[PYTHON/TENSORFLOW] 컨볼루션 오토 인코더 만들기

■ 컨볼루션 오토 인코더를 만드는 방법을 보여준다.

▶ 예제 코드 (PY)


import matplotlib.pyplot as pp
import numpy as np
import tensorflow as tf
import tensorflow.examples.tutorials.mnist as mnist
from tensorflow.python.framework import ops
import warnings
import random
import os

warnings.filterwarnings("ignore")

os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"

ops.reset_default_graph()

# MNIST 데이터를 로드한다.
print("MNIST 데이터 로드를 시작합니다.")

mnistDatasets = mnist.input_data.read_data_sets("data", one_hot = True)

print("MNIST 데이터 로드를 종료합니다.")

# 네트워크 파라미터

inputLayerNodeCount = 784
learningRate = 0.001
batchSize    = 128
epochCount   = 50

n1 = 16
n2 = 32
n3 = 64
ksize = 5

inputLayerTensor           = tf.placeholder(tf.float32, [None, inputLayerNodeCount])
inputLayerTensorReshape    = tf.reshape(inputLayerTensor, shape = [-1, 28, 28, 1])
correctOutputTensor        = tf.placeholder(tf.float32, [None, inputLayerNodeCount])
correctOutputTensorReshape = tf.reshape(correctOutputTensor, shape = [-1, 28, 28, 1])
dropoutRateTensor          = tf.placeholder(tf.float32)

encoderConvolutionLayer1WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, 1 , n1], stddev = 0.1))
encoderConvolutionLayer2WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, n1, n2], stddev = 0.1))
encoderConvolutionLayer3WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, n2, n3], stddev = 0.1))

decoderConvolutionLayer3WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, n2, n3], stddev = 0.1))
decoderConvolutionLayer2WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, n1, n2], stddev = 0.1))
decoderConvolutionLayer1WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, 1 , n1], stddev = 0.1))

encoderConvolutionLayer1BiasVariable = tf.Variable(tf.random_normal([n1], stddev = 0.1))
encoderConvolutionLayer2BiasVariable = tf.Variable(tf.random_normal([n2], stddev = 0.1))
encoderConvolutionLayer3BiasVariable = tf.Variable(tf.random_normal([n3], stddev = 0.1))

decoderConvolutionLayer3BiasVariable = tf.Variable(tf.random_normal([n2], stddev = 0.1))
decoderConvolutionLayer2BiasVariable = tf.Variable(tf.random_normal([n1], stddev = 0.1))
decoderConvolutionLayer1BiasVariable = tf.Variable(tf.random_normal([1 ], stddev = 0.1))

# 인코더
encoderConvolutionLayer1OutputTensor        = tf.nn.sigmoid(tf.add(tf.nn.conv2d(inputLayerTensorReshape                    ,\
    encoderConvolutionLayer1WeightVariable, strides = [1, 2, 2, 1], padding = "SAME"), encoderConvolutionLayer1BiasVariable))
encoderConvolutionLayer1OutputTensorDropout = tf.nn.dropout(encoderConvolutionLayer1OutputTensor, dropoutRateTensor)
encoderConvolutionLayer2OutputTensor        = tf.nn.sigmoid(tf.add(tf.nn.conv2d(encoderConvolutionLayer1OutputTensorDropout,\
    encoderConvolutionLayer2WeightVariable, strides = [1, 2, 2, 1], padding = "SAME"), encoderConvolutionLayer2BiasVariable))
encoderConvolutionLayer2OutputTensorDropout = tf.nn.dropout(encoderConvolutionLayer2OutputTensor, dropoutRateTensor)
encoderConvolutionLayer3OutputTensor        = tf.nn.sigmoid(tf.add(tf.nn.conv2d(encoderConvolutionLayer2OutputTensorDropout,\
    encoderConvolutionLayer3WeightVariable, strides = [1, 2, 2, 1], padding = "SAME"), encoderConvolutionLayer3BiasVariable))
encoderConvolutionLayer3OutputTensorDropout = tf.nn.dropout(encoderConvolutionLayer3OutputTensor, dropoutRateTensor)

# 디코더
decoderConvolutionLayer3OutputTensor        = tf.nn.sigmoid(tf.add(tf.nn.conv2d_transpose(encoderConvolutionLayer3OutputTensorDropout,\
    decoderConvolutionLayer3WeightVariable, tf.stack([tf.shape(inputLayerTensor)[0], 7 , 7 , n2]), strides = [1, 2, 2, 1], padding = "SAME"),\
    decoderConvolutionLayer3BiasVariable))
decoderConvolutionLayer3OutputTensorDropout = tf.nn.dropout(decoderConvolutionLayer3OutputTensor, dropoutRateTensor)
decoderConvolutionLayer2OutputTensor        = tf.nn.sigmoid(tf.add(tf.nn.conv2d_transpose(decoderConvolutionLayer3OutputTensorDropout,\
    decoderConvolutionLayer2WeightVariable, tf.stack([tf.shape(inputLayerTensor)[0], 14, 14, n1]), strides = [1, 2, 2, 1], padding = "SAME"),\
    decoderConvolutionLayer2BiasVariable))
decoderConvolutionLayer2OutputTensorDropout = tf.nn.dropout(decoderConvolutionLayer2OutputTensor, dropoutRateTensor)
decoderConvolutionLayer1OutputTensor        = tf.nn.sigmoid(tf.add(tf.nn.conv2d_transpose(decoderConvolutionLayer2OutputTensorDropout,\
    decoderConvolutionLayer1WeightVariable, tf.stack([tf.shape(inputLayerTensor)[0], 28, 28, 1 ]), strides = [1, 2, 2, 1], padding = "SAME"),\
    decoderConvolutionLayer1BiasVariable))
decoderConvolutionLayer1OutputTensorDropout = tf.nn.dropout(decoderConvolutionLayer1OutputTensor, dropoutRateTensor)

outputLayerOuputTensor = decoderConvolutionLayer1OutputTensorDropout

costTensor = tf.reduce_sum(tf.square(outputLayerOuputTensor - correctOutputTensorReshape))

optimizerOperation = tf.train.AdamOptimizer(learningRate).minimize(costTensor)

# 그래프를 실행한다.
print("학습을 시작합니다.")

with tf.Session() as session:
    session.run(tf.global_variables_initializer())
    meanImageNDArray = np.zeros((inputLayerNodeCount))
    for epoch in range(epochCount):
        totalBatch = mnistDatasets.train.num_examples // batchSize
        for batch in range(totalBatch):
            batchInputNDArray, _ = mnistDatasets.train.next_batch(batchSize)
            trainInputNDArray = np.array([sourceImageNDArray - meanImageNDArray for sourceImageNDArray in batchInputNDArray])
            noiseTrainInputNDArray = trainInputNDArray + 0.3 * np.random.randn(trainInputNDArray.shape[0], inputLayerNodeCount)
            session.run(optimizerOperation, feed_dict = {inputLayerTensor : noiseTrainInputNDArray, correctOutputTensor : trainInputNDArray, dropoutRateTensor : 0.7})
            print("[%02d/%02d] 비용 : %.4f" % (epoch, epochCount, session.run(costTensor, feed_dict = {inputLayerTensor : noiseTrainInputNDArray,\
                correctOutputTensor : trainInputNDArray, dropoutRateTensor: 1.})))
        if (epoch % 1) == 0:
            exampleCount = 5
            testInputNDArray, _ = mnistDatasets.test.next_batch(exampleCount)
            noisyTestInputNDArray = testInputNDArray + 0.3 * np.random.randn(testInputNDArray.shape[0], inputLayerNodeCount)
            outputNDArray = session.run(outputLayerOuputTensor, feed_dict = {inputLayerTensor : noisyTestInputNDArray, dropoutRateTensor : 1.})
            figure, axesNDArray = pp.subplots(2, exampleCount, figsize = (15, 4))
            for example in range(exampleCount):
                 axesNDArray[0][example].matshow(np.reshape(noisyTestInputNDArray[example, :], (28, 28)), cmap = pp.get_cmap("gray"))
                 axesNDArray[1][example].matshow(np.reshape(np.reshape(outputNDArray[example, ...], (inputLayerNodeCount,)) + meanImageNDArray, (28, 28)),\
                     cmap = pp.get_cmap("gray"))
                 pp.show()

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

import matplotlib.pyplot as pp

import numpy as np

import tensorflow as tf

import tensorflow.examples.tutorials.mnist as mnist

from tensorflow.python.framework import ops

import warnings

import random

import os

warnings.filterwarnings("ignore")

os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"

ops.reset_default_graph()

# MNIST 데이터를 로드한다.

print("MNIST 데이터 로드를 시작합니다.")

mnistDatasets = mnist.input_data.read_data_sets("data", one_hot = True)

print("MNIST 데이터 로드를 종료합니다.")

# 네트워크 파라미터

inputLayerNodeCount = 784

learningRate = 0.001

batchSize = 128

epochCount = 50

n1 = 16

n2 = 32

n3 = 64

ksize = 5

inputLayerTensor = tf.placeholder(tf.float32, [None, inputLayerNodeCount])

inputLayerTensorReshape = tf.reshape(inputLayerTensor, shape = [-1, 28, 28, 1])

correctOutputTensor = tf.placeholder(tf.float32, [None, inputLayerNodeCount])

correctOutputTensorReshape = tf.reshape(correctOutputTensor, shape = [-1, 28, 28, 1])

dropoutRateTensor = tf.placeholder(tf.float32)

encoderConvolutionLayer1WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, 1 , n1], stddev = 0.1))

encoderConvolutionLayer2WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, n1, n2], stddev = 0.1))

encoderConvolutionLayer3WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, n2, n3], stddev = 0.1))

decoderConvolutionLayer3WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, n2, n3], stddev = 0.1))

decoderConvolutionLayer2WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, n1, n2], stddev = 0.1))

decoderConvolutionLayer1WeightVariable = tf.Variable(tf.random_normal([ksize, ksize, 1 , n1], stddev = 0.1))

encoderConvolutionLayer1BiasVariable = tf.Variable(tf.random_normal([n1], stddev = 0.1))

encoderConvolutionLayer2BiasVariable = tf.Variable(tf.random_normal([n2], stddev = 0.1))

encoderConvolutionLayer3BiasVariable = tf.Variable(tf.random_normal([n3], stddev = 0.1))

decoderConvolutionLayer3BiasVariable = tf.Variable(tf.random_normal([n2], stddev = 0.1))

decoderConvolutionLayer2BiasVariable = tf.Variable(tf.random_normal([n1], stddev = 0.1))

decoderConvolutionLayer1BiasVariable = tf.Variable(tf.random_normal([1 ], stddev = 0.1))

# 인코더

encoderConvolutionLayer1OutputTensor = tf.nn.sigmoid(tf.add(tf.nn.conv2d(inputLayerTensorReshape ,\

encoderConvolutionLayer1WeightVariable, strides = [1, 2, 2, 1], padding = "SAME"), encoderConvolutionLayer1BiasVariable))

encoderConvolutionLayer1OutputTensorDropout = tf.nn.dropout(encoderConvolutionLayer1OutputTensor, dropoutRateTensor)

encoderConvolutionLayer2OutputTensor = tf.nn.sigmoid(tf.add(tf.nn.conv2d(encoderConvolutionLayer1OutputTensorDropout,\

encoderConvolutionLayer2WeightVariable, strides = [1, 2, 2, 1], padding = "SAME"), encoderConvolutionLayer2BiasVariable))

encoderConvolutionLayer2OutputTensorDropout = tf.nn.dropout(encoderConvolutionLayer2OutputTensor, dropoutRateTensor)

encoderConvolutionLayer3OutputTensor = tf.nn.sigmoid(tf.add(tf.nn.conv2d(encoderConvolutionLayer2OutputTensorDropout,\

encoderConvolutionLayer3WeightVariable, strides = [1, 2, 2, 1], padding = "SAME"), encoderConvolutionLayer3BiasVariable))

encoderConvolutionLayer3OutputTensorDropout = tf.nn.dropout(encoderConvolutionLayer3OutputTensor, dropoutRateTensor)

# 디코더

decoderConvolutionLayer3OutputTensor = tf.nn.sigmoid(tf.add(tf.nn.conv2d_transpose(encoderConvolutionLayer3OutputTensorDropout,\

decoderConvolutionLayer3WeightVariable, tf.stack([tf.shape(inputLayerTensor)[0], 7 , 7 , n2]), strides = [1, 2, 2, 1], padding = "SAME"),\

decoderConvolutionLayer3BiasVariable))

decoderConvolutionLayer3OutputTensorDropout = tf.nn.dropout(decoderConvolutionLayer3OutputTensor, dropoutRateTensor)

decoderConvolutionLayer2OutputTensor = tf.nn.sigmoid(tf.add(tf.nn.conv2d_transpose(decoderConvolutionLayer3OutputTensorDropout,\

decoderConvolutionLayer2WeightVariable, tf.stack([tf.shape(inputLayerTensor)[0], 14, 14, n1]), strides = [1, 2, 2, 1], padding = "SAME"),\

decoderConvolutionLayer2BiasVariable))

decoderConvolutionLayer2OutputTensorDropout = tf.nn.dropout(decoderConvolutionLayer2OutputTensor, dropoutRateTensor)

decoderConvolutionLayer1OutputTensor = tf.nn.sigmoid(tf.add(tf.nn.conv2d_transpose(decoderConvolutionLayer2OutputTensorDropout,\

decoderConvolutionLayer1WeightVariable, tf.stack([tf.shape(inputLayerTensor)[0], 28, 28, 1 ]), strides = [1, 2, 2, 1], padding = "SAME"),\

decoderConvolutionLayer1BiasVariable))

decoderConvolutionLayer1OutputTensorDropout = tf.nn.dropout(decoderConvolutionLayer1OutputTensor, dropoutRateTensor)

outputLayerOuputTensor = decoderConvolutionLayer1OutputTensorDropout

costTensor = tf.reduce_sum(tf.square(outputLayerOuputTensor - correctOutputTensorReshape))

optimizerOperation = tf.train.AdamOptimizer(learningRate).minimize(costTensor)

# 그래프를 실행한다.

print("학습을 시작합니다.")

with tf.Session() as session:

session.run(tf.global_variables_initializer())

meanImageNDArray = np.zeros((inputLayerNodeCount))

for epoch in range(epochCount):

totalBatch = mnistDatasets.train.num_examples // batchSize

for batch in range(totalBatch):

batchInputNDArray, _ = mnistDatasets.train.next_batch(batchSize)

trainInputNDArray = np.array([sourceImageNDArray - meanImageNDArray for sourceImageNDArray in batchInputNDArray])

noiseTrainInputNDArray = trainInputNDArray + 0.3 * np.random.randn(trainInputNDArray.shape[0], inputLayerNodeCount)

session.run(optimizerOperation, feed_dict = {inputLayerTensor : noiseTrainInputNDArray, correctOutputTensor : trainInputNDArray, dropoutRateTensor : 0.7})

print("[%02d/%02d] 비용 : %.4f" % (epoch, epochCount, session.run(costTensor, feed_dict = {inputLayerTensor : noiseTrainInputNDArray,\

correctOutputTensor : trainInputNDArray, dropoutRateTensor: 1.})))

if (epoch % 1) == 0:

exampleCount = 5

testInputNDArray, _ = mnistDatasets.test.next_batch(exampleCount)

noisyTestInputNDArray = testInputNDArray + 0.3 * np.random.randn(testInputNDArray.shape[0], inputLayerNodeCount)

outputNDArray = session.run(outputLayerOuputTensor, feed_dict = {inputLayerTensor : noisyTestInputNDArray, dropoutRateTensor : 1.})

figure, axesNDArray = pp.subplots(2, exampleCount, figsize = (15, 4))

for example in range(exampleCount):

axesNDArray[0][example].matshow(np.reshape(noisyTestInputNDArray[example, :], (28, 28)), cmap = pp.get_cmap("gray"))

axesNDArray[1][example].matshow(np.reshape(np.reshape(outputNDArray[example, ...], (inputLayerNodeCount,)) + meanImageNDArray, (28, 28)),\

cmap = pp.get_cmap("gray"))

pp.show()