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 setParameters(self, W, b): ''' Sets the weights and biases of the layer with the given parameters Arguments W : weights to set b : biases to set ''' assert AuxFunctions.checkNetworkParameters([W], [self.weightPrototype]), 'ERROR:ConvLayer:setParameters: weight dimension does not match the network topology' ; assert AuxFunctions.checkNetworkParameters([b], [self.biasPrototype]), 'ERROR:ConvLayer:setParameters: bias dimension does not match the network topology'; self.weights = W; self.biases = b;
def rollParameters(self, weights, biases): ''' Converts the parameters in matrix form into vector weights : list of weight matrix of each layer biases : list of bias vector of each layer ''' assert AuxFunctions.checkNetworkParameters(weights, self.weightPrototypes), 'ERROR:CNN:rollParameters: weight dimension does not match the network topology' ; assert AuxFunctions.checkNetworkParameters(biases, self.biasPrototypes), 'ERROR:CNN:rollParameters: bias dimension does not match the network topology'; params = np.array([]); for i in range(len(weights)): params = np.hstack((params, weights[i].flatten(), biases[i].flatten())) return params
def doForwardPropagation(self, X, weights, biases): ''' Computes the forward propagation of the input in the CNN. 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 activations : list of activation matrices (h) of each layer (output of neuron after activation function) ''' assert self.isInitialized, 'ERROR:CNN:doForwardPropagation: The instance is not properly initialized' assert AuxFunctions.checkNetworkParameters(weights, self.weightPrototypes), 'ERROR:CNN:doForwardPropagation: weight dimension does not match the network topology' ; assert AuxFunctions.checkNetworkParameters(biases, self.biasPrototypes), 'ERROR:CNN:doForwardPropagation: bias dimension does not match the network topology'; activations = []; # Input to the network indata = X; # Propagate through the convolutional layers for i in range(len(self.layers)): # Compute the activity of the current layer outdata = self.layers[i].doForwardPropagation(indata, weights[i], biases[i]); # Save the activity of the current layer activations.append(outdata); # Set the activity of the current layer as the input to the next layer indata = outdata[INDEX_ACTIVATION_POOL]; # Compute the activity of the softmax (output) layer # Reshape input for the softmax layer indata = np.reshape(indata, [indata.shape[0]*indata.shape[1]*indata.shape[2], indata.shape[3]]); # Compute the activity #outdata = self.softmaxmodel.predict(indata); z = np.dot(weights[-1], indata) + np.repeat(biases[-1], X.shape[2], 1); h = np.exp(z); y = AuxFunctions.doUnbalancedMatrixOperation(h, np.sum(h, 0), 'div', axis=0 ); # Save the activity activations.append(y); return activations;
def unrollParameters(self, params): ''' Converts the vectorized parameters into matrix params: parameters to unroll ''' weights = []; biases = []; read_start = 0; read_end = 0; # Convolutional layers for i in range(len(self.layers)): # set the end index for read read_end = read_start + self.layers[i].filterDim[INDEX_X]*self.layers[i].filterDim[INDEX_Y]*self.layers[i].numFilters; # read the weights for the current layer w = params[read_start:read_end]; # reshape and the weights weights.append( np.reshape(w, (self.layers[i].filterDim[INDEX_X], self.layers[i].filterDim[INDEX_Y], self.layers[i].numFilters)) ); # set the start index for the next read read_start = read_end; # set the end index for the next read read_end = read_start + self.layers[i].numFilters; # read the bias terms b = params[read_start:read_end]; # reshape and store the bias biases.append( np.reshape(b, (self.layers[i].numFilters, 1)) ) # set the start index for the next read read_start = read_end; # Softmax layer read_end = read_start+np.size(self.weights) w = params[read_start:read_end]; weights.append( np.reshape(w, self.weights.shape) ); # set the start index for the next read read_start = read_end; # set the end index for the next read read_end = read_start + len(self.biases); b = params[read_start:read_end]; biases.append(np.reshape(b, self.biases.shape)) assert AuxFunctions.checkNetworkParameters(weights, self.weightPrototypes), 'ERROR:CNN:unrollParameters: dimensions of given parameters do not match the network topology' ; assert AuxFunctions.checkNetworkParameters(biases, self.biasPrototypes), 'ERROR:CNN:unrollParameters: dimensions of given parameters do not match the network topology'; return weights, biases;
def setNetworkParameters(self, weights, biases): ''' Returns the parameters of the network in a stacked form Arguments weights : list weights to set for each layer biases : list of biases to set for each layer ''' assert AuxFunctions.checkNetworkParameters(weights, self.weightPrototypes), 'ERROR:CNN:setNetworkParameters: weight dimension does not match the network topology' ; assert AuxFunctions.checkNetworkParameters(biases, self.biasPrototypes), 'ERROR:CNN:setNetworkParameters: bias dimension does not match the network topology'; for i in range(len(self.layers)): W = weights[i]; b = biases[i]; self.layers[i].setParameters(W, b); # Size check is done in the layer self.weights = weights[-1]; self.biases = biases[-1];
def doForwardPropagation(self, X, weights, biases): ''' Computes the forward propagation of the input in the network. 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 activities : list of activation matrices from convolution and pooling layers, respectively ''' assert self.isInitialized, 'ERROR:ConvLayer:doForwardPropagation: The instance is not properly initialized' assert AuxFunctions.checkNetworkParameters([weights], [self.weightPrototype]), 'ERROR:ConvLayer:doForwardPropagation: weight dimension does not match the network topology' ; assert AuxFunctions.checkNetworkParameters([biases], [self.biasPrototype]), 'ERROR:ConvLayer:doForwardPropagation: bias dimension does not match the network topology'; # Convolution activations_conv = convolve(self.filterDim, self.numFilters, X, weights, biases); # Pooling activations_pool = pool(self.poolDim, activations_conv, self.poolingFunction); return [activations_conv, activations_pool];
def setBiases(self, biases_new):
'''
Updates the biases of the model parameters of the network
Arguments
biases_new : New biases to set
'''
assert self.isInitialized, 'ERROR:SparseAutoencoder:setBiases: The instance is not properly initialized'
assert AuxFunctions.checkNetworkParameters(
biases_new, self.biasPrototypes
), 'ERROR:SparseAutoencoder:setBiases: bias dimension does not match the network topology'
[weights, biases] = self.unrollParameters(self.params)
biases = biases_new
self.params = self.rollParameters(weights, biases)
def rollParameters(self, weights, biases):
'''
Converts the parameters in matrix form into vector
Arguments
weights : list of weight matrices of each layer
biases : list of bias vectors of each layer
Returns
params : parameter vector
'''
assert AuxFunctions.checkNetworkParameters(
weights, self.weightPrototypes
), 'ERROR:SparseAutoencoder:rollParameters: weight dimension does not match the network topology'
assert AuxFunctions.checkNetworkParameters(
biases, self.biasPrototypes
), 'ERROR:SparseAutoencoder:rollParameters: bias dimension does not match the network topology'
params = np.array([])
for i in range(len(weights)):
params = np.hstack(
(params, weights[i].flatten(), biases[i].flatten()))
return params
def rollParameters(self, theta):
'''
Converts a given parameter matrix into a vector
Arguments
theta : parameter matrix
Returns
theta : parameter vector
'''
assert self.isInitialized, 'ERROR:SoftICA:unrollParameters: The instance is not properly initialized'
assert AuxFunctions.checkNetworkParameters(
[theta], [self.weightPrototype]
), 'ERROR:SoftICA:rollParameters: Weight dimension does not match the network topology'
return theta.flatten()
def unstackParameters(self, theta_list):
'''
Converts the model parameters from stacked form into vector form
Arguments
theta_list : list of model parameters of each layer
Returns
theta : vector of combined model parameters of the network
'''
assert self.isInitialized, 'ERROR:DeepNetwork:unstackParameters: The instance is not properly initialized'
assert AuxFunctions.checkNetworkParameters(
theta_list, self.modelParameterPrototype
), 'ERROR:DeepNetwork:unstackParameters: model parameter dimension does not match the network topology'
theta = np.array([])
for i in range(len(theta_list)):
theta = np.hstack((theta, theta_list[i].flatten()))
return theta
def stackParameters(self, theta):
'''
Converts the model parameters from vector form into stacked form
Arguments
theta : vector of combined model parameters of the network
Returns
theta_list : list of model parameters of each layer
'''
assert self.isInitialized, 'ERROR:DeepNetwork:stackParameters: The instance is not properly initialized'
theta_list = []
i_start = 0
for i in range(len(self.modelParameterPrototype)):
i_stop = i_start + self.modelParameterPrototype[i]
theta_list.append(theta[i_start:i_stop])
i_start = i_stop
assert AuxFunctions.checkNetworkParameters(
theta_list, self.modelParameterPrototype
), 'ERROR:DeepNetwork:stackParameters: model parameter dimension does not match the network topology'
return theta_list