inputSize = 4
hiddenSize = 5
lambdaParam = 0.01
data = random.normal(size=(5, inputSize))
labels = array([0, 1, 0, 1, 0])
numClasses = 2
stack = [Layer(1), Layer(2)]
stack[0].W = 0.1 * random.normal(size=(3, inputSize))
stack[0].b = zeros(3)
stack[1].W = 0.1 * random.normal(size=(hiddenSize, 3))
stack[1].b = zeros(hiddenSize)
softmaxTheta = 0.005 * random.normal(size=hiddenSize * numClasses)
(stackParams, netConfig) = stack2params(stack)
stackedAETheta = concatenate([softmaxTheta, stackParams])
def stackedAutoencoderCostCallback(x):
return cost(x, inputSize, hiddenSize, numClasses, netConfig,
lambdaParam, data, labels)
(cost_value, grad) = stackedAutoencoderCostCallback(stackedAETheta)
numgrad = computeNumericalGradient(stackedAutoencoderCostCallback, stackedAETheta)
diff = linalg.norm(numgrad-grad)/linalg.norm(numgrad+grad)
print('%s' % diff)
print('Norm of the difference between numerical and analytical gradient (should be < 1e-9)\n\n')
return argmax(h_data, axis=0)
if __name__ == "__main__":
""" Check correctness of implemenation of softmax cost function
using gradient check
"""
numClasses = 10 # Number of classes (MNIST images fall into 10 classes)
lambdaParam = 1e-4 # Weight decay parameter
inputSize = 8
inputData = random.normal(size=(100, inputSize))
labels = random.randint(10, size=100)
def softmaxCostCallback(x):
return cost(x, numClasses, inputSize, lambdaParam, inputData, labels)
# Randomly initialise theta
thetaParam = 0.005 * random.normal(size=numClasses * inputSize)
(cost_value, grad) = softmaxCostCallback(thetaParam)
numGrad = computeNumericalGradient(softmaxCostCallback, thetaParam)
# Compare numerically computed gradients with those computed analytically
diff = linalg.norm(numGrad - grad) / linalg.norm(numGrad + grad)
print('%s' % diff)
print(
'Norm of the difference between numerical and analytical gradient (should be < 1e-7)\n\n'
)
h_data = exp(thetaParam.dot(data.T))
h_data = h_data / sum(h_data, 0)
return argmax(h_data, axis=0)
if __name__ == "__main__":
""" Check correctness of implemenation of softmax cost function
using gradient check
"""
numClasses = 10 # Number of classes (MNIST images fall into 10 classes)
lambdaParam = 1e-4 # Weight decay parameter
inputSize = 8
inputData = random.normal(size=(100,inputSize))
labels = random.randint(10, size=100)
def softmaxCostCallback(x):
return cost(x, numClasses, inputSize, lambdaParam, inputData, labels)
# Randomly initialise theta
thetaParam = 0.005 * random.normal(size=numClasses * inputSize)
(cost_value, grad) = softmaxCostCallback(thetaParam)
numGrad = computeNumericalGradient(softmaxCostCallback, thetaParam)
# Compare numerically computed gradients with those computed analytically
diff = linalg.norm(numGrad-grad)/linalg.norm(numGrad+grad)
print('%s' % diff)
print('Norm of the difference between numerical and analytical gradient (should be < 1e-7)\n\n')
return theta
if __name__ == "__main__":
""" Check correctness of implemenation of sparse_autoencoder cost function
using gradient check
"""
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
patches = getPatches(numPatches=10, patchSize=patchSize)
# Obtain random parameters theta
thetaParam = initializeParameters(hiddenSize, visibleSize)
def sparseAutoencoderCostCallback(x):
return cost(x, visibleSize, hiddenSize, lambdaParam, sparsityParam, betaParam, patches)
(cost_value, grad) = sparseAutoencoderCostCallback(thetaParam)
numgrad = computeNumericalGradient(sparseAutoencoderCostCallback, thetaParam)
diff = linalg.norm(numgrad - grad) / linalg.norm(numgrad + grad)
print("%s" % diff)
print("Norm of the difference between numerical and analytical gradient (should be < 1e-9)\n\n")
if __name__ == "__main__":
""" Check correctness of implemenation of sparse_autoencoder cost function
using gradient check
"""
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
patches = getPatches(numPatches=10, patchSize=patchSize)
# Obtain random parameters theta
thetaParam = initializeParameters(hiddenSize, visibleSize)
def sparseAutoencoderCostCallback(x):
return cost(x, visibleSize, hiddenSize, lambdaParam, sparsityParam,
betaParam, patches)
(cost_value, grad) = sparseAutoencoderCostCallback(thetaParam)
numgrad = computeNumericalGradient(sparseAutoencoderCostCallback,
thetaParam)
diff = linalg.norm(numgrad - grad) / linalg.norm(numgrad + grad)
print('%s' % diff)
print(
'Norm of the difference between numerical and analytical gradient (should be < 1e-9)\n\n'
)
