sparse_autoencoder.cost(x, visibleSize, hiddenSize, lambdaParam, sparsityParam,
betaParam, patches)
patches = getPatches(numPatches=10000, patchSize=patchSize)
thetaParam = sparse_autoencoder.initializeParameters(hiddenSize, visibleSize)
options = {
'maxiter': 400,
'disp': True,
}
result = optimize.minimize(sparseAutoencoderCostCallback, thetaParam, method='L-BFGS-B', jac=True, options=options)
W1 = result.x[0:hiddenSize*visibleSize].reshape(hiddenSize, visibleSize)
image_filename = 'images.png'
print 'Saving learned features to %s' % image_filename
visualizeNetwork(W1.T, image_filename)
if os.path.exists(sae1OptThetaFilename): sae1OptTheta = load(sae1OptThetaFilename) else: def sparseAutoencoderCostCallbackL1(x): return sparse_autoencoder.cost(x, inputSize, hiddenSizeL1, lambdaParam, sparsityParam, betaParam, trainData, corruptionLevel) sae1Theta = sparse_autoencoder.initializeParameters(hiddenSizeL1, inputSize) result = optimize.minimize(sparseAutoencoderCostCallbackL1, sae1Theta, method='L-BFGS-B', jac=True, options=options) sae1OptTheta = result.x save(sae1OptThetaFilename, sae1OptTheta) W1 = sae1OptTheta[0:hiddenSizeL1*inputSize].reshape(hiddenSizeL1, inputSize) visualizeNetwork(W1.T, results_dir + 'sae1.png') # Train the second sparse autoencoder sae1Features = sparse_autoencoder.feedForward(sae1OptTheta, hiddenSizeL1, inputSize, trainData) sae2OptThetaFilename = results_dir + 'sae2OptTheta.npy' if os.path.exists(sae2OptThetaFilename): sae2OptTheta = load(sae2OptThetaFilename) else: def sparseAutoencoderCostCallbackL2(x): return sparse_autoencoder.cost(x, hiddenSizeL1, hiddenSizeL2, lambdaParam, sparsityParam, betaParam, sae1Features, corruptionLevel) sae2Theta = sparse_autoencoder.initializeParameters(hiddenSizeL2, hiddenSizeL1) result = optimize.minimize(sparseAutoencoderCostCallbackL2, sae2Theta, method='L-BFGS-B', jac=True, options=options)