def test_forward_1(self):
size = 4
inp = np.random.normal(size=(size, 1))
network = NNNetwork(
layers=(FullyConnectedLayer(inp_size=size, out_size=6),
FullyConnectedLayer(inp_size=6, out_size=3)))
activation = network.forward(inp)
print(activation)
self.assertEqual((3, 1), activation.shape)
def test_backprop(self):
layer = FullyConnectedLayer(10, 5)
inp_array = np.random.normal(size=(10, 1))
delta_array = np.random.normal(size=(5, 1))
gradient_w_array, gradient_b_array, back_delta_array = layer.backprop(
inp_array, delta_array)
self.assertEqual((10, 1), back_delta_array.shape)
self.assertEqual((5, 1), gradient_b_array.shape)
self.assertEqual((5, 10), gradient_w_array.shape)
def test_backprop_1(self):
inp_size = 4
out_size = 3
inp = np.random.normal(size=(inp_size, 1))
out = np.random.normal(size=(out_size, 1))
network = NNNetwork(
layers=(FullyConnectedLayer(inp_size=inp_size, out_size=6),
FullyConnectedLayer(inp_size=6, out_size=out_size)))
gradient_w_arrays, gradient_b_arrays = network.backprop(inp, out)
network.update(gradient_w_arrays, gradient_b_arrays)
def test_backprop_3(self):
network = NNNetwork(
layers=(FullyConnectedLayer(inp_size=5, out_size=100),
ConvoLayer()))
inp_size = 5
out_size = 4
inp_array = np.random.normal(size=(inp_size, 1))
out_array = np.random.normal(size=(out_size, out_size))
network.backprop(inp_array, out_array)
def test_forward_3(self):
network = NNNetwork(
layers=(FullyConnectedLayer(inp_size=5, out_size=100),
ConvoLayer()))
size = 5
inp = np.random.normal(size=(size, 1))
activation = network.forward(inp)
print('activation', activation)
print('activation shape', activation.shape)
self.assertEqual((4, 4), activation.shape)
def test_forward_2(self):
size = 10
inp = np.random.normal(size=(size, size))
network = NNNetwork(
layers=(ConvoLayer(),
FullyConnectedLayer(inp_size=16, out_size=3)))
activation = network.forward(inp)
print('activation', activation)
print('activation shape', activation.shape)
self.assertEqual((3, 1), activation.shape)
def test_backprop_2(self):
inp_size = 10
out_size = 3
inputs_array = [
np.random.normal(size=(inp_size, inp_size)) for i in range(100)
]
outputs_array = [
np.random.normal(size=(out_size, 1)) for i in range(100)
]
network = NNNetwork(
layers=(ConvoLayer(),
FullyConnectedLayer(inp_size=16, out_size=out_size)))
for epoch in range(200):
print 'epoc {}'.format(epoch)
for inp, out in zip(inputs_array, outputs_array):
gradient_w_arrays, gradient_b_arrays = network.backprop(
inp, out)
for i, layer in enumerate(network.layers):
network.layers[i].update(gradient_w_arrays[i],
gradient_b_arrays[i], 0.0015)
if i == 1:
# print('network layer {}'.format(i))
# print network.layers[i].weight_array
# print network.layers[i].bias_array
pass
for i, layer in enumerate(network.layers):
if i == 1:
# print('network layer {}'.format(i))
# print 'weight array', network.layers[i].weight_array
# print 'bias array', network.layers[i].bias_array
pass
all_cost = np.zeros((3, 1))
for inp, out in zip(inputs_array, outputs_array):
# forward_result = network.forward(inp)
# print 'network forward result', forward_result
cost = (network.forward(inp) - out)**2
all_cost += cost
print 'all cost'
print all_cost
def test_backprop_with_epoch(self):
input_size = 100
layer = FullyConnectedLayer(10, 5)
vector = np.random.normal(size=(5, 10))
inp_array = [np.random.normal(size=(10, 1)) for i in range(input_size)]
out_array = [np.dot(vector, i) for i in inp_array]
for epoch in range(10000):
activation_array = [layer.forward(inp) for inp in inp_array]
weighted_array = np.asarray([i[1] for i in activation_array])
activations = np.asarray([i[0] for i in activation_array])
cost_array = [x - y for x, y in zip(activations, out_array)]
cost = sum([a.transpose().dot(a)
for a in cost_array]) / (2 * input_size)
delta_array = CrossEntropyCost.delta(weighted_array, activations,
out_array) / input_size
backprop = [
layer.backprop(inp, delta_array)
for inp, delta_array in zip(inp_array, delta_array)
]
gradient_w_array = sum([i[0] for i in backprop]) / input_size
gradient_b_array = sum([i[1] for i in backprop]) / input_size
layer.update(gradient_w_array, gradient_b_array, 15)
print('epoch #{}, cost {}'.format(epoch, cost))
def test_forward(self):
layer = FullyConnectedLayer(10, 5)
inp = np.random.normal(size=(10, 1))
forward = layer.forward(inp)
print('forward, ', forward)
self.assertEqual((5, 1), forward.shape)
def test_init(self):
layer = FullyConnectedLayer(10, 5)
self.assertEqual((5, 10), layer.weight_array.shape)
self.assertEqual((5, 1), layer.bias_array.shape)