def doForwardPropagation(self, X, weights, biases):
'''
Computes the forward propagation of the input in the SMNN:
Z{l+1} = W{l}*H{l} + B{l}
H{l+1} = f(Z{l+1})
where {l} and {l+1} denote layers,
B is the bias matrix, columnwise repetition of the bias vector with the number of samples,
Z is the output matrix of neurons before the activation function is applied,
f(.) is the activation function
H is the output matrix of neurons after the activation function is applied (h{1}=X),
Arguments
X : data matrix in the form [input dim., number of samples]
weights : list of weight matrices of each layer
biases : list of bias vectors of each layer
Returns
outputs : list of output matrices (z) of each layer (output of neuron before activation function)
activities : list of activation matrices (h) of each layer (output of neuron after activation function)
'''
assert self.isInitialized, 'ERROR:SMNN:doForwardPropagation: The instance is not properly initialized'
# Default behaviour is bad implementation
#if len(weights)==0 or len(biases)==0:
# [weights, biases] = self.unrollParameters(self.params);
assert AuxFunctions.checkNetworkParameters(
weights, self.weightPrototypes
), 'ERROR:SMNN:doForwardPropagation: weight dimension does not match the network topology'
assert AuxFunctions.checkNetworkParameters(
biases, self.biasPrototypes
), 'ERROR:SMNN:doForwardPropagation: bias dimension does not match the network topology'
outputs = []
activities = []
for layer in range(self.nLayers - 1):
if layer == 0:
x = X
else:
x = activities[layer - 1]
z = np.dot(weights[layer], x) + np.repeat(biases[layer],
x.shape[1], 1)
if self.activation_fun == SMNN_ACTIVATION_FUNCTIONS[
SMNN_ACTFUN_SIGMOID]:
h = AuxFunctions.sigmoid(z)
else:
# Should not be here
print 'ERROR:SMNN:doForwardPropagation: Wrong activation function'
sys.exit()
outputs.append(z)
activities.append(h)
return [outputs, activities]
def convolve(filterDim, numFilters, X, W, b): ''' Returns the convolution of the features given by W and b with the given data X Arguments filterDim : filter (feature) dimension numFilters : number of feature maps X : input data in the form images(r, c, image number) W : weights i.e. features, is of shape (filterDim,filterDim,numFilters) b : biases, is of shape (numFilters,1) Returns convolvedFeatures : matrix of convolved features in the form convolvedFeatures(imageRow, imageCol, featureNum, imageNum) ''' inputDimX = X.shape[INDEX_X]; inputDimY = X.shape[INDEX_Y]; numData = X.shape[2]; convDimX = inputDimX - filterDim[INDEX_X] + 1; convDimY = inputDimY - filterDim[INDEX_Y] + 1; convolvedFeatures = np.zeros([convDimX, convDimY, numFilters, numData]); for i in range(numData): for filterNum in range (numFilters): # Convolution of image with feature matrix convolvedImage = np.zeros([convDimX, convDimY]); # Obtain the feature (filterDim x filterDim) needed during the convolution filter = W[:,:,filterNum]; # Flip the feature matrix because of the definition of convolution, as explained later filter = np.rot90(filter, 2); # Obtain data data = X[:,:,i]; #Convolve "filter" with "data", adding the result to convolvedImage convolvedImage = scipy.signal.convolve2d(data, filter, mode='valid'); # Add the bias unit # Then, apply the sigmoid function to get the hidden activation convolvedImage = AuxFunctions.sigmoid(convolvedImage + b[filterNum]); convolvedFeatures[:,:,filterNum,i] = convolvedImage; return convolvedFeatures