\ub525\ub7ec\ub2dd\uc740 \uc778\uacf5\uc9c0\ub2a5\uc758 \ud55c \ubd84\uc57c\ub85c, \uc2ec\uce35 \uc2e0\uacbd\ub9dd\uc744 \uc774\uc6a9\ud558\uc5ec \ub370\uc774\ud130\uc5d0\uc11c \ud328\ud134\uc744 \ud559\uc2b5\ud558\uace0 \uc608\uce21\ud558\ub294 \uae30\uc220\uc785\ub2c8\ub2e4. \uc624\ub298\uc740 \uc2ec\uce35 \uc2e0\ub8b0 \uc2e0\uacbd\ub9dd(Deep Belief Network, DBN)<\/strong>\uc744 \uc18c\uac1c\ud558\uace0, \uc774\ub97c \ud30c\uc774\ud1a0\uce58(PyTorch)<\/strong>\ub97c \uc0ac\uc6a9\ud558\uc5ec \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n \uc2ec\uce35 \uc2e0\ub8b0 \uc2e0\uacbd\ub9dd\uc740 \uc5ec\ub7ec \uac1c\uc758 \uce35\uc744 \uac00\uc9c4 \uc778\uacf5 \uc2e0\uacbd\ub9dd\uc73c\ub85c, \ud2b9\ud788 \ub2e4\uc74c\uacfc \uac19\uc740 \ud2b9\uc9d5\uc774 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n DBN\uc740 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc5d0\uc11c \uc911\uc694\ud55c \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 \uc785\ub825 \ub370\uc774\ud130\ub97c \uc5ec\ub7ec \uce35\uc758 \ud655\ub960 \ubd84\ud3ec\ub85c \ud45c\ud604\ud558\uc5ec \ubcf5\uc7a1\ud55c \ud2b9\uc131\uc744 \ud559\uc2b5\ud560 \uc218 \uc788\uac8c \ud569\ub2c8\ub2e4.<\/p>\n Restricted Boltzmann Machine(RBM)\uc740 \ube44\uc9c0\ub3c4 \ud559\uc2b5\uc5d0 \uc0ac\uc6a9\ub418\ub294 \ud655\ub960 \ubaa8\ub378\ub85c, \ub450 \uac1c\uc758 \uce35\uc73c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4:<\/p>\n RBM\uc740 \uac01\uac01\uc758 \uce35\uc758 \ub274\ub7f0 \uc0ac\uc774\uc5d0 \uc5f0\uacb0\uc774 \uc788\uc73c\uba70, \uc774\ub7ec\ud55c \uc5f0\uacb0\uc740 \uc0dd\uc874 \ud655\ub960\uc744 \uae30\ubc18\uc73c\ub85c \ud55c \ud655\ub960 \ubd84\ud3ec\ub97c \ud559\uc2b5\ud558\uac8c \ub429\ub2c8\ub2e4.<\/p>\n \uc774\uc81c \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc2ec\uce35 \uc2e0\ub8b0 \uc2e0\uacbd\ub9dd\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 \uac04\ub2e8\ud55c MNIST \uc22b\uc790 \uc778\uc2dd \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec DBN\uc744 \uad6c\ucd95\ud569\ub2c8\ub2e4.<\/p>\n \uba3c\uc800, MNIST \ub370\uc774\ud130\uc14b\uc744 \ubd88\ub7ec\uc624\uace0 \uc804\ucc98\ub9ac\ud569\ub2c8\ub2e4.<\/p>\n DBN\uc740 \uc5ec\ub7ec \uac1c\uc758 RBM\uc744 \uc313\uc544\uc11c \ub9cc\ub4e4 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798\ub294 \ud30c\uc774\ud1a0\uce58\ub85c DBN\uc744 \uad6c\ud604\ud55c \uc608\uc81c\uc785\ub2c8\ub2e4.<\/p>\n \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \uc2ec\uce35 \uc2e0\ub8b0 \uc2e0\uacbd\ub9dd\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \uc6d0\ub9ac, \uadf8\ub9ac\uace0 \uc774\ub97c \ud30c\uc774\ud1a0\uce58\ub85c \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. DBN\uc740 \ubcf5\uc7a1\ud55c \ub370\uc774\ud130\uc758 \uc7a0\uc7ac\uc801 \uad6c\uc870\ub97c \ud559\uc2b5\ud558\ub294 \ub370 \ub9e4\uc6b0 \uc720\uc6a9\ud55c \ubaa8\ub378\uc785\ub2c8\ub2e4. \ud30c\uc774\ud1a0\uce58\ub97c \uc774\uc6a9\ud558\uba74 \uc774\ub7ec\ud55c \ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \ud6a8\uacfc\uc801\uc73c\ub85c \uad6c\ud604\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \ub354 \uae4a\uc740 \ud559\uc2b5 \ubc0f \ud65c\uc6a9\uc744 \uc704\ud574 \ud30c\uc774\ud1a0\uce58 \uacf5\uc2dd \ubb38\uc11c\uc640 \ub2e4\uc591\ud55c \uc608\uc81c\ub4e4\uc744 \ucc38\uc870\ud558\uc2dc\uae30\ub97c \uad8c\uc7a5\ud569\ub2c8\ub2e4. \ub525\ub7ec\ub2dd\uc758 \uc138\uacc4\uc5d0 \uc624\uc2e0 \uac83\uc744 \ud658\uc601\ud569\ub2c8\ub2e4!<\/p>\n <\/body><\/p>\n","protected":false},"excerpt":{"rendered":"1. \uc2ec\uce35 \uc2e0\ub8b0 \uc2e0\uacbd\ub9dd(Deep Belief Network)\ub780?<\/h2>\n
\n
1.1 Restricted Boltzmann Machine<\/h3>\n
\n
2. \ud30c\uc774\ud1a0\uce58\ub97c \uc774\uc6a9\ud55c DBN \uad6c\ud604<\/h2>\n
2.1 \ub370\uc774\ud130\uc14b \ubd88\ub7ec\uc624\uae30<\/h3>\n
import torch\nfrom torchvision import datasets, transforms\n\n# \ub370\uc774\ud130 \ubcc0\ud658 \uc815\uc758\ntransform = transforms.Compose([\n transforms.ToTensor(),\n transforms.Normalize((0.5,), (0.5,))\n])\n\n# MNIST \ub370\uc774\ud130\uc14b \ub2e4\uc6b4\ub85c\ub4dc\ntrain_dataset = datasets.MNIST(root='.\/data', train=True, transform=transform, download=True)\ntest_dataset = datasets.MNIST(root='.\/data', train=False, transform=transform, download=True)\n\n# \ub370\uc774\ud130 \ub85c\ub354 \uc815\uc758\ntrain_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=64, shuffle=True)\ntest_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=64, shuffle=False)<\/code><\/pre>\n2.2 DBN \uad6c\ud604<\/h3>\n
class RBM:\n def __init__(self, n_visible, n_hidden):\n self.W = torch.randn(n_hidden, n_visible) * 0.1\n self.h_bias = torch.zeros(n_hidden)\n self.v_bias = torch.zeros(n_visible)\n\n def sample_h(self, v):\n h_prob = torch.sigmoid(torch.matmul(self.W, v.t()) + self.h_bias.unsqueeze(1))\n return h_prob, torch.bernoulli(h_prob)\n\n def sample_v(self, h):\n v_prob = torch.sigmoid(torch.matmul(h, self.W) + self.v_bias)\n return v_prob, torch.bernoulli(v_prob)\n\n def train(self, data, lr=0.1, k=1):\n for epoch in range(k):\n v0 = data\n h0, h0_sample = self.sample_h(v0)\n v1, v1_sample = self.sample_v(h0_sample)\n h1, _ = self.sample_h(v1_sample)\n\n # \uc5c5\ub370\uc774\ud2b8\n self.W += lr * (torch.matmul(h0_sample.t(), v0) - torch.matmul(h1.t(), v1_sample)) \/ data.size(0)\n self.h_bias += lr * (h0_sample.mean(0) - h1.mean(0))\n self.v_bias += lr * (v0.mean(0) - v1.mean(0))<\/code><\/pre>\n2.3 \uc5ec\ub7ec \uac1c\uc758 RBM\uc744 \uc313\uc544\uc11c DBN \ub9cc\ub4e4\uae30<\/h3>\n
class DBN:\n def __init__(self, layer_sizes):\n self.RBMs = []\n for i in range(len(layer_sizes) - 1):\n self.RBMs.append(RBM(layer_sizes[i], layer_sizes[i + 1]))\n\n def fit(self, data, lr=0.1, k=1):\n for rbm in self.RBMs:\n rbm.train(data, lr=lr, k=k)\n data, _ = rbm.sample_h(data)\n\n def transform(self, data):\n for rbm in self.RBMs:\n _, data = rbm.sample_h(data)\n return data<\/code><\/pre>\n2.4 DBN \ubaa8\ub378 \ud6c8\ub828<\/h3>\n
# DBN \ud6c8\ub828\ndbn = DBN(layer_sizes=[784, 256, 128])\nfor batch_idx, (data, target) in enumerate(train_loader):\n dbn.fit(data.view(-1, 784), lr=0.1, k=10) # 10\ubc88\uc758 K-\ud559\uc2b5 \uc218\ud589<\/code><\/pre>\n2.5 \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc14b\uc5d0 \ub300\ud55c \ubcc0\ud658\uacfc \ud3c9\uac00<\/h3>\n
test_data = next(iter(test_loader))[0].view(-1, 784)\ntransformed_data = dbn.transform(test_data)\nprint(transformed_data)\n# \uc5ec\uae30\uc11c transformed_data\ub97c \ud6c4\uc18d \ubaa8\ub378\uc5d0 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/code><\/pre>\n3. \uacb0\ub860<\/h2>\n