{"id":2110,"date":"2020-01-02T22:51:07","date_gmt":"2020-01-02T13:51:07","guid":{"rendered":"http:\/\/141.164.34.82\/?p=2110"},"modified":"2020-01-04T08:11:12","modified_gmt":"2020-01-03T23:11:12","slug":"ann-%ed%8f%89%ea%b7%a0%ea%b3%bc-%ed%91%9c%ec%a4%80%ed%8e%b8%ec%b0%a8-%ec%b6%9c%eb%a0%a5-%eb%aa%a8%ed%98%95","status":"publish","type":"post","link":"http:\/\/ds.sumeun.org\/?p=2110","title":{"rendered":"ANN: \ud3c9\uade0\uacfc \ud45c\uc900\ud3b8\ucc28 \ucd9c\ub825 \ubaa8\ud615"},"content":{"rendered":"

keras\uc640 \uc778\uacf5\uc2e0\uacbd\ub9dd(ANN; A<\/strong>rtificial N<\/strong>eural N<\/strong>etwork)<\/h2>\n

\uc778\uacf5\uc2e0\uacbd\ub9dd\uc774 \ub300\ub2e8\ud574 \ubcf4\uc77c\uc9c0\ub77c\ub3c4, \uadf8 \ud575\uc2ec\uc744 \ub530\uc9c0\uace0 \ub4e4\uc5b4\uac04\ub2e4\uba74, \ub2e4\uc74c\uc758 \ud568\uc218\ub97c \ucd94\uc815\ud558\ub294 \uac83\uc774\ub77c\uace0 \ubcfc \uc218 \uc788\ub2e4.<\/p>\n

\\[f(x) = \\mathbb{E}[y|x]\\]<\/p>\n

\uc774\ub294 \uc190\uc218\ud68c\uadc0(Manual Regression)<\/a>\uc5d0\uc11c\ub3c4 \uc124\uba85\ud55c \uac83\ucc98\ub7fc, \uc9c1\uad00\uc801\uc73c\ub85c \uc0dd\uac01\ud574\ub3c4 \uadf8\ub9ac \uc5b4\ub835\uc9c0 \uc54a\uc740 \uac1c\ub150\uc774\ub2e4. \ud558\uc9c0\ub9cc \uace0\ucc28\uc6d0\uc758 \uc124\uba85\ubcc0\uc218 \\(\\vec{x}\\) \uc640 \uc0d8\ud50c\uc758 \uc218\uac00 \ucda9\ubd84\ud558\uc9c0 \uc54a\uc744 \ub54c\uc5d0\ub294 \uc55e\uc5d0\uc11c \uc124\uba85\ud55c \uc190\uc73c\ub85c(?) \ub610\ub294 \uc21c\uc9c4\ud55c(naive?) \ubc29\ubc95\uc73c\ub85c\ub294 \ud55c\uacc4\uac00 \ubd80\ub52a\ud788\uac8c \ub41c\ub2e4.<\/p>\n

\uc5ec\ub7ec \ube14\ub799 \ubc15\uc2a4 \ubaa8\ud615\ub4e4<\/a>\uc740 \uc800\ub9c8\ub2e4\uc758 \uac00\uc815(assumption)\uc744 \uac00\uc9c0\uace0 \uace0\ucc28\uc6d0\uc131\uc744 \ud574\uacb0\ud558\ub824\uace0 \ud55c\ub2e4\uace0 \uc0dd\uac01\ud558\uba74 \ub420 \uac83\uc774\ub2e4. \uadf8 \uc911\uc5d0 \ucd5c\uadfc\uc5d0 \uac01\uad11\uc744 \ubc1b\uace0 \uc788\ub294 \ubc29\ubc95\uc740 \uc2ec\uce35 \uc2e0\uacbd\ub9dd<\/strong>\uc774\ub2e4. keras<\/code>\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc2ec\uce35 \uc2e0\uacbd\ub9dd \ubaa8\ud615\uc744 \ubcf8\uaca9\uc801\uc73c\ub85c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub2e4.<\/p>\n

\uccab \ubc88\uc9f8 \ubaa8\ud615 : \\(f(x) = \\mathbb{E}[y|x]\\)<\/h2>\n

\ub370\uc774\ud130\ub294 \ub2e4\uc74c\uacfc \uac19\ub2e4.<\/p>\n

## Data\r\nN = 1000\r\nx = runif(N, -2, 2)\r\n#y = sin(x*pi) + abs(x)\/2 * rnorm(N)\r\ny = sin(x*pi) + 1\/2*(abs(cos(x))+1\/8) * rnorm(N)\r\n#y = sin(x*pi) + abs(cos(x)) * rnorm(N)\r\n\r\nplot(y ~ x)\r\n<\/code><\/pre>\n
\"plot<\/p>\n
\uadf8\ub9ac\uace0, \uc2e4\uc81c \ud3c9\uade0\uacfc \\(\\pm\\) 1 \ud45c\uc900\ud3b8\ucc28\uc758 \uadf8\ub798\ud504\ub294 \ub2e4\uc74c\uacfc \uac19\ub2e4.<\/p>\n
curve(sin(x*pi), xlim=c(-2,2), ylim=c(-2,2))\r\ncurve(sin(x*pi) + 1\/2*(abs(cos(x))+1\/8), xlim=c(-2,2), add=TRUE, lty='dotted')\r\ncurve(sin(x*pi) - 1\/2*(abs(cos(x))+1\/8), xlim=c(-2,2), add=TRUE, lty='dotted')\r\n<\/code><\/pre>\n
\"plot<\/p>\n
\uc774\ub97c \uc2ec\uce35 \uc778\uacf5\uc2e0\uacbd\ub9dd\uc73c\ub85c \ucd94\uc815\ud574 \ubcf4\uc790.<\/p>\n
library(keras)\r\n<\/code><\/pre>\n
model <- keras_model_sequential()\r\nmodel %>%\r\n  layer_dense(units = 32, input_shape = 1) %>%\r\n  layer_activation_leaky_relu() %>%\r\n  layer_dense(units=32) %>%\r\n  layer_activation_leaky_relu() %>%\r\n  layer_dense(units = 1, activation = 'linear')\r\n<\/code><\/pre>\n
model %>% compile(\r\n  optimizer = \"adam\",\r\n  #lr=0.01,\r\n  loss = \"mse\",\r\n  metrics = \"mae\"\r\n)\r\n<\/code><\/pre>\n
history <-\r\n  model %>% fit(x,\r\n                y,\r\n                batch_size = 100,\r\n                epochs = 100)\r\nplot(history)\r\n<\/code><\/pre>\n
\"plot<\/p>\n
\uc774\uc81c \ubaa8\ud615\uc758 \uc608\uce21\uacfc \uc2e4\uc81c \ud3c9\uade0\uc744 \ud568\uaed8 \uadf8\ub824\ubcf4\uc790. \ud68c\uc0c9\uc740 \uc2e4\uc81c \ud3c9\uade0\uacfc \\(\\pm 1\\) \ud45c\uc900\ud3b8\ucc28\ub97c \ubcf4\uc5ec\uc8fc\uace0, \uad75\uc740 \uac80\uc815\uc0c9 \uc810\uc120\uc740 \uc2ec\uce35\uc2e0\uacbd\ub9dd\uc73c\ub85c \ucd94\uc815\ud55c \ud3c9\uade0(\ud68c\uadc0\uc120)\uc774\ub2e4.<\/p>\n
# PREDICTION\r\ntest_x <- seq(-2,2,0.05)\r\npred_y <- predict(model, test_x)[,1]\r\ndatPred <- data.frame(x=test_x, y=pred_y) \r\n\r\n# TRUE \r\ntrue_y <- sin(test_x*pi)\r\ntrue_yUpper <- sin(test_x*pi) + 1\/2*(abs(cos(test_x))+1\/8)\r\ntrue_yLower <- sin(test_x*pi) - 1\/2*(abs(cos(test_x))+1\/8)\r\ndatTrue <- data.frame(x=test_x, y=true_y, ymax=true_yUpper, ymin=true_yLower)\r\n\r\nrequire(ggplot2)\r\nggplot() + \r\n  geom_line(data=datPred, aes(x=x, y=y), linetype='dotted') +\r\n  geom_line(data=datTrue, aes(x=x, y=y), col='grey20') + \r\n  geom_ribbon(data=datTrue, aes(x=x, ymin=ymin, ymax=ymax), alpha=0.15)\r\n<\/code><\/pre>\n
\"plot<\/p>\n
\uc790\ub8cc\uc758 \uc911\uc2ec\ubd80\uc5d0\uc11c\ub294 \uc0c1\ub2f9\ud788 \uc815\ud655\ud558\uc9c0\ub9cc, \uc591\ub05d\uc5d0\uc11c\ub294 \uadf8\ub9ac \uc815\ud655\ud558\uc9c0 \ubabb\ud588\ub2e4. \ud558\uc9c0\ub9cc \uc5ec\uae30\uc11c \ud55c \uac00\uc9c0 \ub354 \uc8fc\ubaa9\ud560 \uc810\uc740 \uc2e4\uc81c \ub370\uc774\ud130\ub294 \\(\\mathbb{V}\\textrm{ar}[y|x]\\) \uc5ed\uc2dc \ubcc0\ud558\uace0 \uc788\ub2e4\ub294 \uc810\uc774\ub2e4. \uc6b0\ub9ac\uac00 \ub2e4\ub8ec \uc704\uc758 \uc2ec\uce35 \ubaa8\ud615\uc740 \uc870\uac74\ubd80 \ubd84\uc0b0\uc740 \ub2e4\ub8e8\uc9c0 \ubabb\ud55c\ub2e4.<\/p>\n
\ub450 \ubc88\uc9f8 \ubaa8\ud615 : \\(f_1(x) = \\mathbb{E}[y|x], f_2(x) = \\mathbb{V}\\textrm{ar}[y|x]\\)<\/h2>\n
\uc2ec\uce35\uc2e0\uacbd\ub9dd\uc758 \ucd9c\ub825\uc744 \ub450 \uac1c \uc774\uc0c1\uc73c\ub85c \ub9cc\ub4e4 \uc218 \uc788\ub2e4. \ud754\ud788 \ub2e4\uc911 \ucd9c\ub825(multiple output, multi-head)\ub77c\uace0 \uc598\uae30\ud55c\ub2e4.<\/p>\n
keras<\/code>\ub97c \uc0ac\uc6a9\ud558\uba74 \uc27d\ub2e4. \uc704\uc758 \ubaa8\ud615\uc5d0\uc11c \ub9c8\uc9c0\ub9c9 output\ub9cc \ubc14\uafd4\uc8fc\uba74 \ub41c\ub2e4.<\/p>\n
model %>%\r\n  layer_dense(units = 32, input_shape = 1) %>%\r\n  layer_activation_leaky_relu() %>%\r\n  layer_dense(units=32) %>%\r\n  layer_activation_leaky_relu() %>%\r\n  layer_dense(units = 2, activation = 'linear')\r\n<\/code><\/pre>\n
\ud3c9\uade0-\ud45c\uc900\ud3b8\ucc28 \ucd9c\ub825 \ubaa8\ud615\uc758 \uc190\uc2e4\ud568\uc218<\/h3>\n
\ubb38\uc81c\ub294 loss<\/code>\uc774\ub2e4.<\/p>\n
\uac00\uc7a5 \uc190\uc26c\uc6b4 \ubc29\ubc95\uc740 \\(-\\log \\mathbb{P}(y|x)\\) \ub85c \uc124\uc815\ud558\ub294 \uac83\uc774\ub2e4. \uc774\ub54c \\(\\mathbb{P}(y|x)\\) \uc5d0 \\(\\log\\) \ub97c \ucde8\ud558\ub294 \uac83\uc740 \uc6b0\ub9ac\uac00 \uc774 \uc190\uc2e4\ud569\uc758 \ucd5c\uc18c\ud654\ud558\ub294 \ubaa8\uc218(\uc2e0\uacbd\ub9dd weights)\ub97c \uad6c\ud558\uae30 \ub54c\ubb38\uc774\ub2e4. \ub2e4\uc74c \uc2dd\uc5d0\uc11c \ud655\uc778\ud560 \uc218 \uc788\ub4ef\uc774, \uc774\ub294<\/p>\n
\\[\\log \\mathbb{P}(y_1|x_1) + \\log \\mathbb{P}(y_2|x_2) = \\log \\mathbb{P}(y_1|x_1)\\mathbb{P}(y_2|x_2)\\]<\/p>\n
\uacb0\uad6d \ubaa8\ub4e0 \uad00\ucc30 \\((x_1, y_1), (x_2, y_2), \\cdots\\) \uc774 \uc11c\ub85c \ub3c5\ub9bd\uc77c \ub54c, \ubaa8\ub4e0 \uad00\ucc30\uc5d0 \ub300\ud574 \ud655\ub960\uc744 \ucd5c\ub300\ub85c \ud558\ub294 \ucd5c\ub300\uc6b0\ub3c4\ucd94\uc815\uac12\uc744 \uad6c\ud558\ub294 \uac83\uacfc \uac19\ub2e4.<\/p>\n
\uc5ec\uae30\uc11c \\(y\\) \uc758 \uc870\uac74\ubd80 \ud655\ub960 \ubd84\ud3ec \\(\\mathbb{P}(y|x)\\) \ub294 \ub2e4\uc74c\uacfc \uac19\uc774 \uac00\uc7a5 \ub9cc\ub9cc\ud55c \uc815\uaddc\ubd84\ud3ec\ub97c \uac00\uc815\ud558\uc790.<\/p>\n
\\[\\mathbb{P}(y | x) \\sim \\mathcal{N}(\\mu(x), \\sigma^2(x))\\]<\/p>\n
\uadf8\ub9ac\uace0 \uc2e0\uacf5\uc2e0\uacbd\ub9dd\uc758 \ub450 \ucd9c\ub825 \\(y_1, y_2\\) \ub97c \\(\\mu(x), \\sigma(x)\\) \ub85c \uc0ac\uc6a9\ud560 \uac83\uc774\ub2e4.<\/p>\n
\\[\\mathbb{P}(y | x) = \\frac{1}{\\sigma\\sqrt{2\\pi}}\\exp\\left(-\\frac{(x-\\mu)^2}{2\\sigma^2}\\right)\\]<\/p>\n
\\[\\log \\mathbb{P}(y | x) = -\\log{\\sigma}-\\frac{1}{2}\\log{2\\pi}-\\frac{(x-\\mu)^2}{2\\sigma^2}\\]<\/p>\n
\\[-\\log \\mathbb{P}(y | x) = \\log{\\sigma}+\\frac{1}{2}\\log{2\\pi}+\\frac{(x-\\mu)^2}{2\\sigma^2}\\]<\/p>\n
\uc5ec\uae30\uc11c \\(\\frac{1}{2}\\log{2\\pi}\\) \ub294 \uc0c1\uc218\uc774\ubbc0\ub85c \\(-\\log \\mathbb{P}(y | x)\\) \uc758 \ud569\uc744 \ucd5c\uc18c\ud654\ud558\ub294 \\(\\mu(x)\\) , \\(\\sigma(x)\\) \ub97c \ucc3e\ub294 \uac83\uc740 \ub2e4\uc74c\uc744 \ucd5c\uc18c\ud654\ud558\ub294 \\(\\mu(x)\\) , \\(\\sigma(x)\\) \ub97c \ucc3e\ub294 \ubb38\uc81c\uc640 \ub3d9\uc77c\ud558\ub2e4.<\/p>\n
\\[ \\log{\\sigma} + \\frac{(x-\\mu)^2}{2\\sigma^2}\\]<\/p>\n
\uc190\uc2e4\ud568\uc218 \uad6c\ud604<\/h3>\n
\uc704\uc758 \uc190\uc2e4\ud568\uc218\ub294 keras\uc5d0\uc11c \ub2e4\uc74c\uacfc \uac19\uc774 \uad6c\ud604\ud560 \uc218 \uc788\ub2e4.<\/p>\n
gauss_loss <- function(t, pred) {\r\n  #k_mean(k_square(pred[1,] - t[1,]) + pred[2,])\r\n  #k_mean(k_square((pred[,1] - t[,1])\/pred[,2])\/2 + pred[,2])\r\n  k_mean(k_square((pred[,1] - t[,1])\/k_exp(pred[,2]))\/2 + pred[,2] )\r\n}\r\n\r\nmodel %>% compile(\r\n  optimizer = \"adam\",\r\n  loss = function(y_true, y_pred)\r\n#    tilted_loss(quantile, y_true, y_pred),\r\n    gauss_loss(y_true, y_pred),\r\n  metrics = \"mae\"\r\n)\r\n<\/code><\/pre>\n
\uc870\uae08\ub9cc \uc124\uba85\ud574 \ubcf4\uc790\uba74, loss=function(y_true, y_pred) {}<\/code> \uc5d0\uc11c y_true<\/code>\uc640 y_pred<\/code>\ub294 \ubaa8\ub450 \uae38\uc774 2\uc758 \ubca1\ud130\uc774\ub2e4. y_pred<\/code>\ub294 \uc778\uacf5\uc2e0\uacbd\ub9dd\uc774 \ucd9c\ub825\ud558\ub294 \ub450 \uac12\uc744 \ub2f4\uace0 \uc788\ub294 \uae38\uc774 2\uc758 \ubca1\ud130\uac00 \ub418\uace0, y_true<\/code>\ub294 \ub3d9\uc77c\ud558\uac8c \uae38\uc774 2\uc758 \ubca1\ud130\ub85c \uc0c1\uc815\ud588\ub2e4. \ud558\uc9c0\ub9cc \uc6b0\ub9ac\uac00 \uac00\uc9c0\uace0 \uc788\ub294 \ub370\uc774\ud130\ub294 \\((x,y)\\) \ud615\ud0dc\ub85c \\(y\\) \ubfd0\uc774\uae30 \ub54c\ubb38\uc5d0, y_true<\/code>\uc758 \ub450 \ubc88\uc9f8 \uc6d0\uc18c\ub294 0\uc73c\ub85c \ucc44\uc6cc\ub123\uc5c8\ub2e4.<\/p>\n
\uadf8\ub9ac\uace0 gauss_loss<\/code>\ub294 y_true<\/code>\uc640 y_pred<\/code>\ub97c \ubc1b\uc544, \uc704\uc5d0\uc11c \uc18c\uac1c\ud55c \uc190\uc2e4\ud568\uc218\ub97c \uad6c\ud558\uac8c \ub41c\ub2e4. \uc544, \uc544\ub2c8\ub2e4. \uc778\uacf5\uc2e0\uacbd\ub9dd\uc758 \ucd9c\ub825\uc740 \uc74c\uc218\ub3c4 \uac00\ub2a5\ud558\uae30 \ub54c\ubb38\uc5d0 \\(\\exp\\) \ub97c \uc801\uc6a9\ud55c \uac12\uc774 \ud45c\uc900\ud3b8\ucc28\uac00 \ub418\ub3c4\ub85d \uc870\uc815\uc744 \ud558\uc600\ub2e4.<\/p>\n
train_Y <- matrix(c(y, rep(0, length(y))), length(y), 2)\r\nhistory <-\r\n  model %>% fit(x,\r\n                train_Y,\r\n                batch_size = 100,\r\n                epochs = 1000)\r\nplot(history)\r\n<\/code><\/pre>\n
\"plot<\/p>\n
\uc774\uc81c \ub2e4\uc2dc \ud55c\ubc88 \uc2e4\uc81c \ubaa8\ud615\uacfc \ucd94\uc815\ub41c \ubaa8\ud615\uc744 \ube44\uad50\ud574\ubcf4\uc790.<\/p>\n
# PREDICTION\r\ntest_x <- seq(-2,2,0.05)\r\npred_y <- predict(model, test_x)[,1]\r\npred_sd <- predict(model, test_x)[,2]\r\ndatPred <- data.frame(x=test_x, y=pred_y,\r\n                      y1=pred_y+exp(pred_sd), \r\n                      y2=pred_y-exp(pred_sd)) \r\n\r\n# TRUE \r\ntrue_y <- sin(test_x*pi)\r\ntrue_yUpper <- sin(test_x*pi) + 1\/2*(abs(cos(test_x))+1\/8)\r\ntrue_yLower <- sin(test_x*pi) - 1\/2*(abs(cos(test_x))+1\/8)\r\ndatTrue <- data.frame(x=test_x, y=true_y, ymax=true_yUpper, ymin=true_yLower)\r\n\r\nrequire(ggplot2)\r\nggplot() + \r\n  geom_line(data=datPred, aes(x=x, y=y), linetype='dotted') +\r\n  geom_line(data=datPred, aes(x=x, y=y1), linetype='dotted') +\r\n  geom_line(data=datPred, aes(x=x, y=y2), linetype='dotted') +\r\n  geom_line(data=datTrue, aes(x=x, y=y), col='grey20') + \r\n  geom_ribbon(data=datTrue, aes(x=x, ymin=ymin, ymax=ymax), alpha=0.15)\r\n<\/code><\/pre>\n
\"plot<\/p>\n
\uc5b4\uc9f8 output\uc774 \ud558\ub098\uc600\ub358 \uacbd\uc6b0\ubcf4\ub2e4 \ub354 \uc815\ud655\ud55c \ub4ef \ubcf4\uc778\ub2e4. multi-head\uc778 \uacbd\uc6b0\uc5d0 regularization \ud6a8\uacfc\uac00 \uc788\ub2e4\uace0 \uc54c\ub824\uc838 \uc788\ub2e4.<\/p>\n
\ube45\ub370\uc774\ud130 \uc2dc\ubbac\ub808\uc774\uc158<\/h2>\n
## Data\r\nN = 10000\r\nx = runif(N, -2, 2)\r\n#y = sin(x*pi) + abs(x)\/2 * rnorm(N)\r\ny = sin(x*pi) + 1\/2*(abs(cos(x))+1\/8) * rnorm(N)\r\n#y = sin(x*pi) + abs(cos(x)) * rnorm(N)\r\n\r\nplot(y ~ x)\r\n<\/code><\/pre>\n
\"plot<\/p>\n
## To initialize the weights\r\nmodel <- keras_model_sequential()\r\nmodel %>%\r\n  layer_dense(units = 32, input_shape = 1) %>%\r\n  layer_activation_leaky_relu() %>%\r\n  layer_dense(units=32) %>%\r\n  layer_activation_leaky_relu() %>%\r\n  layer_dense(units = 2, activation = 'linear')\r\n\r\ngauss_loss <- function(t, pred) {\r\n  #k_mean(k_square(pred[1,] - t[1,]) + pred[2,])\r\n  #k_mean(k_square((pred[,1] - t[,1])\/pred[,2])\/2 + pred[,2])\r\n  k_mean(k_square((pred[,1] - t[,1])\/k_exp(pred[,2]))\/2 + pred[,2] )\r\n}\r\n\r\nmodel %>% compile(\r\n  optimizer = \"adam\",\r\n  loss = function(y_true, y_pred)\r\n#    tilted_loss(quantile, y_true, y_pred),\r\n    gauss_loss(y_true, y_pred),\r\n  metrics = \"mae\"\r\n)\r\n\r\ntrain_Y <- matrix(c(y, rep(0, length(y))), length(y), 2)\r\nhistory <-\r\n  model %>% fit(x,\r\n                train_Y,\r\n                batch_size = 100,\r\n                epochs = 100)\r\nplot(history)\r\n<\/code><\/pre>\n
\"plot<\/p>\n
# PREDICTION\r\ntest_x <- seq(-2,2,0.05)\r\npred_y <- predict(model, test_x)[,1]\r\npred_sd <- predict(model, test_x)[,2]\r\ndatPred <- data.frame(x=test_x, y=pred_y,\r\n                      y1=pred_y+exp(pred_sd), \r\n                      y2=pred_y-exp(pred_sd)) \r\n\r\n# TRUE \r\ntrue_y <- sin(test_x*pi)\r\ntrue_yUpper <- sin(test_x*pi) + 1\/2*(abs(cos(test_x))+1\/8)\r\ntrue_yLower <- sin(test_x*pi) - 1\/2*(abs(cos(test_x))+1\/8)\r\ndatTrue <- data.frame(x=test_x, y=true_y, ymax=true_yUpper, ymin=true_yLower)\r\n\r\nrequire(ggplot2)\r\nggplot() + \r\n  geom_line(data=datPred, aes(x=x, y=y), linetype='dotted') +\r\n  geom_line(data=datPred, aes(x=x, y=y1), linetype='dotted') +\r\n  geom_line(data=datPred, aes(x=x, y=y2), linetype='dotted') +\r\n  geom_line(data=datTrue, aes(x=x, y=y), col='grey20') + \r\n  geom_ribbon(data=datTrue, aes(x=x, ymin=ymin, ymax=ymax), alpha=0.15)\r\n<\/code><\/pre>\n
\"plot<\/p>\n
\uac70\uc758 \uc644\ubcbd\ud558\ub2e4!<\/p>\n
\n
\uc0dd\uac01\ud560 \ubb38\uc81c<\/h3>\n