{"id":191,"date":"2020-06-16T19:09:12","date_gmt":"2020-06-16T10:09:12","guid":{"rendered":"http:\/\/cedartrees.co.kr\/?p=191"},"modified":"2021-04-03T19:21:02","modified_gmt":"2021-04-03T10:21:02","slug":"pytorchs-embedding","status":"publish","type":"post","link":"http:\/\/blog.cedartrees.co.kr\/index.php\/2020\/06\/16\/pytorchs-embedding\/","title":{"rendered":"PyTorch’s Embedding()"},"content":{"rendered":"\n

\ub2e8\uc5b4 \uc784\ubca0\ub529(Word Embedding)\uc774\ub780 \ub9d0\ubb49\uce58\uc758 \uac01 \ub2e8\uc5b4\uc5d0 \uc77c\ub300\uc77c\ub85c \ub300\uc751\ud558\ub294 \uc2e4\uc218 \ubca1\ud130\uc758 \uc9d1\ud569\uc774\ub098 \ud639\uc740 \uc774\ub7f0 \uc9d1\ud569\uc744 \uad6c\ud558\ub294 \ud589\uc704\ub97c Word Embedding\uc774\ub77c\uace0 \ud569\ub2c8\ub2e4. Word2Vec\ub3c4 \uc774\ub7f0 \uc6cc\ub4dc \uc784\ubca0\ub529\uc758 \ud55c \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/p>\n\n\n\n

\uadf8\ub807\ub2e4\uba74 \uc65c \uc774\ub7f0 \uc6cc\ub4dc \uc784\ubca0\ub529 \ubc29\ubc95\uc774 \ud544\uc694\ud560\uae4c\uc694? \uadf8 \uc774\uc720\ub294 \ucef4\ud4e8\ud130\uac00 \uc790\uc5f0\uc5b4\ub97c \uc774\ud574\ud558\uc9c0 \ubabb\ud558\uae30 \ub54c\ubb38\uc785\ub2c8\ub2e4. \uadf8\ub807\ub2e4\uba74 \ucef4\ud4e8\ud130\uac00 \uc774\ud574\ud560 \uc218 \uc788\ub294 \ud615\ud0dc, \uc989 \uc22b\uc790\ub85c \ub2e8\uc5b4\ub97c \ubc14\uafd4\uc11c \uc785\ub825\ud574\uc918\uc57c \ud569\ub2c8\ub2e4. \uadf8\ub807\ub2e4\uba74 \uc5b4\ub5bb\uac8c \ud574\uc57c \ud6a8\uacfc\uc801\uc73c\ub85c \ub2e8\uc5b4\ub97c \uc22b\uc790\uc758 \ud615\ud0dc\ub85c \ubc14\uafc0 \uc218 \uc788\uc744\uae4c\uc694? \uc774\ub7ec\ud55c \uace0\ubbfc\uc5d0\uc11c \ub098\uc628 \uac83\uc774 \uc6cc\ub4dc \uc784\ubca0\ub529\uc785\ub2c8\ub2e4.<\/p>\n\n\n\n

\uba3c\uc800 \uc6cc\ub4dc\ub97c \uc22b\uc790\ub85c \ubc14\uafb8\ub294 \uac00\uc7a5 \uac04\ub2e8\ud55c \ubc29\ubc95\uc740 \uc6d0-\ud56b-\uc778\ucf54\ub529(One-Hot-Encoding)\uc785\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4 “\ub098\ub294 \ud559\uad50\uc5d0 \uac11\ub2c8\ub2e4” \uc774 \ubb38\uc7a5\uc744 3\uac1c\uc758 \ub2e8\uc5b4\ub85c \uad6c\ubd84\ud558\uace0 \uac01 \ub2e8\uc5b4\uc758 \uc704\uce58\ub97c \ud45c\uc2dc\ud558\ub294 \uac83\uc774\uc8e0. \uc774\ub807\uac8c \ud558\uba74 \uc77c\ub2e8 \ubb38\uc790\ub97c \uc22b\uc790\ub85c \ubc14\uafb8\ub294\ub370\ub294 \uc131\uacf5\ud588\uc2b5\ub2c8\ub2e4. <\/p>\n\n\n\n

1<\/td>\ub098\ub294<\/td>1,0,0<\/td><\/tr>
2<\/td>\ud559\uad50\uc5d0<\/td>0,1,0<\/td><\/tr>
3<\/td>\uac11\ub2c8\ub2e4<\/td>0,0,1<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

\uadf8\ub7ec\ub098 \uc774\ub7ec\ud55c \ubc29\ubc95\uc5d0\ub294 \ub2e8\uc810\uc774 \uc788\uc2b5\ub2c8\ub2e4. \uac00\uc7a5 \ud070 \ub2e8\uc810\uc740 \ubca1\ud130\uc758 \uc0ac\uc774\uc988\uac00 \ub108\ubb34 \ucee4\uc9c4\ub2e4\ub294 \uac83\uacfc \ubca1\ud130\uc758 \ub0b4\uc6a9\uc774 \ud558\ub098\uc758 1\uc744 \uc81c\uc678\ud55c \ub098\uba38\uc9c0 \ub0b4\uc6a9\uc774 \ubaa8\ub450 0\uc73c\ub85c \ucc44\uc6cc\uc9c4\ub2e4\ub294 \uac83\uc785\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4 “\ub098\ub294 \ud559\uad50\uc5d0 \uac11\ub2c8\ub2e4” 3\uac1c\uc758 \ub2e8\uc5b4\uc774\uc9c0\ub9cc \ucc45\uacfc \uac19\uc740 \ub300\uaddc\ubaa8\uc758 \ub9d0\ubb49\uce58\uc5d0 \ub4f1\uc7a5\ud558\ub294 \ub2e8\uc5b4\ub294 \uc218\ub9cc\uac1c\uac00 \ub41c\ub2e4\ub294 \uac83\uc774\uc8e0. \uadf8\ub807\uac8c \ub418\uba74 \ubca1\ud130\uc758 \ud06c\uae30\ub294 \uc218\ub9cc\uac1c\uac00 \ub118\ub294 \uc0ac\uc774\uc988\uc5d0 \ub300\ubd80\ubd84 0\uc778 \ubca1\ud130\uac00 \ub9cc\ub4e4\uc5b4\uc9c0\uae30 \ub54c\ubb38\uc5d0 \uc774\ub97c \ucc98\ub9ac\ud558\ub294\ub370 \ud070 \ubb38\uc81c\uac00 \uc0dd\uae41\ub2c8\ub2e4. \uc774\uac83\uc740 \ud76c\uc18c \ubca1\ud130(Sparse Vector)\ub77c\uace0 \ud569\ub2c8\ub2e4. \ub610 \ud558\ub098\uc758 \ubb38\uc81c\ub294 \uac01 \ub2e8\uc5b4\uc758 \uac12\ub4e4\uc740 \ubaa8\ub450 \ub3d9\uc77c\ud55c \uac70\ub9ac(Distance)\ub97c \uac00\uc9c4\ub2e4\ub294 \uac83\uc785\ub2c8\ub2e4. \uc758\ubbf8\ub860\uc801\uc778 \uad6c\ubd84\uc774 \ubd88\uac00\ub2a5\ud558\ub2e4\ub294 \uac83\uc774\uc8e0.<\/p>\n\n\n\n

\uadf8\ub807\uae30 \ub54c\ubb38\uc5d0 \ud544\uc694\ud55c \uac83\uc740 \ub2e8\uc5b4\uc758 \ud06c\uae30\uc640 \uc0c1\uad00 \uc5c6\ub294 \ucc28\uc6d0\uc758 \ubca1\ud130\uc640 0\uacfc 1\uc774 \uc544\ub2cc \uc2e4\uc218\uac12\uc744 \uac00\uc9c0\ub294 \uc0c8\ub85c\uc6b4 \ubca1\ud130\uac00 \ud544\uc694\ud569\ub2c8\ub2e4. \uc774\uac83\uc744 \ubc00\uc9d1 \ubca1\ud130(Dense Vector)\ub77c\uace0 \ud569\ub2c8\ub2e4. \ub610 \uac01 \ub2e8\uc5b4\uac00 \uac00\uc9c0\ub294 \ubca1\ud130\uc5d0 \ubc29\ud5a5\uc131\uc774\ub098 \uc720\uc0ac\ub3c4\uc5d0 \ub530\ub77c\uc11c \uac70\ub9ac\uac00 \uac00\uae5d\uac70\ub098 \uba40\uac70\ub098 \ud558\ub294 \ud2b9\uc9d5\uc744 \uac00\uc9c0\ub3c4\ub85d \ud45c\ud604\ud560 \ud544\uc694\uac00 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n

\uc6cc\ub4dc \uc784\ubca0\ub529\uc740 \uc774\ub7ec\ud55c \ubc00\uc9d1\ub41c \ud615\ud0dc\uc758 \ubca1\ud130\ub97c \ub9cc\ub4dc\ub294 \uacfc\uc815\uc774\ub77c\uace0 \ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n\n\n\n

PyTorch\ub294 \uc785\ub825 \ud14d\uc2a4\ud2b8\ub97c \ubc1b\uc544\uc11c \uc784\ubca0\ub529 \ubca1\ud130\ub97c \uc0dd\uc131\ud558\ub294 nn.Eembedding()\uc744 \uc81c\uacf5\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. index \uac12\uc774 \ubd80\uc5ec\ub418\uc5b4 \uc788\ub294 \ub2e8\uc5b4\ub97c \uc785\ub825 \ubc1b\uc2b5\ub2c8\ub2e4. \uc5ec\uae30\uc11c index\ub294 \uace0\uc720\ud55c \uac12\uc774 \ub429\ub2c8\ub2e4. \uc774 Index\ub97c \ucc38\uc870 \ud14c\uc774\ube14(look-up table)\uc5d0\uc11c \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4. \uc989, |?|\u00d7? \ud06c\uae30\uc758 \ud589\ub82c\uc5d0 \ub2e8\uc5b4 \uc784\ubca0\ub529\uc744 \uc800\uc7a5\ud558\ub294\ub370, D \ucc28\uc6d0\uc758 \uc784\ubca0\ub529 \ubca1\ud130\uac00 \ud589\ub82c\uc758 ?i \ubc88\uc9f8 \ud589\uc5d0 \uc800\uc7a5\ub418\uc5b4\uc788\uc5b4 ?i \ub97c \uc778\ub371\uc2a4\ub85c \ud65c\uc6a9\ud574 \uc784\ubca0\ub529 \ubca1\ud130\ub97c \ucc38\uc870\ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uc5ec\uae30\uc11c |?|\ub294 vocabulary\uc758 \uc218\uc774\uace0 D\ub294 \ucc28\uc6d0\uc815\ubcf4\uc785\ub2c8\ub2e4.<\/p>\n\n\n\n

\"\"\/
https:\/\/wikidocs.net\/64779<\/figcaption><\/figure><\/div>\n\n\n\n

\uc704\uc758 \uadf8\ub9bc\uc740 \ub2e8\uc5b4 great\uc774 \uc815\uc218 \uc778\ucf54\ub529 \ub41c \ud6c4 \ud14c\uc774\ube14\ub85c\ubd80\ud130 \ud574\ub2f9 \uc778\ub371\uc2a4\uc5d0 \uc704\uce58\ud55c \uc784\ubca0\ub529 \ubca1\ud130\ub97c \uaebc\ub0b4\uc624\ub294 \ubaa8\uc2b5\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uc704\uc758 \uadf8\ub9bc\uc5d0\uc11c\ub294 \uc784\ubca0\ub529 \ubca1\ud130\uc758 \ucc28\uc6d0\uc774 4\ub85c \uc124\uc815\ub418\uc5b4\uc838 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\ub9ac\uace0 \ub2e8\uc5b4 great\uc740 \uc815\uc218 \uc778\ucf54\ub529 \uacfc\uc815\uc5d0\uc11c 1,918\uc758 \uc815\uc218\ub85c \uc778\ucf54\ub529\uc774 \ub418\uc5c8\uace0 \uadf8\uc5d0 \ub530\ub77c \ub2e8\uc5b4 \uc9d1\ud569\uc758 \ud06c\uae30\ub9cc\ud07c\uc758 \ud589\uc744 \uac00\uc9c0\ub294 \ud14c\uc774\ube14\uc5d0\uc11c \uc778\ub371\uc2a4 1,918\ubc88\uc5d0 \uc704\uce58\ud55c \ud589\uc744 \ub2e8\uc5b4 great\uc758 \uc784\ubca0\ub529 \ubca1\ud130\ub85c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uc774 \uc784\ubca0\ub529 \ubca1\ud130\ub294 \ubaa8\ub378\uc758 \uc785\ub825\uc774 \ub418\uace0, \uc5ed\uc804\ud30c \uacfc\uc815\uc5d0\uc11c \ub2e8\uc5b4 great\uc758 \uc784\ubca0\ub529 \ubca1\ud130\uac12\uc774 \ud559\uc2b5\ub429\ub2c8\ub2e4.

\ub8e9\uc5c5 \ud14c\uc774\ube14\uc758 \uac1c\ub150\uc744 \uc774\ub860\uc801\uc73c\ub85c \uc6b0\uc120 \uc811\ud558\uace0, \ucc98\uc74c \ud30c\uc774\ud1a0\uce58\ub97c \ubc30\uc6b8 \ub54c \uc5b4\ub5a4 \ubd84\ub4e4\uc740 \uc784\ubca0\ub529 \uce35\uc758 \uc785\ub825\uc774 \uc6d0-\ud56b \ubca1\ud130\uac00 \uc544\ub2c8\uc5b4\ub3c4 \ub3d9\uc791\ud55c\ub2e4\ub294 \uc810\uc5d0 \ud5f7\uac08\ub824 \ud569\ub2c8\ub2e4. \ud30c\uc774\ud1a0\uce58\ub294 \ub2e8\uc5b4\ub97c \uc815\uc218 \uc778\ub371\uc2a4\ub85c \ubc14\uafb8\uace0 \uc6d0-\ud56b \ubca1\ud130\ub85c \ud55c\ubc88 \ub354 \ubc14\uafb8\uace0\ub098\uc11c \uc784\ubca0\ub529 \uce35\uc758 \uc785\ub825\uc73c\ub85c \uc0ac\uc6a9\ud558\ub294 \uac83\uc774 \uc544\ub2c8\ub77c, \ub2e8\uc5b4\ub97c \uc815\uc218 \uc778\ub371\uc2a4\ub85c\ub9cc \ubc14\uafbc\ucc44\ub85c \uc784\ubca0\ub529 \uce35\uc758 \uc785\ub825\uc73c\ub85c \uc0ac\uc6a9\ud574\ub3c4 \ub8e9\uc5c5 \ud14c\uc774\ube14 \ub41c \uacb0\uacfc\uc778 \uc784\ubca0\ub529 \ubca1\ud130\ub97c \ub9ac\ud134\ud569\ub2c8\ub2e4.[1]<\/p>\n\n\n\n

\ucc38\uace0 : Word Embeddings in Pytorch<\/strong><\/p>https:\/\/pytorch.org\/tutorials\/beginner\/nlp\/word_embeddings_tutorial.html<\/a><\/cite><\/blockquote>\n\n\n\n

\ud30c\uc774\ud1a0\uce58 \uacf5\uc2dd\ud648\uc5d0 \uc788\ub294 \ub0b4\uc6a9\uc744 \uc0ac\uc6a9\ud574\uc11c \uac04\ub2e8\ud55c \ud14c\uc2a4\ud2b8 \ucf54\ub4dc\ub97c \ub9cc\ub4e4\uc5b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n

import torch\nimport torch.nn as nn\n\ntrain_data = '\ud0dc\ucd08\uc5d0 \ud558\ub098\ub2d8\uc774 \ucc9c\uc9c0\ub97c \ucc3d\uc870\ud558\uc2dc\ub2c8\ub77c \ucc3d\uc138\uae30 1\uc7a5 1\uc808'.split(' ')\nword_set = list(set(train_data))\nword_to_ix = {tkn:i for i, tkn in enumerate(word_set)}\nword_to_ix<\/pre>\n\n\n\n
\uc785\ub825 \ub370\uc774\ud130\ub97c \uae00\uc790 \ub2e8\uc704\ub85c \ubd84\ub9ac\ud558\uc5ec dict\uc5d0 \uc800\uc7a5\ud569\ub2c8\ub2e4. \uc800\uc7a5\ub41c \ub0b4\uc6a9\uc744 \ucd9c\ub825\ud574\ubcf4\uba74 \uc544\ub798\uc758 \ub0b4\uc6a9\uacfc \uac19\uc2b5\ub2c8\ub2e4. \uc911\ubcf5\uc744 \uc81c\uac70\ud558\uae30 \uc704\ud574 set() \uc790\ub8cc\ud615\uc744 \uc0ac\uc6a9\ud574\uc11c \ub2e8\uc5b4\uc758 \uc21c\uc11c\ub294 \ubb34\uc2dc\ub418\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n
{'\ucc9c\uc9c0\ub97c': 0, '\ud558\ub098\ub2d8\uc774': 1, '\ucc3d\uc870\ud558\uc2dc\ub2c8\ub77c': 2, '\ucc3d\uc138\uae30': 3, '\ud0dc\ucd08\uc5d0': 4, '1\uc808': 5, '1\uc7a5': 6}<\/pre>\n\n\n\n
\uc774\uc81c \uc774\ub807\uac8c \ubd84\ub9ac\ud55c \ub2e8\uc5b4\ub97c \uc544\ub798\uc640 \uac19\uc740 \ubc29\ubc95\uc73c\ub85c one-hot \ud615\ud0dc\ub85c \ub098\ud0c0\ub0b4\uba74 \ub2e4\uc74c\uacfc \uac19\uc740 \ud615\ud0dc\ub85c \ud45c\uc2dc\ub429\ub2c8\ub2e4.<\/p>\n\n\n\n
one_hot = []\nfor i, tkn in enumerate(word_to_ix):\n    one_hot.append(np.eye(len(vocab), dtype='int')[word_to_ix[tkn]])<\/pre>\n\n\n\n
[array([1, 0, 0, 0, 0, 0, 0]), #\ucc9c\uc9c0\ub97c\n array([0, 1, 0, 0, 0, 0, 0]), #\ud558\ub098\ub2d8\uc774\n array([0, 0, 1, 0, 0, 0, 0]), #\ucc3d\uc870\ud558\uc2dc\ub2c8\ub77c\n array([0, 0, 0, 1, 0, 0, 0]), #\ucc3d\uc138\uae30\n array([0, 0, 0, 0, 1, 0, 0]), #\ud0dc\ucd08\uc5d0\n array([0, 0, 0, 0, 0, 1, 0]), #1\uc808\n array([0, 0, 0, 0, 0, 0, 1])] #1\uc7a5<\/pre>\n\n\n\n
\uadf8\ub7ec\ub098 \uc11c\ub450\uc5d0 \uc5b8\uae09\ud588\ub4ef\uc774 \uc774\ub807\uac8c \ud45c\ud604\ub41c \ubca1\ud130 \ub370\uc774\ud130\ub97c \uc9c1\uc811 \ud559\uc2b5\uc5d0 \uc0ac\uc6a9\ud558\uae30\ub294 \uc801\uc808\ud558\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4. \uadf8\ub798\uc11c \ucc28\uc6d0\uc744 \uc785\ub825 \ub370\uc774\ud130\uc758 \ucc28\uc6d0\uc744 \ub0ae\ucdb0\uc8fc\uace0 Sparse\ud55c \ub370\uc774\ud130\ub97c Dense\ud55c \ud615\ud0dc\ub85c \ubcc0\uacbd\ud560 \ud544\uc694\uac00 \uc788\uc2b5\ub2c8\ub2e4.  \uadf8\ub54c \uc0ac\uc6a9\ud558\ub294 \uac83\uc774 Embedding\uc774\ub77c\uace0 \ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n\n\n\n
torch.nn.Embedding()\uc740 \uc774\ub7ec\ud55c \uc791\uc5c5\uc744 \uc27d\uac8c \ud560 \uc218 \uc788\ub3c4\ub85d \ub3c4\uc640\uc90d\ub2c8\ub2e4.<\/p>\n\n\n\n
embeds = nn.Embedding(len(vocab), 3)\nlookup_tensor = torch.tensor([word_to_ix[\"\ud0dc\ucd08\uc5d0\"]], dtype=torch.long)\nw = embeds(lookup_tensor)\nprint(w)<\/pre>\n\n\n\n
\uc704\uc640 \uac19\uc740 \ubc29\ubc95\uc744 \uc0ac\uc6a9\ud558\uba74 [1, 0, 0, 0, 0, 0, 0] \ud615\ud0dc\uc758 one-hot \ubca1\ud130\ub97c [-1.0998, -1.0605, -0.5849] \ud615\ud0dc\uc758 \ub370\uc774\ud130\ub85c \ubcc0\uacbd\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. <\/p>\n\n\n\n
#\uc0dd\uc131\ub41c weight \ubca1\ud130\nParameter containing:\ntensor([[ 0.5765,  0.2391, -0.1834],\n        [-0.1860, -0.0754,  0.4587],\n        [-0.9538, -0.6950,  0.5682],\n        [-2.1076, -0.4070,  0.2598],\n        [-1.0998, -1.0605, -0.5849],\n        [ 1.1632, -0.8139,  0.1154],\n        [ 0.9705,  0.3963,  0.8804]], requires_grad=True)<\/pre>\n\n\n\n
\uc544\ub798\uc758 \ub9c1\ud06c\ub294 Word2Vec\uc744 \uc2e4\uc81c\ub85c \uc5b4\ub5bb\uac8c \uc0ac\uc6a9\ud558\ub294\uc9c0\uc5d0 \ub300\ud55c \uc608\uc81c\uc785\ub2c8\ub2e4. \uad81\uae08\ud558\uc2e0 \ubd84\ub4e4\uc740 \ucc38\uace0\ud558\uc2dc\uae30 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n\n\n\n
\n
Word2Vec \uad6c\ud604<\/a><\/blockquote>