15 KiB
15 KiB
Zajęcia 8: Vector representations
Wszystkie zadania ćwiczeniowe należy rozwiązywać w języku Python w kopii Jupyter Notebook'a dla danych zajęć w wyznaczonych miejscach (komórki z komentarzem # Solution).
Nie należy usuwać komórek z treścią zadań.
Należy wyświetlać outputy przy pomocy print
Dla chętnych! (może się przydać od ambitniejszych projektów końcowych)
https://github.com/huggingface/smol-course - kurs finetune'owania LLMów do własnych zadań
https://github.com/unslothai/unsloth - biblioteka do efektywnego finetune'owania LLMów (są gotowe notebooki z kodem na platformie Colab)
Co to jest wektor?
Wektor - jednowymiarowa macierz
[0, 1, 0, 0, 0] - one hot encoding - tylko wartości 0/1
[0, 2, 0, 5, 1, 100] - frequency encoding - liczby całkowite >= 0
[-1.5, 0.0002, 5000.01] - wektor
Zadanie 1
Dokonaj preprocessingu tekstów https://git.wmi.amu.edu.pl/ryssta/spam-classification/src/branch/master/train/in.tsv (preprocessing - proces wstępnej "obróbki tekstów - sprowadzenie do małych liter, tokenizacja itd..) i dokonaj:
- a) one hot encodingu
- b) frequency encodingu
dla dwóch przykładowych zdań, które zawierają przynajmniej 2 wystąpienia słowa, które znajduje się w słowniku (czyli występuje w korpusie z pliku in.tsv). Ze względu na dużą liczbę unikalnych słów w korpusie, proszę nie printować całych wektorów, tylko indeksy oraz wartości.
Zadanie 2
Na podstawie pliku https://git.wmi.amu.edu.pl/ryssta/spam-classification/src/branch/master/train/in.tsv oraz pliku https://git.wmi.amu.edu.pl/ryssta/spam-classification/src/branch/master/train/expected.tsv podziel teksty względem klasy spam/nie spam. Oblicz wartość IDF osobno dla tekstów klasy spam oraz dla tekstów klasy nie spam, dla słów:
- free
- send
- are
- the
oraz 2-3 własnoręcznie wybranych słów.
Zadanie 3
Na podstawie warstwy embedding modelu gpt2 wypisz 15 najbardziej podobnych (względem miary podobieństwa cosinuowego) tokenów do słów:
- cat
- tree
oraz 2 własnoręcznie wybranych tokenów.
from transformers import GPT2Tokenizer, GPT2Model
import torch
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
model = GPT2Model.from_pretrained('gpt2')
embedding_layer = model.wte
cos_sim = torch.nn.CosineSimilarity()c:\Users\ryssta\AppData\Local\anaconda3\Lib\site-packages\transformers\tokenization_utils_base.py:1601: FutureWarning: `clean_up_tokenization_spaces` was not set. It will be set to `True` by default. This behavior will be depracted in transformers v4.45, and will be then set to `False` by default. For more details check this issue: https://github.com/huggingface/transformers/issues/31884 warnings.warn(
[
[0.1, 0.2, 0.3], # Ala
[-0.5, 0.5, 0.9], # ma
...
# 50254
...
[0.1, -0.1, -0.2] # w GPT2 jest 768 wartości w pojedynczym wektorze, a nie 3
]print("Tekst 'cat' jest konwertowany do tokenu 9246")
print("\nTokenizacja")
print(tokenizer("computer"))
print("\nDetokenizacja")
print(tokenizer.decode([33215]))
print("\nLiczba tokenów w słowniku")
print(len(tokenizer))Tekst 'cat' jest konwertowany do tokenu 9246
Tokenizacja
{'input_ids': [33215], 'attention_mask': [1]}
Detokenizacja
computer
Liczba tokenów w słowniku
50257
print("Embedding tokenu: 9246")
cat_embedding = embedding_layer(torch.LongTensor([9246]))
print("\nRozmiar embeddingu (wektora)")
print(cat_embedding.shape)
print("\nWartości embeddingu")
print(cat_embedding)Embedding tokenu: 9246
Rozmiar embeddingu (wektora)
torch.Size([1, 768])
Wartości embeddingu
tensor([[-0.0164, -0.0934, 0.2425, 0.1398, 0.0388, -0.2592, -0.2724, -0.1625,
0.1683, 0.0829, 0.0136, -0.2788, 0.1493, 0.1408, 0.0557, -0.3691,
0.2200, -0.0428, 0.2206, 0.0865, 0.1237, -0.1499, 0.1446, -0.1150,
-0.1425, -0.0715, -0.0526, 0.1550, -0.0678, -0.2059, 0.2065, -0.0297,
0.0834, -0.0483, 0.1207, 0.1975, -0.3193, 0.0124, 0.1067, -0.0473,
-0.3037, 0.1139, 0.0949, -0.2175, 0.0796, -0.0941, -0.0394, -0.0704,
0.2033, -0.1555, 0.2928, -0.0770, 0.0787, 0.1214, 0.1528, -0.1464,
0.4247, 0.1921, -0.0415, -0.0850, -0.2787, 0.0656, -0.2026, 0.1856,
0.1353, -0.0820, -0.0639, 0.0701, 0.1680, 0.0597, 0.3265, -0.1100,
0.1056, 0.1845, -0.1156, 0.0054, 0.0663, 0.1842, -0.1069, 0.0491,
-0.0853, -0.2519, 0.0031, 0.1805, 0.1505, 0.0442, -0.2427, 0.1104,
0.0970, 0.1123, -0.1519, -0.1444, 0.2323, -0.0241, -0.0677, 0.1157,
-0.2668, -0.1229, 0.1120, 0.0601, -0.0535, 0.1259, -0.0966, -0.1975,
-0.2031, 0.1323, 0.0176, -0.1332, 0.1159, 0.1037, 0.0722, 0.1644,
-0.0775, -0.0227, 0.1146, 0.0060, 0.3959, -0.0828, 0.0125, 0.0415,
-0.0147, -0.1352, -0.0579, 0.0423, -0.0793, -0.2702, 0.2806, -0.0744,
0.1118, 0.0908, 0.0639, -0.0882, -0.0190, -0.1386, -0.0490, -0.1785,
0.1416, 0.0497, -0.0461, -0.1544, 0.0662, -0.0538, 0.0992, 0.1308,
-0.0885, -0.2840, -0.0297, -0.0882, -0.0340, -0.1495, 0.0295, 0.0700,
0.0661, -0.1282, -0.0546, -0.1392, -0.1368, 0.0353, 0.0814, -0.1436,
-0.0559, 0.1523, -0.0780, 0.2562, 0.0164, -0.0433, 0.0468, 0.2896,
0.0069, 0.2136, 0.0378, -0.1625, -0.0421, -0.0109, -0.0386, 0.0453,
0.2572, 0.0323, -0.1206, 0.0135, -0.0171, 0.0404, -0.2544, 0.1455,
0.3265, -0.0545, 0.0887, -0.0321, 0.1485, -0.0699, -0.0606, 0.2177,
-0.1566, -0.0619, -0.2655, 0.2471, -0.1213, -0.0741, 0.1074, -0.0243,
-0.2747, 0.1828, 0.0046, 0.1426, 0.0201, 0.0413, -0.0189, -0.0070,
-0.0338, -0.1384, 0.0269, 0.1447, -0.0216, -0.0042, -0.0363, -0.0579,
-0.0909, -0.1359, 0.1407, 0.1421, 0.0041, 0.0100, 0.0997, -0.0718,
-0.0958, 0.0051, -0.2576, 0.1980, 0.1545, 0.0226, -0.2521, -0.1091,
-0.1467, -0.1140, 0.1383, -0.1952, 0.0554, -0.0036, -0.2915, -0.1645,
0.0469, -0.2251, 0.2000, 0.1070, 0.1651, -0.0781, 0.1511, -0.0095,
0.0925, 0.0776, 0.1631, 0.1563, 0.0286, -0.1157, 0.0349, 0.0033,
-0.0870, -0.0864, 0.1233, -0.0691, -0.0458, -0.0601, 0.0501, -0.1450,
-0.2425, -0.0773, 0.1182, 0.1351, 0.1904, 0.1746, 0.0925, -0.1253,
-0.1149, -0.1312, -0.2170, 0.0682, -0.0121, -0.1774, -0.0857, -0.1906,
0.2842, -0.0410, 0.0530, 0.0480, -0.0641, -0.0911, 0.2907, -0.2503,
-0.1085, 0.1753, 0.0610, 0.0466, 0.0097, -0.1300, -0.0273, 0.0498,
-0.0619, -0.1867, -0.0769, -0.1091, -0.0410, -0.0617, -0.0537, 0.0582,
-0.0986, -0.2655, 0.1236, -0.0026, 0.0444, -0.1018, -0.1652, 0.0174,
-0.2561, 0.0440, 0.2048, 0.0049, -0.0220, -0.1031, -0.1387, 0.0493,
0.2048, 0.0473, 0.1630, -0.0195, -0.0714, -0.0713, -0.2524, 0.0852,
-0.1682, 0.0713, 0.1301, -0.1088, -0.0188, 0.0092, 0.0078, 0.2213,
0.0638, -0.1617, -0.0365, -0.0923, -0.1052, 0.1108, -0.1175, -0.0016,
-0.0258, 0.0902, 0.1089, 0.1685, -0.2664, -0.0309, -0.0187, -0.0678,
-0.1424, -0.0026, -0.0623, -0.0575, -0.1009, 0.0142, -0.1950, 0.0085,
-0.1402, 0.0371, -0.4072, -0.0478, 0.4013, 0.3212, 0.1051, 0.0349,
-0.1302, -0.0298, -0.1738, 0.0692, -0.2638, 0.1268, 0.1773, -0.1094,
0.0737, -0.0460, 0.1870, -0.0605, -0.1308, -0.0920, -0.0290, -0.0542,
0.1214, -0.0308, -0.1173, -0.2127, 0.0209, 0.2911, -0.1751, 0.0469,
0.0740, -0.1323, 0.0283, 0.2125, 0.1870, 0.0978, 0.1799, 0.2669,
0.1709, -0.1191, -0.2022, 0.0445, 0.0601, -0.1820, 0.0224, 0.1902,
-0.3199, -0.2551, 0.0795, 0.0814, 0.1245, 0.0871, -0.0455, -0.2342,
0.1167, 0.0870, 0.0257, -0.2073, 0.1849, -0.0184, 0.0498, -0.1423,
-0.0682, 0.1386, 0.0406, -0.0325, 0.2179, -0.0567, -0.2568, -0.0935,
-0.0453, -0.1317, 0.0682, -0.2721, -0.2026, -0.0565, -0.0134, -0.0423,
-0.0415, -0.0560, -0.1522, 0.1617, -0.0753, -0.1967, 0.0536, -0.0988,
0.0539, -0.0489, 0.0296, 0.1272, 0.1567, 0.0185, -0.0855, -0.2336,
0.1859, 0.1528, -0.1824, -0.0834, -0.1414, -0.0526, 0.1744, -0.0290,
-0.0753, -0.0100, -0.1702, -0.0676, 0.0856, 0.0493, -0.1256, -0.1652,
-0.1317, 0.0677, 0.0209, -0.0346, 0.0048, 0.1209, 0.1959, 0.1520,
0.0793, -0.1492, 0.3141, 0.1526, -0.1732, -0.0914, 0.1339, -0.1410,
-0.0595, -0.0250, -0.1136, -0.1206, -0.1126, -0.0470, 0.1898, 0.0565,
-0.2058, 0.0389, 0.0177, -0.2718, 0.2021, -0.0779, 0.1444, 0.1047,
-0.2096, -0.0210, 0.1791, -0.4005, -0.1931, 0.1083, 0.2465, 0.1026,
-0.0503, 0.1047, 0.0299, -0.1043, 0.0964, 0.0852, -0.2067, 0.1263,
0.2064, 0.2248, 0.2739, -0.1881, -0.0745, 0.0769, 0.2994, 0.2803,
0.0063, 0.2585, -0.0176, 0.2318, -0.0432, 0.1889, -0.0766, 0.0751,
-0.0157, 0.0517, 0.1274, -0.2235, -0.0450, 0.1606, 0.0876, 0.1240,
0.4417, -0.0625, 0.0591, -0.0181, 0.1996, 0.0959, -0.2623, -0.2826,
0.0023, 0.1835, 0.1931, -0.1054, 0.1816, -0.1599, -0.0871, 0.0115,
0.2386, 0.0161, 0.0580, -0.0558, 0.0963, 0.1206, -0.3461, 0.0726,
0.0301, 0.1058, 0.0532, 0.0515, 0.0216, -0.0531, -0.0217, 0.0539,
-0.0191, 0.0636, -0.1527, -0.1670, 0.0756, 0.0167, -0.0437, 0.0050,
-0.1861, 0.0304, 0.2442, -0.0126, -0.2314, 0.1562, 0.1635, -0.1206,
-0.0428, 0.1079, 0.1216, -0.0113, 0.1757, -0.0235, 0.2049, -0.3030,
0.0067, -0.3157, 0.1435, -0.1737, -0.1698, 0.2276, -0.0360, 0.0048,
-0.2974, 0.2021, 0.1380, 0.1129, -0.0626, 0.1347, 0.0729, 0.0481,
-0.1397, 0.0197, -0.0932, 0.1717, -0.1519, -0.0554, -0.0344, 0.0201,
0.1316, 0.0743, -0.1189, 0.2787, 0.0597, -0.2073, -0.3555, -0.0645,
-0.1326, 0.1094, 0.1512, 0.0241, 0.0608, 0.0334, 0.1340, -0.0510,
-0.0197, 0.0681, -0.1494, 0.2410, -0.1016, 0.1148, -0.1280, 0.0576,
-0.0101, 0.1957, -0.2854, 0.0231, -0.0282, -0.0101, 0.0383, -0.0086,
-0.1152, 0.1464, 0.0351, -0.2189, 0.2156, 0.0722, 0.0881, 0.2360,
0.1182, 0.0676, 0.0506, -0.0226, -0.2842, -0.2781, 0.0791, 0.1958,
0.0470, -0.0645, 0.0129, 0.2878, 0.1339, -0.2548, -0.0855, -0.1515,
0.1442, -0.1694, -0.2864, -0.1720, -0.0473, -0.1325, 0.2390, 0.0736,
-0.1975, -0.1141, -0.1409, 0.1901, -0.0673, 0.0887, 0.2132, 0.1869,
-0.2635, 0.1148, -0.0100, 0.0344, -0.1273, -0.0943, 0.1698, 0.1511,
-0.2248, 0.0495, 0.1384, 0.1260, 0.0787, 0.3357, 0.2288, -0.0885,
0.0622, -0.2151, 0.0553, 0.0690, 0.2568, -0.2637, -0.0460, 0.0911,
-0.0214, 0.1712, 0.0342, -0.0438, -0.1457, -0.0953, 0.0289, -0.0830,
-0.0591, -0.0244, 0.1259, 0.0075, -0.1052, 0.1119, 0.0621, -0.2382,
0.1763, -0.1132, -0.1201, 0.0162, -0.1807, 0.1419, 0.0749, -0.2318,
0.0132, 0.2532, -0.1028, 0.0556, -0.0214, 0.1974, -0.1573, 0.1861,
-0.1729, -0.0737, -0.1499, -0.1533, 0.0011, 0.0681, -0.1828, 0.0519,
-0.0837, 0.0671, -0.1867, 0.0012, 0.0377, 0.1061, -0.1713, -0.1579]],
grad_fn=<EmbeddingBackward0>)
print("Podobieństwo tego samego embeddingu (powinno wyjść 1)")
print(cos_sim(cat_embedding, cat_embedding))Podobieństwo tego samego embeddingu (powinno wyjść 1) tensor([1.0000], grad_fn=<SumBackward1>)