pull/1/head
E. Almqvist 4 years ago
parent 2a597c1a34
commit 3c9bffd797
  1. 22
      rgbAI/lib/func.py

@ -53,31 +53,34 @@ class AIlib:
dCost = cost2 - cost1 dCost = cost2 - cost1
return dCost return dCost
def compareInstance( obj, theta, layerIndex, neuronIndex=0 ): def compareInstance( obj, theta, layerIndex, neuronIndex_X=0, neuronIndex_Y=0 ):
# Create new instances of the object # Create new instances of the object
obj2_w = copy(obj) # annoying way to create a new instance of the object obj2_w = copy(obj) # annoying way to create a new instance of the object
obj2_b = copy(obj) obj2_b = copy(obj)
obj2_w.weights[layerIndex][neuronIndex] += theta # mutate the second objects neuron obj2_w.weights[layerIndex][neuronIndex_X][neuronIndex_Y] += theta # mutate the second objects neuron
dCost_weight = AIlib.compareAIobjects( obj, obj2_w ) # compare the two and get the dCost with respect to the weights dCost_weight = AIlib.compareAIobjects( obj, obj2_w ) # compare the two and get the dCost with respect to the weights
obj2_b.bias[layerIndex][neuronIndex] += theta # do the same thing for the bias obj2_b.bias[layerIndex][neuronIndex_X][neuronIndex_Y] += theta # do the same thing for the bias
dCost_bias = AIlib.compareAIobjects( obj, obj2_b ) dCost_bias = AIlib.compareAIobjects( obj, obj2_b )
return dCost_weight, dCost_bias return dCost_weight, dCost_bias
def getChangeInCost( obj, theta, layerIndex ):
mirrorObj = copy(obj)
# Fill the buffer with None so that the dCost can replace it later
mirrorObj.weights[layerIndex].fill(None)
mirrorObj.bias[layerIndex].fill(None)
def gradient( inp:np.array, obj, theta:float, maxLayer:int, layerIndex: int=0, grads=None, obj1=None, obj2=None ): # Calculate the gradient for that prop def gradient( inp:np.array, obj, theta:float, maxLayer:int, layerIndex: int=0, grads=None, obj1=None, obj2=None ): # Calculate the gradient for that prop
# Check if grads exists, if not create the buffer # Check if grads exists, if not create the buffer
if( not grads ): if( not grads ):
grads = [None] * (maxLayer+1) grads = [None] * (maxLayer+1)
# Create the change in variable (which is constant to theta)
dWeight = np.zeros(shape=obj.weights[layerIndex].shape).fill(theta) # since (x + theta) - (x) = theta then just fill it with theta
dBias = np.zeros(shape=obj.bias[layerIndex].shape).fill(theta)
# Calculate the gradient for the layer # Calculate the gradient for the layer
weightDer = AIlib.propDer( dCost, dWeight ) weightDer = AIlib.propDer( dCost_W, theta )
biasDer = AIlib.propDer( dCost, dBias ) biasDer = AIlib.propDer( dCost_B, theta )
# Append the gradients to the list # Append the gradients to the list
grads[layerIndex] = { grads[layerIndex] = {
@ -112,6 +115,7 @@ class AIlib:
while( not curCost or curCost > targetCost ): # targetCost is the target for the cost function while( not curCost or curCost > targetCost ): # targetCost is the target for the cost function
maxLen = len(obj.bias) maxLen = len(obj.bias)
grads, res, curCost = AIlib.gradient( inp, obj, theta, maxLen - 1 ) grads, res, curCost = AIlib.gradient( inp, obj, theta, maxLen - 1 )
obj = AIlib.mutateProps( obj, maxLen, grads ) # mutate the props for next round obj = AIlib.mutateProps( obj, maxLen, grads ) # mutate the props for next round
print("Cost:", curCost, "|", inp, res) print("Cost:", curCost, "|", inp, res)

Loading…
Cancel
Save