def test():
sess = fl.Session()
sess.fan_in = 3
sess.fan_out = 1
x = fl.Placeholder(sess, train_x, 'x')
y = fl.Placeholder(sess, train_y, 'y')
# We can choose to apply kernel to x_input or not
x2 = fl.concat(x, fl.square(x), 1)
# x2 = x
S0, W0, b0 = fl.fully_conntected(x2,
30,
activation=fl.sigmoid,
initializer=fl.xavier_initializer())
S1, W1, b1 = fl.fully_conntected(S0,
1,
activation=fl.sigmoid,
initializer=fl.xavier_initializer())
y_ = S1
E = fl.l2loss(y, y_)
optimizer = fl.AdamOptimizer(sess, [E], lr=0.01)
anim = fl.make_animation2d(x,
y,
y_,
E,
optimizer, (0, 7), (0, 7),
interval=1,
blit=True)
plt.show()
def test():
num_images, images, num_rows, num_cols = read_x('data/mnist/train_x')
_, labels = read_y('data/mnist/train_y')
vec_size = 28 * 28
batch_size = 6000
class_num = 10
hidden_sizes = 256, 128
lr = 0.001
epoch = 15
num_test, test_images, _, _ = read_x('data/mnist/test_x')
_, test_labels = read_y('data/mnist/test_y')
print(num_images, num_test, 'Images Read')
images = np.reshape(images, (num_images, vec_size))
test_images = np.reshape(test_images, (num_test, vec_size))
sess = fl.Session()
sess.fan_in = vec_size
sess.fan_out = class_num
input_x = fl.Placeholder(sess, (None, vec_size), 'x')
output_y = fl.Placeholder(sess, (None, class_num), 'y')
H = input_x
for hs in hidden_sizes:
H, _, _ = fl.fully_conntected(H,
hs,
activation=fl.relu,
initializer=fl.xavier_initializer())
y_, _, _ = fl.fully_conntected(H,
class_num,
activation=fl.relu,
initializer=fl.xavier_initializer())
E = fl.softmax_cross_entropy_loss(output_y, y_, 1)
# E = fl.l2loss(output_y, y_)
optimizer = fl.AdamOptimizer(sess, [E], lr=lr)
for _ in range(epoch):
for batch_x, batch_y in mini_batch(images, labels, batch_size):
input_x.set_result(batch_x)
output_y.set_result(batch_y)
optimizer.minimize()
print('E:', E.get_result() / batch_size)
input_x.set_result(test_images)
output_y.set_result(test_labels)
print('E:', E.get_result() / num_test)
print(
'acc:',
np.sum(
np.argmax(y_.get_result(), axis=1) == np.argmax(test_labels, 1)) /
num_test)
def test():
sess = fl.Session()
# Vx + b = y
train_x = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
train_y = np.array([[0], [1], [1], [0]])
x = fl.Placeholder(sess, train_x, 'x')
y = fl.Placeholder(sess, train_y, 'y')
def initializer(*shape):
return fl.xavier(shape, 2, 2)
V0 = fl.Variable(sess, initializer(2, 2))
b0 = fl.Variable(sess, np.zeros(2))
S0 = fl.sigmoid(fl.matmul(x, V0) + b0)
V1 = fl.Variable(sess, initializer(2, 1))
b1 = fl.Variable(sess, np.zeros(1))
S1 = fl.sigmoid(fl.matmul(S0, V1) + b1)
y_ = S1
E = fl.sum(fl.square(y_ - y), axis=0)
optimizer = fl.AdamOptimizer(sess, [E], lr=0.1)
if True: # Pre-calculate before animation
print('start error:', E.get_result())
epoch = 1000
with pb.ProgressBar(max_value=epoch) as bar:
for i in range(epoch):
optimizer.minimize()
bar.update(i)
print('last error:', E.get_result())
anim = fl.make_animation2d(x,
y,
y_,
E,
optimizer, (-1, 2), (-1, 2),
epoch_per_frame=50,
frames=50,
interval=80,
blit=True)
if True:
plt.show()
else:
anim.save('static/xor.gif', writer='imagemagick')
def test(train=True):
sess = fl.Session()
sess.fan_in = 1
sess.fan_out = 1
x = fl.Placeholder(sess, train_x, 'x')
y = fl.Placeholder(sess, train_y, 'y')
S = x
for _ in range(7):
S, _, _ = fl.fully_conntected(S,
100,
activation=fl.tanh,
initializer=fl.xavier_initializer())
S, _, _ = fl.fully_conntected(S,
1,
activation=None,
initializer=fl.xavier_initializer())
y_ = S
# E = fl.avg(fl.avg(fl.square(y - y_), 0), 0)
E = fl.l2loss(y, y_)
optimizer = fl.AdamOptimizer(sess, [E], lr=0.001)
if False: # Pre-training before animation
for _ in pb.progressbar(range(epoch)):
train()
anim = fl.make_animation1d(x,
y,
y_,
E,
optimizer, (-4, 4), (-2, 2),
answer,
interval=1,
blit=True)
plt.show()
def ones(sess, shape, name='ones'):
return fl.Placeholder(sess, np.ones(shape), name)
def zeros(sess, shape, name='zero'):
return fl.Placeholder(sess, np.zeros(shape), name)
def empty_like(sess, a, name='empty'):
return fl.Placeholder(sess, np.empty_like(a.shape), name)
def ones_like(sess, a, name='ones'):
return fl.Placeholder(sess, np.ones_like(a.shape), name)
def zeros_like(sess, a, name='zero'):
return fl.Placeholder(sess, np.zeros_like(a.shape), name)
def empty(sess, shape, name='empty'):
return fl.Placeholder(sess, np.empty(shape), name)