def getImageData(numInputVectors):
from htmresearch.algorithms.image_sparse_net import ImageSparseNet
DATA_PATH = "../sparse_net/data/IMAGES.mat"
DATA_NAME = "IMAGES"
DEFAULT_SPARSENET_PARAMS = {
"filterDim": 64,
"outputDim": 64,
"batchSize": numInputVectors,
"numLcaIterations": 75,
"learningRate": 2.0,
"decayCycle": 100,
"learningRateDecay": 1.0,
"lcaLearningRate": 0.1,
"thresholdDecay": 0.95,
"minThreshold": 1.0,
"thresholdType": 'soft',
"verbosity": 0, # can be changed to print training loss
"showEvery": 500,
"seed": 42,
}
network = ImageSparseNet(**DEFAULT_SPARSENET_PARAMS)
print "Loading training data..."
images = network.loadMatlabImages(DATA_PATH, DATA_NAME)
nDim1, nDim2, numImages = images.shape
binaryImages = np.zeros(images.shape)
for i in range(numImages):
binaryImages[:, :, i] = convertToBinaryImage(images[:, :, i])
inputVectors = network._getDataBatch(binaryImages)
inputVectors = inputVectors.T
return inputVectors
def getImageData(numInputVectors):
from htmresearch.algorithms.image_sparse_net import ImageSparseNet
DATA_PATH = "../sparse_net/data/IMAGES.mat"
DATA_NAME = "IMAGES"
DEFAULT_SPARSENET_PARAMS = {
"filterDim": 64,
"outputDim": 64,
"batchSize": numInputVectors,
"numLcaIterations": 75,
"learningRate": 2.0,
"decayCycle": 100,
"learningRateDecay": 1.0,
"lcaLearningRate": 0.1,
"thresholdDecay": 0.95,
"minThreshold": 1.0,
"thresholdType": 'soft',
"verbosity": 0, # can be changed to print training loss
"showEvery": 500,
"seed": 42,
}
network = ImageSparseNet(**DEFAULT_SPARSENET_PARAMS)
print "Loading training data..."
images = network.loadMatlabImages(DATA_PATH, DATA_NAME)
nDim1, nDim2, numImages = images.shape
binaryImages = np.zeros(images.shape)
for i in range(numImages):
binaryImages[:, :, i] = convertToBinaryImage(images[:, :, i])
inputVectors = network._getDataBatch(binaryImages)
inputVectors = inputVectors.T
return inputVectors
def runExperiment():
print "Creating network..."
network = ImageSparseNet(**DEFAULT_SPARSENET_PARAMS)
print "Loading training data..."
images = network.loadMatlabImages(DATA_PATH, DATA_NAME)
print
print "Training {0}...".format(network)
network.train(images, numIterations=5000)
print "Saving loss history and function basis..."
network.plotLoss(filename=LOSS_HISTORY_PATH)
network.plotBasis(filename=BASIS_FUNCTIONS_PATH)
if capnp:
print
print "Saving model..."
proto1 = SparseNetProto.new_message()
network.write(proto1)
with open(SERIALIZATION_PATH, 'wb') as f:
proto1.write(f)
print "Loading model..."
with open(SERIALIZATION_PATH, 'rb') as f:
proto2 = SparseNetProto.read(f)
newNetwork = ImageSparseNet.read(proto2)
print "Checking that loaded model is the same as before..."
if newNetwork != network:
raise ValueError("Model is different!")
else:
print "Model is the same."
print
print "Training {0} again...".format(network)
newNetwork.train(images, numIterations=5000)
print "Saving loss history and function basis..."
newNetwork.plotLoss(filename=LOSS_HISTORY_PATH2)
newNetwork.plotBasis(filename=BASIS_FUNCTIONS_PATH2)
def runExperiment():
print "Creating network..."
network = ImageSparseNet(**DEFAULT_SPARSENET_PARAMS)
print "Loading training data..."
images = network.loadMatlabImages(DATA_PATH, DATA_NAME)
print
print "Training {0}...".format(network)
network.train(images, numIterations=5000)
print "Saving loss history and function basis..."
network.plotLoss(filename=LOSS_HISTORY_PATH)
network.plotBasis(filename=BASIS_FUNCTIONS_PATH)
if capnp:
print
print "Saving model..."
proto1 = SparseNetProto.new_message()
network.write(proto1)
with open(SERIALIZATION_PATH, 'wb') as f:
proto1.write(f)
print "Loading model..."
with open(SERIALIZATION_PATH, 'rb') as f:
proto2 = SparseNetProto.read(f)
newNetwork = ImageSparseNet.read(proto2)
print "Checking that loaded model is the same as before..."
if newNetwork != network:
raise ValueError("Model is different!")
else:
print "Model is the same."
print
print "Training {0} again...".format(network)
newNetwork.train(images, numIterations=5000)
print "Saving loss history and function basis..."
newNetwork.plotLoss(filename=LOSS_HISTORY_PATH2)
newNetwork.plotBasis(filename=BASIS_FUNCTIONS_PATH2)
