\uc81c\ud55c\ub41c \ubcfc\uce20\ub9cc \uba38\uc2e0(Restricted Boltzmann Machine, RBM)\uc740 \ube44\uc9c0\ub3c4 \ud559\uc2b5 \uc54c\uace0\ub9ac\uc998\uc758 \uc77c\uc885\uc73c\ub85c, \uc0dd\uc131 \ubaa8\ub378\ub85c\ub3c4 \uc54c\ub824\uc838 \uc788\uc2b5\ub2c8\ub2e4. RBM\uc740 \ub2e4\ub7c9\uc758 \uc785\ub825 \ub370\uc774\ud130\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \ud559\uc2b5\ud560 \uc218 \uc788\uc73c\uba70, \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud65c\uc6a9\ub429\ub2c8\ub2e4. \ubcf8 \ubb38\uc11c\uc5d0\uc11c\ub294 RBM\uc758 \uae30\ubcf8 \uc6d0\ub9ac, \ud30c\uc774\uc36c\uc73c\ub85c \uad6c\ud604\ud558\ub294 \ubc29\ubc95, \uadf8\ub9ac\uace0 PyTorch \ud504\ub808\uc784\uc6cc\ud06c\ub97c \ud65c\uc6a9\ud55c \uc608\uc81c\ub97c \ud1b5\ud574 \uae4a\uc774 \uc788\ub294 \uc774\ud574\ub97c \ub3d5\uace0\uc790 \ud569\ub2c8\ub2e4.<\/p>\n
RBM\uc740 \ud1b5\uacc4 \ubb3c\ub9ac\ud559\uc5d0\uc11c \uc720\ub798\ub41c \ubaa8\ub378\ub85c, ‘\ubcfc\uce20\ub9cc \uba38\uc2e0’\uc758 \uac1c\ub150\uc744 \uae30\ubc18\uc73c\ub85c \ud569\ub2c8\ub2e4. RBM\uc740 \ub450 \uc885\ub958\uc758 \ub178\ub4dc\ub85c \uad6c\uc131\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4: \uac00\uc2dc \ub178\ub4dc(visible nodes)\uc640 \uc740\ub2c9 \ub178\ub4dc(hidden nodes). \uc774 \ub450 \ub178\ub4dc \uc0ac\uc774\uc5d0\ub294 \uc5f0\uacb0\uc774 \uc788\uc9c0\ub9cc, \uc740\ub2c9 \ub178\ub4dc\ub07c\ub9ac\ub294 \uc5f0\uacb0\uc774 \uc5c6\ub294 \uc81c\ud55c\ub41c \uad6c\uc870\ub97c \uac00\uc9d1\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uad6c\uc870 \ub355\ubd84\uc5d0 RBM\uc740 \ub354\uc6b1 \ud6a8\uc728\uc801\uc778 \ud559\uc2b5\uc774 \uac00\ub2a5\ud569\ub2c8\ub2e4.<\/p>\n
RBM\uc740 \ub2e4\uc74c\uacfc \uac19\uc740 \uad6c\uc131 \uc694\uc18c\ub97c \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
RBM\uc758 \ud559\uc2b5\uc740 \uc5d0\ub108\uc9c0 \ud568\uc218(Energy Function)\ub97c \ucd5c\uc18c\ud654\ud558\ub294 \uacfc\uc815\uc744 \ud1b5\ud574 \uc774\ub8e8\uc5b4\uc9d1\ub2c8\ub2e4. \uc5d0\ub108\uc9c0 \ud568\uc218\ub294 \uac00\uc2dc \ub178\ub4dc\uc640 \uc740\ub2c9 \ub178\ub4dc\uc758 \uc0c1\ud0dc\uc5d0 \ub530\ub77c \ub2e4\uc74c\uacfc \uac19\uc774 \uc815\uc758\ub429\ub2c8\ub2e4:<\/p>\n
E(v, h) = -\u2211 vi<\/sub> * bi<\/sub> - \u2211 hj<\/sub> * cj<\/sub> - \u2211 vi<\/sub> * hj<\/sub> * wij<\/sub><\/code><\/pre>\n\uc5ec\uae30\uc11c, \\( v \\)\ub294 \uac00\uc2dc \ub178\ub4dc, \\( h \\)\ub294 \uc740\ub2c9 \ub178\ub4dc, \\( b \\)\ub294 \uac00\uc2dc \ub178\ub4dc\uc758 \ubc14\uc774\uc5b4\uc2a4, \\( c \\)\ub294 \uc740\ub2c9 \ub178\ub4dc\uc758 \ubc14\uc774\uc5b4\uc2a4, \\( w \\)\ub294 \uac00\uc911\uce58\uc785\ub2c8\ub2e4.<\/p>\n
2. \uc81c\ud55c\ub41c \ubcfc\uce20\ub9cc \uba38\uc2e0\uc758 \ud559\uc2b5 \uacfc\uc815<\/h2>\n
RBM\uc758 \ud559\uc2b5 \uacfc\uc815\uc740 \ub2e4\uc74c\uacfc \uac19\uc774 \uc9c4\ud589\ub429\ub2c8\ub2e4:<\/p>\n
\n- \ub370\uc774\ud130\uc14b\uc5d0\uc11c \uac00\uc2dc \ub178\ub4dc\ub97c \ucd08\uae30\ud654\ud569\ub2c8\ub2e4.<\/li>\n
- \uc740\ub2c9 \ub178\ub4dc\uc758 \ud655\ub960\uc744 \uacc4\uc0b0\ud569\ub2c8\ub2e4.<\/li>\n
- \uc740\ub2c9 \ub178\ub4dc\ub97c \uc0d8\ud50c\ub9c1\ud569\ub2c8\ub2e4.<\/li>\n
- \uac00\uc2dc \ub178\ub4dc\uc758 \uc7ac\uad6c\uc131\uc744 \ud1b5\ud574 \uc0c8\ub85c\uc6b4 \uac00\uc2dc \ub178\ub4dc\uc758 \ud655\ub960\uc744 \uacc4\uc0b0\ud569\ub2c8\ub2e4.<\/li>\n
- \uc740\ub2c9 \ub178\ub4dc\uc758 \uc7ac\uad6c\uc131\uc744 \ud1b5\ud574 \uc0c8\ub85c\uc6b4 \uc740\ub2c9 \ub178\ub4dc\uc758 \ud655\ub960\uc744 \uacc4\uc0b0\ud569\ub2c8\ub2e4.<\/li>\n
- \uac00\uc911\uce58\uc640 \ubc14\uc774\uc5b4\uc2a4\ub97c \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n
2.1 Contrastive Divergence \uc54c\uace0\ub9ac\uc998<\/h3>\n
RBM\uc758 \ud559\uc2b5\uc740 Contrastive Divergence(CD) \uc54c\uace0\ub9ac\uc998\uc744 \ud1b5\ud574 \uc774\ub8e8\uc5b4\uc9d1\ub2c8\ub2e4. CD\ub294 \ub450 \uac00\uc9c0 \uc8fc\uc694 \ub2e8\uacc4\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4:<\/p>\n
\n- \uc704\uc0c1 \uc804\uc774 \ub2e8\uacc4(Positive Phase)<\/strong>: \uc785\ub825 \ub370\uc774\ud130\ub85c\ubd80\ud130 \uc740\ub2c9 \ub178\ub4dc\uc758 \ud65c\uc131\ub3c4\ub97c \ud30c\uc545\ud558\uace0, \uc774 \uac12\uc744 \ubc14\ud0d5\uc73c\ub85c \uac00\uc911\uce58\ub97c \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/li>\n
- \ubd80\uc815\uc801 \uc704\uc0c1(Pseudo Negative Phase)<\/strong>: \uc0d8\ud50c\ub9c1\ub41c \uc740\ub2c9 \ub178\ub4dc\ub85c\ubd80\ud130 \uac00\uc2dc \ub178\ub4dc\ub97c \uc7ac\uad6c\uc131\ud55c \ud6c4, \ub2e4\uc2dc \uc740\ub2c9 \ub178\ub4dc\ub97c \uc0d8\ud50c\ub9c1\ud558\uc5ec \uc720\uc0ac\uc131\uc744 \uac10\uc18c\uc2dc\ud0a4\ub294 \ubc29\ud5a5\uc73c\ub85c \uac00\uc911\uce58\ub97c \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/li>\n<\/ol>\n
3. PyTorch\ub85c RBM \uad6c\ud604\ud558\uae30<\/h2>\n
\uc774\ubc88 \uc7a5\uc5d0\uc11c\ub294 PyTorch\ub97c \uc774\uc6a9\ud558\uc5ec RBM\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc744 \uc124\uba85\ud569\ub2c8\ub2e4. \uc6b0\uc120, \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uace0 \ub370\uc774\ud130\uc14b\uc744 \uc900\ube44\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
3.1 \ub77c\uc774\ube0c\ub7ec\ub9ac \uc124\uce58 \ubc0f \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n!pip install torch torchvision<\/code><\/pre>\nMNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec RBM\uc758 \ud559\uc2b5\uc744 \uc9c4\ud589\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \ub370\uc774\ud130\uc14b\uc740 \uc190\uae00\uc528 \uc22b\uc790 \uc774\ubbf8\uc9c0\ub85c \uad6c\uc131\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
import torch\nfrom torchvision import datasets, transforms\n\n# MNIST \ub370\uc774\ud130\uc14b \ub2e4\uc6b4\ub85c\ub4dc \ubc0f \ubcc0\ud658\ntransform = transforms.Compose([transforms.ToTensor(), transforms.Lambda(lambda x: x.view(-1))])\nmnist = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ntrain_loader = torch.utils.data.DataLoader(dataset=mnist, batch_size=64, shuffle=True)<\/code><\/pre>\n3.2 RBM \ud074\ub798\uc2a4 \uc815\uc758<\/h3>\n
\uc774\uc81c RBM \ud074\ub798\uc2a4\ub97c \uc815\uc758\ud569\uc2dc\ub2e4. \ud074\ub798\uc2a4\ub294 \uac00\uc911\uce58 \ucd08\uae30\ud654, \uac00\uc911\uce58 \uc5c5\ub370\uc774\ud2b8, \ud559\uc2b5 \ub4f1\uc758 \uba54\uc18c\ub4dc\ub97c \ud3ec\ud568\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
class RBM:\n def __init__(self, visible_units, hidden_units, learning_rate=0.1):\n self.visible_units = visible_units\n self.hidden_units = hidden_units\n self.learning_rate = learning_rate\n self.weights = torch.randn(visible_units, hidden_units) * 0.1\n self.visible_bias = torch.zeros(visible_units)\n self.hidden_bias = torch.zeros(hidden_units)\n\n def sample_hidden(self, visible):\n activation = torch.mm(visible, self.weights) + self.hidden_bias\n probabilities = torch.sigmoid(activation)\n return probabilities, torch.bernoulli(probabilities)\n\n def sample_visible(self, hidden):\n activation = torch.mm(hidden, self.weights.t()) + self.visible_bias\n probabilities = torch.sigmoid(activation)\n return probabilities, torch.bernoulli(probabilities)\n\n def train(self, train_loader, num_epochs=10):\n for epoch in range(num_epochs):\n for data, _ in train_loader:\n # \uac00\uc2dc \ub178\ub4dc \uc0d8\ud50c\ub9c1\n v0 = data\n h0, h0_sample = self.sample_hidden(v0)\n\n # \ubd80\uc815\uc801 \uc704\uc0c1 \ub2e8\uacc4\n v1, v1_sample = self.sample_visible(h0_sample)\n h1, _ = self.sample_hidden(v1_sample)\n\n # \uac00\uc911\uce58 \uc5c5\ub370\uc774\ud2b8\n self.weights += self.learning_rate * (torch.mm(v0.t(), h0) - torch.mm(v1.t(), h1)) \/ v0.size(0)\n self.visible_bias += self.learning_rate * (v0 - v1).mean(0)\n self.hidden_bias += self.learning_rate * (h0 - h1).mean(0)\n\n print('Epoch: {} - Loss: {:.4f}'.format(epoch, torch.mean((v0 - v1) ** 2).item()))<\/code><\/pre>\n3.3 RBM \ud559\uc2b5 \uc218\ud589<\/h3>\n
\uc774\uc81c \uc815\uc758\ud55c RBM \ud074\ub798\uc2a4\ub97c \uc774\uc6a9\ud558\uc5ec \ubaa8\ub378\uc744 \ud559\uc2b5\uc2dc\ucf1c\ubd05\uc2dc\ub2e4.<\/p>\n
visible_units = 784 # MNIST\uc758 \uacbd\uc6b0 28x28 \ud53d\uc140\nhidden_units = 256 # \uc740\ub2c9 \ub178\ub4dc \uc218\nrbm = RBM(visible_units, hidden_units)\nrbm.train(train_loader, num_epochs=10)<\/code><\/pre>\n4. \uacb0\uacfc \ubc0f \ud574\uc11d<\/h2>\n
\ud559\uc2b5\uc774 \uc9c4\ud589\ub418\uba74 \uac01 epoch\ub9c8\ub2e4 \uc190\uc2e4 \uac12\uc774 \ucd9c\ub825\ub429\ub2c8\ub2e4. \uc190\uc2e4 \uac12\uc740 \uac00\uc2dc \ub178\ub4dc \uc7ac\uad6c\uc131\uc774 \uc740\ub2c9 \uc0c1\ud0dc\uc640 \uc5bc\ub9c8\ub098 \uc720\uc0ac\ud55c\uc9c0\ub97c \ub098\ud0c0\ub0b4\ubbc0\ub85c, \uc190\uc2e4 \uac12\uc758 \uac10\uc18c\ub294 \ubaa8\ub378\uc758 \uc131\ub2a5 \ud5a5\uc0c1\uc744 \uc758\ubbf8\ud569\ub2c8\ub2e4. \uc8fc\ubaa9\ud560 \uc810\uc740 \ubcfc\uce20\ub9cc \uba38\uc2e0\uc774 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc758 \uae30\ucd08\uac00 \ub418\uc5b4 \ub9ce\uc740 \ub2e4\ub978 \uc54c\uace0\ub9ac\uc998\uacfc \uacb0\ud569\ub418\uc5b4 \uc0ac\uc6a9\ub41c\ub2e4\ub294 \uc810\uc785\ub2c8\ub2e4.<\/p>\n
5. \uacb0\ub860<\/h2>\n
\uc774\ubc88 \ud3ec\uc2a4\ud2b8\uc5d0\uc11c\ub294 \uc81c\ud55c\ub41c \ubcfc\uce20\ub9cc \uba38\uc2e0\uc758 \uac1c\ub150\ubd80\ud130 \ud559\uc2b5 \uacfc\uc815, \uadf8\ub9ac\uace0 PyTorch\ub97c \uc0ac\uc6a9\ud55c \uc2e4\uc81c \uad6c\ud604 \uc0ac\ub840\uae4c\uc9c0 \ub2e4\ub8e8\uc5b4 \ubcf4\uc558\uc2b5\ub2c8\ub2e4. RBM\uc740 \ub370\uc774\ud130\uc758 \ub0b4\uc7ac\ub41c \uad6c\uc870\ub97c \ud559\uc2b5\ud558\ub294 \ub370 \ub9e4\uc6b0 \uc720\ud6a8\ud55c \ub3c4\uad6c\uc785\ub2c8\ub2e4. \uadf8\ub7fc\uc5d0\ub3c4 \ubd88\uad6c\ud558\uace0 RBM\uc740 \ud604\uc7ac\uc758 \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\uc5d0\uc11c\ub294 \uc8fc\ub85c \uc0ac\uc804 \ud559\uc2b5(pre-training) \ub610\ub294 \ub2e4\ub978 \uc544\ud0a4\ud14d\ucc98\uc640\uc758 \uacb0\ud569\uc5d0 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \uc55e\uc73c\ub85c \ub2e4\uc591\ud55c \uc0dd\uc131 \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5f0\uad6c\uac00 \uc774\ub8e8\uc5b4\uc9c8 \uac83\uc73c\ub85c \uae30\ub300\ub429\ub2c8\ub2e4.<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"