Tensorflow training을 save하고 restore하는 방법

training modeling 을 저장했다가 다시 불러서 사용하게 되면,

처음부터 training을 할 필요가 없어져서 매우 유용합니다.

그 방법 역시 매우 간단한데요. 아래 예제가 있습니다.

한번 활용해보세요.

optimizer = tf.train.AdamOptimizer(learning_rate)

train = optimizer.minimize(cost)

saver = tf.train.Saver()

sess = tf.Session()

sess.run(tf.global_variables_initializer())

# training 을 저장할 file 명

mytrain= "./mytain.ckpt"

if os.path.exists(mytrain+".meta"):

# 파일에서 loading

    saver.restore(sess, mytrain)

else :

    for step in range(2001) :

        cost_val, hy_val, _ = sess.run([cost, hypothesis, train], feed_dict={X: trainX, Y: trainY})

        if step % 10 ==0 :

            print(step, "Cost: ", cost_val, "\nPrediction:\n", hy_val)

#file에 저장

    saver.save(sess, mytrain)

test_predict = sess.run(hypothesis, feed_dict= {X: testX})

아래와 같이 graph 까지 복원해서 사용할 수 있는 방법도 있습니다.

saver = tf.train.import_meta_graph(mytrain+".meta")

saver.restore(sess, mytrain)

ex)

save

train_size = int(len(dataY) * 0.7)

test_size =  len(dataY) - train_size

trainX = np.array(dataX[0:train_size])

testX = np.array(dataX[train_size:len(dataX)])

trainY = np.array(dataY[0:train_size])

testY = np.array(dataY[train_size:len(dataY)])

input_len = data_dim*seq_length

X = tf.placeholder(tf.float32, [None, input_len], name='X')

Y = tf.placeholder(tf.float32, [None, 1], name='Y')

W = tf.Variable(tf.random_normal([input_len, 1]), name='weight')

b = tf.Variable(tf.random_normal([1]), name='bias')

#hypothesis = tf.matmul(X, W) + b

hypothesis = tf.add(b, tf.matmul(X, W), name="h")

cost = tf.reduce_mean(tf.square(hypothesis - Y))

optimizer = tf.train.AdamOptimizer(learning_rate)

train = optimizer.minimize(cost)

saver = tf.train.Saver()

sess = tf.Session()

sess.run(tf.global_variables_initializer())

for step in range(2001) :

    cost_val, hy_val, _ = sess.run([cost, hypothesis, train], feed_dict={X: trainX, Y: trainY})

    if step % 10 ==0 :

        print(step, "Cost: ", cost_val, "\nPrediction:\n", hy_val)

saver.save(sess, mytrain)

resotre

if os.path.exists(mytrain+".meta"):

    sess = tf.Session()

    sess.run(tf.global_variables_initializer())

    saver = tf.train.import_meta_graph(mytrain+".meta")

    saver.restore(sess, mytrain)

    graph = tf.get_default_graph()

    X = graph.get_tensor_by_name('X:0')

    W = graph.get_tensor_by_name('weight:0')

    b = graph.get_tensor_by_name('bias:0')

    hypothesis = graph.get_tensor_by_name('h:0')   #tf.matmul(X, W) + b

    test_predict = sess.run(hypothesis, feed_dict={X: testX})