텐서플로우 가이드북
https://tensorflowkorea.gitbooks.io/tensorflow-kr/content/g3doc/
tf 모델 저장/복원
ckpt(체크포인트) : 가중치 저장
pb(포토버프) : 네트워크 그래프
tfrecords : 데이터
import tensorflow as tf
state1 = tf.Variable(0, name = 'state1')
state2 = tf.Variable(10, name = 'state2')
init_op = tf.global_variables_initializer()
saver = tf.train.Saver({"mys2": state2})
one = tf.constant(1)
new_value = tf.add(state2, one)
update = tf.assign(state2, new_value)
with tf.Session() as sess:
sess.run(init_op)
sess.run(update)
save_path = saver.save(sess, '/tmp/model1.ckpt')
print(" Model saved in file: ", save_path)
print(sess.run(state2))
#----------------------
import tensorflow as tf
state1 = tf.Variable(0, name = 'state1')
state2 = tf.Variable(20, name = 'state2')
# one = tf.constant(1)
# new_value = tf.add(state2, one)
# update = tf.assign(state2, new_value)
init_op = tf.global_variables_initializer()
saver = tf.train.Saver({"mys2": state2})
with tf.Session() as sess:
#sess.run(init_op)
saver.restore(sess, '/tmp/model1.ckpt')
#sess.run(init_op)
print(" Model restored from file ")
print(sess.run(state2))tf2에서 tf1 함수 섞어쓰기
import tensorflow as tf
import tensorflow.compat.v1 as tf1
tf1.disable_eager_execution()
state = tf.Variable(0, name= 'counter')
one = tf.constant(1)
new_value = tf.add(state, one)
update = tf1.assign(state, new_value)
init_op = tf1.global_variables_initializer()
writer = tf.summary.create_file_writer('/tmp/log')
with tf1.Session() as sess:
sess.run(init_op)
print(sess.run(state))
for _ in range(3):
sess.run(update)
print(sess.run(update))tensorboard에서 flow graph 확인
/tmp/log 이런식으로 디렉토리 안에 그래프 저장 후, 터미널에서
tensorboard --logdir=/tmp/log 치고,
ctrl 눌러서 링크 클릭하면 flow graph 확인 가능
tf1.reset_default_graph()
with tf1.name_scope("DefineVariable"):
x_data = np.random.rand(100).astype('f')
y_data = x_data * .1 + .3
W = tf.Variable(tf.random.uniform([1], -1., 1.))
b = tf.Variable(tf.zeros([1]))
y = W * x_data + b
with tf1.name_scope("Logit"):
loss = tf.reduce_mean(tf.square(y - y_data))
optimizer = tf1.train.GradientDescentOptimizer(.5)
train = optimizer.minimize(loss)
init = tf1.global_variables_initializer()
sess = tf1.Session()
sess.run(init)
for step in range(201):
sess.run(train)
writer = tf1.summary.FileWriter("/tmp/log2", graph = sess.graph)
if step % 20 == 0:
print(step, 'W =', sess.run(W),
'b =', sess.run(b))
sess.close()Computation Graph (Using tf2.x)
import tensorflow as tf
import numpy as np
import tensorflow.compat.v1 as tf1
tf1.enable_eager_execution()
x_data = np.random.rand(100).astype('f')
y_data = x_data * .1 + .3
init = lambda : (tf.Variable(tf.random.uniform([1], -1, 1),
name = 'Weight'), tf.Variable(tf.zeros([1])))
W, b = init()
yp = lambda x : W * x + b
mse = lambda yp, yt : tf.reduce_mean(tf.square(yp-yt))
way = 1
if way == 1:
for step in range(2001):
with tf.GradientTape() as tape:
y = yp(x_data)
loss = mse(y, y_data)
grads = tape.gradient(loss, [W, b])
if step % 200 == 0:
print('\n\rW = ', W.numpy(), '\n\rb =', b.numpy(),
'\n\rg[0] =', grads[0].numpy(), '\n\rg[1] =', grads[1].numpy())
W.assign(W - 0.2 * grads[0].numpy())
b.assign(b - 0.2 * grads[1].numpy())
elif way == 2:
optimizer = tf.optimizers.SGD(learning_rate = 0.2)
for step in range(201):
with tf.GradientTape() as tape:
y = yp(x_data)
loss = mse(y, y_data)
grads = tape.gradient(loss, [W, b])
process_grads = [g for g in grads]
grads_and_vars = zip(process_grads, [W, b])
if step % 20 == 0:
print('\n\rW = ', W.numpy(), '\n\rb =', b.numpy(),
'\n\rg[0] =', grads[0].numpy(), '\n\rg[1] =', grads[1].numpy())
optimizer.apply_gradients(grads_and_vars)
elif way == 3:
yp, yt = yp(x_data), y_data
optimizer = tf.optimizers.SGD(learning_rate = 0.1)
for step in range(201):
if step % 20 == 0:
print('\n\rW = ', W.numpy(), '\n\rb =', b.numpy())
mse_min = lambda : tf.reduce_mean(tf.square((W * x_data + b) - yt))
optimizer.minimize(mse_min, var_list = [W, b])Cost and Gradient
import tensorflow as tf
import matplotlib.pyplot as plt
X = [1., 2., 3.]
Y = [1., 2., 3.]
# W = tf.placeholder(tf.float32)
hypothesis = lambda W: tf.multiply(W, X)
cost = lambda hx: tf.reduce_sum(tf.pow(hx-Y, 2)) / len(X)
W_val, cost_val = [], []
for i in range(-30, 51):
xPos = i * 0.1
#xPos = i
yPos = cost(hypothesis(xPos))
print('\n\rW =', xPos, '\n\rCost =', yPos)
W_val.append(xPos)
cost_val.append(yPos)
# sess.close()
plt.plot(W_val, cost_val, 'ro')
plt.ylabel('Cost')
plt.xlabel('W')
plt.show()Gradient Descent Optimization (1)
import tensorflow as tf
x_data = [1., 2., 3.]
y_data = [1., 2., 3.]
W = tf.Variable(tf.random.uniform([1], -1.0, 1.0))
b = tf.Variable(tf.random.uniform([1], -1.0, 1.0))
hx = lambda : W * x_data + b
cost = lambda : tf.reduce_mean(tf.square(hx() - y_data))
rate = tf.Variable(0.1)
opt = tf.optimizers.SGD(rate)
for step in range(2001):
if step % 20 == 0:
print('{:4} {} {} {}'.format(step, cost(), W.numpy(), b.numpy()))
opt.minimize(cost, var_list = [W, b])Gradient Descent Optimization (2)
import tensorflow as tf
x_data = [1., 2., 3., 4.]
y_data = [2., 4., 6., 8.]
W = tf.Variable(tf.random.uniform([1], -100.0, 100.0))
b = tf.Variable(tf.random.uniform([1], -100.0, 100.0))
hx = lambda x_data : W * x_data + b
cost = lambda y_data : tf.reduce_mean(tf.square(hx(x_data) - y_data))
rate = tf.Variable(0.1)
opt = tf.optimizers.SGD(rate)
for step in range(601):
cost_min = lambda : cost(y_data)
opt.minimize(cost_min, var_list = [W, b])
if step % 20 == 0:
print('{:4} {} {} {}'.format(step, cost(y_data), W.numpy(), b.numpy()))
print('\n\rx = 2.5:', hx(2.5).numpy(), '\n\rx = 10:', hx(10).numpy(), '\n\rx = 2.5, 10', hx([2.5, 10]).numpy())Logistic Regression
# import tensorflow as tf
# import numpy as np
# tf.get_logger().setLevel('ERROR')
xy = np.loadtxt('train.csv', delimiter = ',', unpack = True, dtype = 'f')
x_data = xy[:-1]
y_data = xy[-1]
W = tf.Variable(tf.random.uniform([1, len(x_data)], -1., 1.))
h = lambda X : tf.matmul(W, X)
hx = lambda X : tf.math.divide(1., 1. + tf.math.exp(-h(X)))
cost = lambda Y : -tf.reduce_mean(Y * tf.math.log(hx(x_data)) + (1-Y) * tf.math.log(1 - hx(x_data)))
rate = tf.Variable(.1)
opt = tf.optimizers.SGD(rate)
for step in range(2001):
with tf.GradientTape() as tape:
y = hx(x_data)
loss = cost(y_data)
grads = tape.gradient(loss, [W])
process_grads = [g for g in grads]
grads_and_vars = zip(process_grads, [W])
opt.apply_gradients(grads_and_vars)
if step % 200 == 0:
print(step, grads[0].numpy(), '\n\rW = ', W.numpy())
print('-'*60)
print('[1, 2, 2] :', hx([[1.], [2.], [2.]]))
print('[1, 5, 5] :', hx([[1.], [5.], [5.]]))
print('[1, 2, 2], [1, 5, 5] :', hx([[1., 1.], [2., 5.], [2., 5.]]) >.5)
Backend Web Developer