# in layer 2 grab the 5th column
r, c = out[i].shape
stats = np.zeros((noOfStatsTests, c))
for j in range(c):
r = out[i][:, j].shape[0]
y = out[i][:, j]
# stats[0, j] = a.local_charactor(y)
# stats[1, j] = a.global_charactor(y)
# ynorm = a.normalise_to_zero_one_interval(y, layerMinMaxVal[i][0], layerMinMaxVal[i][1])
# print(stats)
doPredictions = 1
## This sections figures out how good the model is by getting it to predict the answers for the train
## and test sets
if doPredictions == 1:
d.prediction_tester(model, XTrain, TTrain, name='Training data')
if noOfTestData != 0:
d.prediction_tester(model,
XTest,
TTest,
name='Test data',
example_no=0)
from keras.utils.visualize_util import plot
plot(model, to_file='model.png')
d.write_out(model, "model.json")
doGap = 0
if doGap == 1:
verbose = 1
def GeneralisationTest(noOfTestData=500, doPredictions=1, doMatLabResults=False):
"""Function to create a disjoing from the training set test set"""
X= np.load("allInputDataCurrent.npy")
T= np.load("allOutputDataCurrent.npy")
from keras.models import load_model
model = load_model("Random_model.h5")
# things we can calc from this:
noOfTrainData = len(X)
assert len(X) == len(T)
lenOfInput = len(X[3])
lenOfOutput = len(T[3])
lenOfBlock = int(lenOfInput / noOfPrototypes)
noOfExamples = noOfTrainData //noOfPrototypes
noOfNewExamples = noOfTestData // noOfPrototypes
lenOfR = lenOfInput - lenOfBlock
weightOfX = int(sum(X[0]))
weightOfR = weightOfX - lenOfBlock
inverseWeightOfR = lenOfR - weightOfR
denom=lenOfInput-(lenOfInput/noOfPrototypes) # denom is the floating point length of R
assert int(denom) == lenOfR
fractionalWeightOfR = weightOfR / denom
fractionalInverseWeightOfR = inverseWeightOfR / denom
weight = [fractionalWeightOfR, fractionalInverseWeightOfR]
weightOfT = int(sum(T[3]))
if lenOfOutput == noOfPrototypes:
use1HOT = 1
else:
use1HOT = 0
if categories == True:
noOfOutputs = noOfPrototypes
if use1HOT == 1:
sizeOfOutput = noOfPrototypes
print('Overwriting output vector size to length {}'.format(noOfPrototypes))
else:
noOfOutputs = noOfTrainData
print('Random vector, R, has weight {0}'.format(weightOfR))
#Test_X = code.make_prototyped_random_codes(M=noOfTestData, n=lenOfInput, p=noOfPrototypes, weight=[fractionalWeightOfR],
# k=2, symbolList=None, verbose=verbose, decay_templates=decay)
#### testing code
#this gives you matlab files of the codes so you can play with them if you want
if doMatLabResults:
Test_X = code.make_prototyped_random_codes(M=500, n=lenOfInput, p=noOfPrototypes, weight=[fractionalWeightOfR],
k=2, symbolList=None, verbose=verbose, decay_templates=decay)
sio.savemat('Test_X5000.mat', {'Test_X':Test_X})
R = code.make_random_codes(M=500, n=501, weight=weight, k=2,symbolList=[1,0], verbose=True)
sio.savemat('R3.mat', {'R':R})
#######
Test_X, All_X = code.get_test_x(X=X, noOfTestData=noOfTestData, lenOfInput=lenOfInput, noOfPrototypes=noOfPrototypes,
weight=[fractionalWeightOfR, fractionalInverseWeightOfR], k=2, symbolList=None, verbose=verbose, decay_templates=decay)
###### get T
######
## Now we get the correct sized Test_T
Test_T, prototypeOutputCodes = code.get_test_t(T,
noOfPrototypes=noOfPrototypes,
noOfTestData=noOfTestData,
lenOfOutput=len(T[0]),
verbose=False)
## This sections figures out how good the model is by getting it to predict the answers for the train
## and test sets
if doPredictions == 1:
d.prediction_tester(model, X, T, name='Training data')
if noOfTestData != 0:
d.prediction_tester(model, Test_X, Test_T, name='Test data', example_no=0)
np.save("GeneralisantionInputDataTest.npy", Test_X)
np.save("GeneralisationOutputDataTest.npy", Test_T)
return Test_X, Test_T