def dbnunfoldtonn(dbn, outputsize):
#DBNUNFOLDTONN Unfolds a DBN to a NN
# dbnunfoldtonn(dbn, outputsize ) returns the unfolded dbn with a final
# layer of size outputsize added.
if outputsize is not None:
size = np.append(dbn.dbnSizes, outputsize)
else:
size = np.array(dbn.sizes)
nn = NeuralNetwork.NN(size)
for i in range(0, len(dbn.RBM)):
nn.W[i][:, 0] = dbn.RBM[i].C.squeeze()
nn.W[i][:, 1:] = dbn.RBM[i].W
return nn
def test_example_NN():
dataSet = importMat.loadMatFile()
if dataSet is not None:
keys = dataSet.keys()
for key in keys:
if key == "train_x":
train_x = dataSet[key]
train_x = train_x / 255.0
elif key == "train_y":
train_y = dataSet[key]
train_y = train_y
elif key == "test_x":
test_x = dataSet[key]
test_x = test_x / 255.0
elif key == "test_y":
test_y = dataSet[key]
test_y = test_y
train_x, mu, sigma = zscore.zscore(train_x)
test_x = normalize.normalize(test_x, mu, sigma)
# ex1 vanilla nerual net
np.random.seed(
1
) # Setting the random seed so that the weights are generated same as in matlab code - default is twister algorithm
nn = NeuralNetwork.NN(np.array([784, 100, 10]))
# numepochs = Number of full sweep through data
# batchsize = Mean gradient step over this many samples
options = {"numepochs": 5, "batchsize": 100, "plot": 1}
nn, L = nntrain.nntrain(nn, train_x, train_y, options, None, None, 1)
er, bad = nntest.nntest(nn, test_x, test_y)
assert (er < 0.08), "Too big error"
del options
del nn
# ex2 neural net with L2 weight decay
np.random.seed(
1
) # Setting the random seed so that the weights are generated same as in matlab code - default is twister algorithm
nn = NeuralNetwork.NN(np.array([784, 100, 10]))
nn.WeightPenaltyL2 = 1e-4
options = {"numepochs": 1, "batchsize": 100, "plot": 1}
nn, L = nntrain.nntrain(nn, train_x, train_y, options, None, None, 2)
er, bad = nntest.nntest(nn, test_x, test_y)
assert (er < 0.08), "Too big error"
del nn
# ex3 nerual net with dropout
np.random.seed(
1
) # Setting the random seed so that the weights are generated same as in matlab code - default is twister algorithm
nn = NeuralNetwork.NN(np.array([784, 100, 10]))
nn.DropoutFraction = 0.5
nn, L = nntrain.nntrain(nn, train_x, train_y, options, None, None, 3)
er, bad = nntest.nntest(nn, test_x, test_y)
assert (er < 0.1), "Too big error"
del nn
# ex4 nerual net with sigmoid activation function
np.random.seed(
1
) # Setting the random seed so that the weights are generated same as in matlab code - default is twister algorithm
nn = NeuralNetwork.NN(np.array([784, 100, 10]))
nn.ActivationFunction = "sigm"
nn.LearningRate = 1
nn, L = nntrain.nntrain(nn, train_x, train_y, options, None, None, 4)
er, bad = nntest.nntest(nn, test_x, test_y)
assert (er < 0.1), "Too big error"
del nn
# ex5 nerual net with softmax activation function and plotting functionality
np.random.seed(
1
) # Setting the random seed so that the weights are generated same as in matlab code - default is twister algorithm
nn = NeuralNetwork.NN(np.array([784, 100, 10]))
nn.Output = "softmax" #Use the softmax output
options = {"numepochs": 1, "batchsize": 1000, "plot": 1}
nn, L = nntrain.nntrain(nn, train_x, train_y, options, None, None, 5)
er, bad = nntest.nntest(nn, test_x, test_y)
assert (er < 0.1), "Too big error"
del nn
# ex6 nerual net with sigmoid activation function and plotting functionality and training error
# Split training data into training and validation
vx = train_x[0:10000, :]
tx = train_x[9999:-1, :]
vy = train_y[0:10000, :]
ty = train_y[9999:-1, :]
np.random.seed(
1
) # Setting the random seed so that the weights are generated same as in matlab code - default is twister algorithm
nn = NeuralNetwork.NN(np.array([784, 100, 10]))
nn.Output = "softmax" #Use the softmax output
options = {"numepochs": 5, "batchsize": 1000, "plot": 1}
nn, L = nntrain.nntrain(nn, tx, ty, options, vx, vy, 6)
er, bad = nntest.nntest(nn, test_x, test_y)
assert (er < 0.1), "Too big error"
del nn
print("test_example_NN completed")
input("Press [enter] to continue.")