def FullyConnectedLayerTest():
inputLayer = InputLayer(numInput + 1) # The +1 is the bias unit.
hiddenLayer = FullyConnectedLayer(numInput + 1, numHidden, nonlinear, nonlinearDeriv)
outputLayer = FullyConnectedLayer(numHidden, numOutput, nonlinear, nonlinearDeriv)
# Until we're done,
for i in xrange(numIterations):
# go through our samples:
for sample in samplesToProc:
inputs = sample[0:2]
# Do forward propagation on the sample.
inputOut = inputLayer.forward_prop(inputs)
hiddenOut = hiddenLayer.forward_prop(inputOut)
outOut = outputLayer.forward_prop(hiddenOut)
# Now, do backward propagation.
# Note that computation of weight deltas requires information outside of a layer.
outWeights = outputLayer.W
outDelta = sample[2] - outOut
outBack = outputLayer.backward_prop(hiddenOut, outDelta, learningRate)
hiddenDelta = np.dot(outWeights, outBack) # Needs weights of output layer BEFORE update!
hiddenBack = hiddenLayer.backward_prop(inputOut, hiddenDelta, learningRate)
# Print the current error we're experiencing.
if i % 100 == 0:
print "error on iteration %d: %0.3f" % (i, Error(sample[2], outOut)[0])
# Print our final values.
print "FINAL VALUES:"
for sample in samplesToProc:
inputs = sample[0:2]
print "input: ", inputs
# Get the final value.
inputOut = inputLayer.forward_prop(inputs)
hiddenOut = hiddenLayer.forward_prop(inputOut)
outOut = outputLayer.forward_prop(hiddenOut)
# Print final along with expected, and error.
print "output: ", outOut
print "expected output: ", sample[2]
print "error: %0.2f" % Error(sample[2], outOut)[0]
