def findBestRegularization(s, x_sub, y_sub):
regs = np.linspace(0, 10, 20)
reg_acc_cv = []
reg_acc_train = []
max_acc = 0
best_reg = 0
for r in regs:
th1, th2 = Train.trainSciPy2(s, x_sub, y_sub, r)
acc_cv = accuracy_score(y_cv, [SimpleNN2.predictClass(s, th1, th2, w) for w in x_cv])
acc_train = accuracy_score(y_sub, [SimpleNN2.predictClass(s, th1, th2, w) for w in x_sub])
reg_acc_cv.append(acc_cv)
reg_acc_train.append(acc_train)
if max_acc < acc_cv:
max_acc = acc_cv
best_reg = r
print("Validating regularization parameter [{0}]; Train accuracy: [{1}] CV accuracy: [{2}]"
.format(r, acc_train, acc_cv))
print("Best reg param: {0} with accuracy on CV dataset: {1}".format(best_reg, max_acc))
plt.plot(regs, reg_acc_cv);plt.plot(regs, reg_acc_train)
plt.show()
return best_reg
def test3():
(x, y) = DataModel.loadData("..\\train.csv")
(x_train, x_cv, y_train, y_cv) = DataModel.splitData(x, y)
x_sub = x_train[:20000,:]
y_sub = y_train[:20000]
s = SimpleNN2.NeuralNetConfig(784, 70, 10)
regLambda = 6.84
#s = Train.trainGradientDescent(s, x_sub, y_sub, 5)
th1, th2 = Train.trainSciPy2(s, x_sub, y_sub, regLambda)
#th1, th2 = Train.trainGradientDescent2(s, x_sub, y_sub, 5)
acc_cv = accuracy_score(y_cv, [SimpleNN2.predictClass(s, th1, th2, w) for w in x_cv])
print("Accuracy on CV set: {0}".format(acc_cv))
def trainFullAndSave():
(x, y) = DataModel.loadData("..\\train.csv")
(x_train, x_cv, y_train, y_cv) = DataModel.splitData(x, y)
s = SimpleNN2.NeuralNetConfig(784, 70, 10)
regLambda = 6.84
print("Training neural network on full dataset")
#s = Train.trainGradientDescent(s, x_sub, y_sub, 5)
th1, th2 = Train.trainSciPy2(s, x_train, y_train, regLambda)
#th1, th2 = Train.trainGradientDescent2(s, x_sub, y_sub, 5)
print("Training complete, checking accuracy on CV data")
acc_cv = accuracy_score(y_cv, [SimpleNN2.predictClass(s, th1, th2, w) for w in x_cv])
print("Accuracy on CV set: {0}".format(acc_cv))
SimpleNN2.saveNetwork(s, th1, th2, "..\\NeuralNetwork.bin")
