class Test(unittest.TestCase):
trainData = np.random.rand(100, 784)
trainLabel = np.random.randint(10, size=(100,), dtype=np.uint8)
train = [trainData, trainLabel]
testData = np.random.rand(50, 784)
testLabel = np.random.randint(10, size=(50,), dtype=np.uint8)
test = [testData, testLabel]
nn = NN.NN(training=train, testing=test, lr=0.003, mu=.99, minibatch=100, dropout=1, disableLog=True)
def test_num_hidden_layers(self):
self.nn.addLayers([10, 10], ['relu', 'relu'])
self.assertEqual(self.nn.nHidden, 2)
def test_relu(self):
z = np.zeros((5, 5))
r = af.ReLU(z)
self.assertTrue(r.min() == r.max())
def test_sigmoid_zero(self):
self.assertEqual(af.sigmoid(0), 0.5)
def test_sigmoid_max(self):
self.assertAlmostEqual(af.sigmoid(1000), 1)
def test_sigmoid_min(self):
self.assertAlmostEqual(af.sigmoid(-1000), 0)
def mod_sel():
n1 = 20
n2 = 40
while n2 <= 300:
while n1 <= 300:
i = 0
acc = []
for train_index, test_index in kf.split(x):
train_set = [
TRAINING[0][train_index], TRAINING[1][train_index]
]
val_set = [TRAINING[0][test_index], TRAINING[1][test_index]]
nn = NN.NN(training=train_set,
testing=val_set,
lr=0.003,
mu=.99,
minibatch=100)
nn.addLayers([n1, n2], ['relu', 'relu'])
_, acc_val = nn.train(stop_function=0, num_epochs=120)
i += 1
acc.append(acc_val)
if i == end:
print(np.mean(acc))
m = [str([n1, n2]), np.mean(acc)]
print('Mean accuracy on validation: ' + str(m))
means.append(m)
n1 += 20
n2 += 20
n1 = 20
return means
def mod_sel(n1, n2):
kf = KFold(n_splits=3, random_state=42, shuffle=True)
x = TRAINING[0]
means = []
end = kf.get_n_splits(x)
i = 0
acc = []
for train_index, test_index in kf.split(x):
train_set = [TRAINING[0][train_index], TRAINING[1][train_index]]
val_set = [TRAINING[0][test_index], TRAINING[1][test_index]]
nn = NN.NN(training=train_set,
testing=val_set,
lr=0.003,
mu=.99,
minibatch=100)
nn.addLayers([n1, n2], ['relu', 'relu'])
_, acc_val = nn.train(stop_function=0, num_epochs=120)
i += 1
acc.append(acc_val)
if i == end:
m = [str([n1, n2]), round(np.mean(acc), 2)]
print('Mean accuracy on validation: ' + str(m))
means.append(m)
return means
for n in [[250,100]]:
nn = NN.NN(training=TRAINING, testing=TESTING, lr=0.003, mu=.99, minibatch=100)
NN.NN.update_layers = normal_upd
nn.addLayers(n, ['relu', 'relu','tanh'])
a,b=nn.train(stop_function=0, num_epochs=150)
w = (nn.getWeigth())
for p in [10,20]:
print("Pruning="+str(p)+"%")
w1=np.copy(w)
pr.set_pruned_layers(nn, p, w1)
nn.train(stop_function=0, num_epochs=50)'''
for n in [[250, 100]]:
nn = NN.NN(training=TRAINING,
testing=TESTING,
lr=0.003,
mu=.99,
minibatch=100)
NN.NN.update_layers = normal_upd
nn.addLayers(n, ['leakyrelu', 'leakyrelu', 'sigmoid'])
a, b = nn.train(stop_function=0, num_epochs=150)
w = (nn.getWeigth())
for p in [10, 20]:
print("Pruning=" + str(p) + "%")
w1 = np.copy(w)
pr.set_pruned_layers(nn, p, w1)
nn.train(stop_function=0, num_epochs=50)
for n in [[250, 100]]:
nn = NN.NN(training=TRAINING,
testing=TESTING,
labels = np.linspace(1, len(bin_data), num=len(bin_data), dtype=np.float64)
labels = labels / len(bin_data)
labels = np.reshape(labels, (-1, 1))
p = np.random.RandomState(seed=42).permutation(dim_set)
bin_data_perm = bin_data[p]
labels_perm = labels[p]
now = time.time()
nn = NN.NN(training=[bin_data_perm, labels_perm],
testing=[[0], [0]],
lr=0.005,
mu=0.9,
lambd=1e-6,
minibatch=64,
disableLog=True)
#file3, lr=0.005 --> 4 48s
#file7, lr=0.005 --> 40 28s
nn.addLayers([256], ['leakyrelu', 'leakyrelu'])
nn.train(stop_function=2, num_epochs=20000)
later = time.time()
difference = int(later - now)
#print("R2: {}".format(r2_score(labels, nn.predict(bin_data))))
max_err = 0
mean_err = 0
for j in range(dim_set):
dim_set = len(bin_data)
labels = np.linspace(1, len(bin_data), num=len(bin_data), dtype=np.float64)
labels = labels / len(bin_data)
labels = np.reshape(labels, (-1, 1))
p = np.random.RandomState(seed=42).permutation(dim_set)
bin_data_perm = bin_data[p]
labels_perm = labels[p]
lr = 0.003 if i > 2 else 0.0005
nn = NN.NN(training=[bin_data_perm, labels_perm],
testing=[[0], [0]],
lr=lr,
mu=0.9,
lambd=1.e-5,
minibatch=64,
disableLog=True)
nn.addLayers([256], ['leakyrelu', 'tanh'])
nn.train(stop_function=2, num_epochs=20000)
print("R2: {}".format(r2_score(labels, nn.predict(bin_data))))
max_err = 0
mean_err = 0
for j in range(dim_set):
pr = floor(nn.predict(bin_data[j])[0] * dim_set)
val = abs(pr - labels[j] * dim_set)
if val > max_err:
max_err = val
mean_err += val