SparseAutoencoder.ACTIVATION_FUNCTION_SIGMOID,
SparseAutoencoder.ACTIVATION_FUNCTION_SIGMOID
]
params_ae['debug'] = 1
# Parameters for softmax
params_softmax = {}
params_softmax['id'] = 'softmax'
params_softmax['outputDim'] = outputDim
params_softmax['featureDim'] = 200
params_softmax['debug'] = 1
layerParams = [params_ae, params_ae, params_softmax]
# Read data from file
labels_training = DataInputOutput.loadMNISTLabels(
mnist_lbl_filename_training, nSamples_max_train)
images_training = DataInputOutput.loadMNISTImages(
mnist_img_filename_training, nSamples_max_train)
labels_test = DataInputOutput.loadMNISTLabels(mnist_lbl_filename_test,
nSamples_max_test)
images_test = DataInputOutput.loadMNISTImages(mnist_img_filename_test,
nSamples_max_test)
# Normalize data
images_training = images_training / 255.0
images_test = images_test / 255.0
DNN = DeepNetwork(inputDim,
layerParams,
lambd=lambd,
doFineTuning=doFineTuning,
return pred
if __name__ == '__main__':
# --------------------------
# Example:
# Housing price prediction
# --------------------------
housing_filename = '/home/cem/develop/UFL/data/housing.bin'
debug = 1
doPlot = 1
# Read data from file
data = DataInputOutput.loadHousingData(housing_filename)
# Include a row of 1s as an additional intercept feature.
data = np.vstack((np.ones((1, data.shape[1])), data))
# Shuffle examples.
data = data[:, np.random.permutation(data.shape[1])]
# Split into train and test sets
# The last row of 'data' is the median home price.
data_train_X = data[0:-1, 0:400]
data_train_y = data[-1:-2:-1, 0:400]
data_test_X = data[0:-1, 401:-1]
data_test_y = data[-1:-2:-1, 401:-1]
imChannels = 1
inputDim = patchWidth * patchHeight * imChannels
numFeatures = 25
lambda_w = 0.0001
# weight decay parameter
beta = 3
# weight of sparsity penalty term
sparsityParam = 0.01
actFunctions = [
ACTIVATION_FUNCTION_SIGMOID, ACTIVATION_FUNCTION_SIGMOID
]
# Read data from file
data = scipy.io.loadmat(filename_data)
patches = DataInputOutput.samplePatches(data['IMAGES'], patchWidth,
patchHeight, numPatches)
# Normalize data
patches = DataNormalization.normZeroToOne(patches)
if debug > 1:
Visualization.displayNetwork(patches[:, 0:100])
if debug:
print 'Number of samples: ', patches.shape[1]
dimLayers = [inputDim, numFeatures, inputDim]
SAE = SparseAutoencoder(dimLayers=dimLayers,
lambda_w=lambda_w,
beta=beta,
# Test gradient computation?
debug = 1
numImages = 10000
imWidth = 28
imHeight = 28
numPatches = 10000
imageChannels = 1
patchWidth = 9
patchHeight = 9
inputDim = patchWidth * patchHeight * imageChannels
numFeatures = 50
epsilon = 1e-2
lambd = 0.99
# Read data from file
images_training = DataInputOutput.loadMNISTImages(
mnist_img_filename_training, numImages)
images_training = np.reshape(images_training,
[imHeight, imWidth, images_training.shape[1]])
# Sample patches
patches = DataInputOutput.samplePatches(images_training, patchWidth,
patchHeight, numPatches)
if debug > 1:
Visualization.displayNetwork(patches[:, 0:100])
# Normalize data: ZCA whiten patches
patches = patches / 255.0
instance_pca = PCA.PCA(inputDim, 0.99, debug)
patches_ZCAwhite = instance_pca.doZCAWhitening(patches)
inputDim_img = imWidth * imHeight * imageChannels
numFeatures = 32
nClasses = 10
epsilon = 1e-2
lambd = 0.99
poolDim = 5
#-------------------------
# Load Data
#-------------------------
if debug: print "Loading data..."
# Read data from file
numImages = numImages_unlabeled + numImages_training + numImages_test
images = DataInputOutput.loadMNISTImages(mnist_img_filename_training,
numImages)
images = np.reshape(images, [imHeight, imWidth, images.shape[1]])
images_unlabeled = images[:, :, 0:numImages_unlabeled]
images_training = images[:, :, numImages_unlabeled:numImages_unlabeled +
numImages_training]
images_test = images[:, :, numImages_unlabeled +
numImages_training:numImages_unlabeled +
numImages_training + numImages_test]
labels = DataInputOutput.loadMNISTLabels(mnist_lbl_filename_training,
numImages)
labels_training = labels[numImages_unlabeled:numImages_unlabeled +
numImages_training]
labels_test = labels[numImages_unlabeled +
numImages_training:numImages_unlabeled +
numImages_training + numImages_test]
patchWidth = 28;
patchHeight = 28;
outputDim = 10;
params_layer1 = {}
params_layer1['numFilters'] = 2;
params_layer1['filterDim'] = [2, 2];
params_layer1['poolDim'] = [3, 3];
params_layer1['poolingFunction'] = POOLING_MEAN;
params_layer1['debug'] = 1;
params_layers = [params_layer1];
#testlabel = DataInputOutput.loadMNISTLabels(mnist_lbl_filename_training, numPatches);
testlabel = np.array([1,2,3,4,3,2,1,2,3,4])-1;
testdata = DataInputOutput.loadMNISTImages(mnist_img_filename_training, numPatches);
testdata = testdata / 255.0;
testdata = np.reshape(testdata, [patchWidth, patchHeight, testdata.shape[1]]);
ConvNet_test = CNN( [patchWidth, patchHeight], outputDim, params_layers, debug=debug);
print 'Checking gradient...'
ConvNet_test.testGradient(testdata, testlabel);
debug = 1;
nSamples_max_train = 10000;
nSamples_max_test = 10000;
imWidth = 28;
imHeight = 28;
outputDim = 10;