\n1. \uc804\uc774 \ud559\uc2b5\uc774\ub780?<\/h2>\n
\n \uc804\uc774 \ud559\uc2b5\uc740 \uae30\uacc4 \ud559\uc2b5\uacfc \ub525\ub7ec\ub2dd \ubd84\uc57c\uc5d0\uc11c \uc774\ubbf8 \ud559\uc2b5\ub41c \ubaa8\ub378\uc744 \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc5d0 \ud65c\uc6a9\ud558\uc5ec \ud559\uc2b5\uc744 \uac00\uc18d\ud654\ud558\ub294 \ubc29\ubc95\ub860\uc785\ub2c8\ub2e4.
\n \uc774\ub294 \ubaa8\ub378\uc774 \uc774\uc804\uc5d0 \ud559\uc2b5\ud55c \uc815\ubcf4\ub97c \ud65c\uc6a9\ud558\uc5ec \uc0c8\ub85c\uc6b4 \ud658\uacbd\uc5d0\uc11c \ub354 \ube60\ub974\uace0 \ud6a8\uc728\uc801\uc73c\ub85c \ud559\uc2b5\ud560 \uc218 \uc788\ub3c4\ub85d \ud569\ub2c8\ub2e4.
\n \uc8fc\ub85c \ud6c8\ub828 \ub370\uc774\ud130\uac00 \ubd80\uc871\ud558\uac70\ub098 \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc774 \uae30\uc874 \uc791\uc5c5\uacfc \uc720\uc0ac\ud560 \ub54c \ud6a8\uacfc\uc801\uc785\ub2c8\ub2e4.\n <\/p>\n
\uc608\ub97c \ub4e4\uc5b4, \uc774\ubbf8\uc9c0 \ubd84\ub958 \ubaa8\ub378\uc758 \uacbd\uc6b0, \uc22b\uc790 \uc778\uc2dd\uc744 \uc704\ud55c \ubaa8\ub378\uc774 \ud2b9\uc815\ud55c \uc885\ub958\uc758 \uc774\ubbf8\uc9c0 \uc778\uc2dd\uc5d0 \uc5bc\ub9c8\ub098 \ube60\ub974\uac8c \uc801\uc751\ud560 \uc218 \uc788\ub294\uc9c0\ub97c \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n<\/section>\n\n2. \uc804\uc774 \ud559\uc2b5\uc758 \ud544\uc694\uc131<\/h2>\n
\n \uae30\ubcf8\uc801\uc778 \uba38\uc2e0\ub7ec\ub2dd \ubaa8\ub378\uc744 \ucc98\uc74c\ubd80\ud130 \ub05d\uae4c\uc9c0 \ud559\uc2b5\uc2dc\ud0a4\ub294 \ub370\uc5d0\ub294 \ub9ce\uc740 \uc2dc\uac04\uacfc \uc790\uc6d0\uc774 \ud544\uc694\ud569\ub2c8\ub2e4.
\n \ud2b9\ud788 \ub300\uaddc\ubaa8 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\ub294 \uacbd\uc6b0, \ubaa8\ub378 \ud559\uc2b5\uc5d0 \ud544\uc694\ud55c \uacc4\uc0b0\ub7c9\uc740 \uae30\ud558\uae09\uc218\uc801\uc73c\ub85c \uc99d\uac00\ud569\ub2c8\ub2e4.
\n \uc804\uc774 \ud559\uc2b5\uc740 \uc774\ub7ec\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uba70, \uae30\uc874\uc5d0 \uc774\ubbf8 \ud559\uc2b5\ub41c \ubaa8\ub378(\uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378)\uc744 \ud65c\uc6a9\ud558\uc5ec \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\uc14b\uc5d0 \ub9de\uac8c \uc218\uc815\ud558\ub294 \ubc29\ubc95\uc73c\ub85c, \ub450 \uac00\uc9c0 \uc911\uc694\ud55c \uc7a5\uc810\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4.\n <\/p>\n
\n- \ud559\uc2b5 \uc2dc\uac04 \ub2e8\ucd95:<\/strong> \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378\uc744 \uae30\ubc18\uc73c\ub85c \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc744 \uc218\ud589\ud568\uc73c\ub85c\uc368 \ud559\uc2b5 \uc2dc\uac04\uc774 \ub300\ud3ed \uc904\uc5b4\ub4ed\ub2c8\ub2e4.<\/li>\n
- \ub370\uc774\ud130 \ud6a8\uc728\uc131:<\/strong> \ub370\uc774\ud130 \uc591\uc774 \uc801\uc740 \uacbd\uc6b0\uc5d0\ub3c4 \ud6a8\uacfc\uc801\uc73c\ub85c \uc131\ub2a5 \ud5a5\uc0c1\uc774 \uac00\ub2a5\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n<\/section>\n\n
3. \uc804\uc774 \ud559\uc2b5\uc758 \uc811\uadfc \ubc29\uc2dd<\/h2>\n
\n \uc77c\ubc18\uc801\uc73c\ub85c \uc804\uc774 \ud559\uc2b5\uc5d0\uc11c\ub294 \ub450 \uac00\uc9c0 \uc811\uadfc \ubc29\ubc95\uc774 \uc788\uc2b5\ub2c8\ub2e4:\n <\/p>\n
\n- \uace0\uc815\ub41c \ubca0\uc774\uc2a4:<\/strong> \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130\ub97c \ub3d9\uacb0(freeze)\ud558\uace0, \uc0c8\ub86d\uac8c \ucd94\uac00\ub41c \ub808\uc774\uc5b4\ub9cc \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/li>\n
- \ubbf8\uc138 \uc870\uc815:<\/strong> \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378\uc758 \ud2b9\uc815 \ub808\uc774\uc5b4\ub97c \ubbf8\uc138 \uc870\uc815(fine-tuning)\ud558\uc5ec \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc5d0 \ub9de\uac8c \uc870\uc815\ud569\ub2c8\ub2e4.<\/li>\n<\/ol>\n
\ubaa8\ub378\uc5d0\uc11c \uc5b4\ub5a4 \uc811\uadfc \ubc29\uc2dd\uc744 \uc0ac\uc6a9\ud560\uc9c0\ub294 \ub370\uc774\ud130\uc14b\uc758 \ud06c\uae30\uc640 \uc720\uc0ac\uc131\uc5d0 \ub530\ub77c \uacb0\uc815\ub429\ub2c8\ub2e4.<\/p>\n<\/section>\n\n4. \uc804\uc774 \ud559\uc2b5\uc758 \uc2e4\uc81c \uc0c1\ud669<\/h2>\n
\n \ub2e4\uc74c\uc740 \uc804\uc774 \ud559\uc2b5\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc774\ubbf8\uc9c0 \ubd84\ub958 \ubaa8\ub378\uc744 \ud559\uc2b5\ud558\ub294 \uc608\uc81c\uc785\ub2c8\ub2e4.
\n TensorFlow\uc640 Keras \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc804\uc774 \ud559\uc2b5\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc744 \uc18c\uac1c\ud558\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
4.1. \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
\n \uba3c\uc800, \uc0ac\uc6a9\ud560 \ub370\uc774\ud130\uc14b\uc744 \uc900\ube44\ud569\ub2c8\ub2e4. CIFAR-10 \ub370\uc774\ud130\uc14b\uacfc \uac19\uc740 \uac04\ub2e8\ud55c \uc774\ubbf8\uc9c0 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.
\n \uc774 \ub370\uc774\ud130\uc14b\uc740 10\uac1c\uc758 \ud074\ub798\uc2a4\uac00 \ud3ec\ud568\ub41c 60,000\uac1c\uc758 32×32 \uceec\ub7ec \uc774\ubbf8\uc9c0\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4.\n <\/p>\n
4.2. \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378 \uc120\ud0dd<\/h3>\n
\n TensorFlow\uc758 Keras API\uc5d0\uc11c\ub294 \uc5ec\ub7ec \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 VGG16 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4.
\n \uc774 \ubaa8\ub378\uc740 \ub300\uaddc\ubaa8 \uc774\ubbf8\uc9c0\ub137 \ub370\uc774\ud130\uc14b\uc5d0\uc11c \ud559\uc2b5\ub418\uc5c8\uc2b5\ub2c8\ub2e4.\n <\/p>\n
4.3. \ubaa8\ub378 \uad6c\ud604<\/h3>\n
\uc608\uc81c \ucf54\ub4dc:<\/strong><\/p>\n\nimport tensorflow as tf\nfrom tensorflow.keras import layers, models\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\n\n# \ub370\uc774\ud130\uc14b \ub85c\ub4dc \ubc0f \uc804\ucc98\ub9ac\n(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()\nx_train, x_test = x_train \/ 255.0, x_test \/ 255.0\n\n# \ub370\uc774\ud130 \uc99d\uac15\ndatagen = ImageDataGenerator(rotation_range=20, width_shift_range=0.2,\n height_shift_range=0.2, shear_range=0.2,\n zoom_range=0.2, horizontal_flip=True,\n fill_mode='nearest')\n\n# \uc0ac\uc804 \ud559\uc2b5\ub41c VGG16 \ubaa8\ub378 \ub85c\ub4dc (\uc0c1\ub2e8 \ub808\uc774\uc5b4 \uc81c\uc678)\nbase_model = tf.keras.applications.VGG16(include_top=False, weights='imagenet',\n input_shape=(32, 32, 3))\n\n# \uc0c8\ub86d\uac8c \ucd94\uac00\ud560 \ub808\uc774\uc5b4 \uad6c\uc131\nmodel = models.Sequential()\nmodel.add(base_model)\nmodel.add(layers.Flatten())\nmodel.add(layers.Dense(256, activation='relu'))\nmodel.add(layers.Dense(10, activation='softmax'))\n\n# \ubaa8\ub378 \ucef4\ud30c\uc77c\nmodel.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\n\n# \ubaa8\ub378 \ud6c8\ub828\nmodel.fit(datagen.flow(x_train, y_train, batch_size=32), epochs=10)\n <\/code><\/pre>\n<\/section>\n\n5. \ubbf8\uc138 \uc870\uc815(Fine-Tuning)<\/h2>\n
\n \ubaa8\ub378\uc758 \uc0c1\uc704 \ub808\uc774\uc5b4\ub97c \uace0\uc815\ud55c \ud6c4, \ud2b9\uc815 \ub808\uc774\uc5b4\uc758 \ubbf8\uc138 \uc870\uc815\uc744 \ud1b5\ud574 \uc131\ub2a5\uc744 \ub354\uc6b1 \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4.
\n \uc608\ub97c \ub4e4\uc5b4, VGG16\uc758 \ub9c8\uc9c0\ub9c9 \uba87 \uac1c\uc758 \ub808\uc774\uc5b4\ub9cc \ud559\uc2b5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\uc608\uc81c \ucf54\ub4dc:<\/strong><\/p>\n\n# \ud2b9\uc815 \ub808\uc774\uc5b4\uc758 \uac00\uc911\uce58 \ub3d9\uacb0\nfor layer in base_model.layers[:-4]:\n layer.trainable = False\n\n# \ubaa8\ub378 \uc7ac\ucef4\ud30c\uc77c\nmodel.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\n\n# \ubaa8\ub378 \uc7ac\ud6c8\ub828\nmodel.fit(datagen.flow(x_train, y_train, batch_size=32), epochs=10)\n <\/code><\/pre>\n<\/section>\n\n6. \uc804\uc774 \ud559\uc2b5\uc758 \uc801\uc6a9 \ubc94\uc704<\/h2>\n
\n \uc804\uc774 \ud559\uc2b5\uc740 \uc774\ubbf8\uc9c0 \ubd84\ub958 \uc678\uc5d0\ub3c4 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0 \uc801\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubd84\uc57c\uc5d0\uc11c\ub294
\n BERT\ub098 GPT\uc640 \uac19\uc740 \uc0ac\uc804 \ud559\uc2b5\ub41c \uc5b8\uc5b4 \ubaa8\ub378\uc744 \ud2b9\uc815\ud55c \ud14d\uc2a4\ud2b8 \ubd84\ub958 \uacfc\uc81c\uc5d0 \uc801\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \uac01\uac01\uc758 \uc791\uc5c5\uc5d0 \ub9de\uac8c \ud2b9\uc815 \ub808\uc774\uc5b4\ub97c \ub3d9\uacb0\ud558\uace0 \ubbf8\uc138 \uc870\uc815\ud558\ub294 \ubc29\uc2dd\uc73c\ub85c \uc0ac\uc6a9\ud558\ub294 \ub370\uc774\ud130\uc14b\uc5d0 \ucd5c\uc801\ud654\ub41c \uc131\ub2a5\uc744 \ub04c\uc5b4\ub0bc \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n<\/section>\n\n7. \uc804\uc774 \ud559\uc2b5 \uc2dc \uace0\ub824\uc0ac\ud56d<\/h2>\n
\n \uc804\uc774 \ud559\uc2b5\uc744 \uc9c4\ud589\ud560 \ub54c, \uba87 \uac00\uc9c0 \uace0\ub824\ud574\uc57c \ud560 \uc810\uc774 \uc788\uc2b5\ub2c8\ub2e4:\n <\/p>\n
\n- \ub370\uc774\ud130\uc14b \uc720\uc0ac\uc131:<\/strong> \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378\uacfc \uc0c8\ub85c\uc6b4 \uc791\uc5c5 \uac04\uc758 \ub370\uc774\ud130\uc14b \uc720\uc0ac\uc131\uc774 \ub192\uc744\uc218\ub85d \uc88b\uc740 \uc131\ub2a5\uc744 \uae30\ub300\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n
- \ud30c\ub77c\ubbf8\ud130 \uc870\uc815:<\/strong> \ud559\uc2b5\ub960 \ubc0f \ub808\uc774\uc5b4 \uc218\uc640 \uac19\uc740 \ud30c\ub77c\ubbf8\ud130 \uc870\uc815\uc774 \ud544\uc694\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubbf8\uc138 \uc870\uc815\uc758 \uacbd\uc6b0 \ud559\uc2b5\ub960\uc744 \ub0ae\ucd94\ub294 \uac83\uc774 \uc77c\ubc18\uc801\uc785\ub2c8\ub2e4.<\/li>\n<\/ul>\n<\/section>\n