L'API PaLM est obsolète et peut être utilisée avec les services et outils d'IA de Google (mais pas pour Vertex AI). Pour en savoir plus sur cet abandon, le calendrier associé et la migration vers Gemini, consultez le guide d'abandon de l'API PaLM.

Cette page a été traduite par l'API Cloud Translation.

API PaLM: Guide de démarrage rapide sur les représentations vectorielles continues avec Python

Voir sur ai.google.dev

Exécuter dans Google Colab

Afficher la source sur GitHub

Dans ce notebook, vous allez apprendre à utiliser l'API PaLM, qui vous donne accès aux derniers grands modèles de langage de Google. Vous allez apprendre à utiliser les fonctionnalités de génération de représentations vectorielles continues de l'API PaLM et voir un exemple de ce que vous pouvez faire avec ces représentations vectorielles continues.

Préparation

Commencez par télécharger et installer la bibliothèque Python de l'API PaLM.

pip install -U google-generativeai

import numpy as np
import google.generativeai as palm

Obtenir une clé API

Pour commencer, vous devez créer une clé API.

palm.configure(api_key='PALM_KEY')

Que sont les représentations vectorielles continues ?

La technique des représentations vectorielles continues permet de représenter du texte (comme des mots, des phrases ou des paragraphes entiers) sous la forme d'une liste de nombres à virgule flottante dans un tableau. Ces chiffres ne sont pas aléatoires. L'idée principale est que les textes ayant des significations similaires auront des représentations vectorielles continues similaires. Vous pouvez utiliser la relation entre eux pour de nombreuses tâches importantes.

Génération de représentations vectorielles continues

Dans cette section, vous allez voir comment générer des représentations vectorielles continues pour un texte à l'aide de la fonction palm.generate_embeddings de l'API PaLM. Voici une liste de modèles compatibles avec cette fonction.

for model in palm.list_models():
  if 'embedText' in model.supported_generation_methods:
    print(model.name)

models/embedding-gecko-001

Utilisez la fonction palm.generate_embeddings et transmettez le nom du modèle ainsi qu'un texte. Vous obtenez une liste de valeurs à virgule flottante. Commencez par une requête "Que mangent les écureuils ?" et observez la relation entre deux chaînes différentes.

x = 'What do squirrels eat?'

close_to_x = 'nuts and acorns'

different_from_x = 'This morning I woke up in San Francisco, and took a walk to the Bay Bridge. It was a good, sunny morning with no fog.'

model = "models/embedding-gecko-001"

# Create an embedding
embedding_x = palm.generate_embeddings(model=model, text=x)
embedding_close_to_x = palm.generate_embeddings(model=model, text=close_to_x)
embedding_different_from_x = palm.generate_embeddings(model=model, text=different_from_x)

print(embedding_x)

{'embedding': [-0.025894878, -0.02103396, 0.003574992, 0.00822288, 0.03276648, -0.10068223, -0.037702546, 0.01079403, 0.0001406235, -0.029412385, 0.01919925, 0.0048481044, 0.070619866, -0.013349887, 0.028378602, -0.018658886, -0.038629908, 0.056883123, 0.06332366, 0.039849922, -0.085393265, -0.016251814, -0.025535949, 0.0049480307, 0.048581485, -0.11295683, 0.033869933, 0.015498774, -0.07306243, 0.000857902, -0.022031788, -0.005298939, -0.08311722, -0.027091762, 0.042790364, 0.023175264, 0.011238991, -0.02432924, -0.0044626957, 0.05167071, 0.023430848, 0.027325166, -0.01492389, -0.018770715, -0.003783692, 0.040971957, -0.044652887, 0.033220302, -0.05659744, -0.055191413, -0.0023204528, -0.043687623, 0.030044463, -0.015966717, -0.04318426, 0.015735775, -0.038352676, -0.005009736, -0.03289721, 0.016246213, -0.005696393, -0.0010992853, -0.02768714, -0.03534994, -0.045970507, 0.05784305, -0.026696421, -0.013302212, 0.007055761, -0.05885901, 0.03330113, 0.04399591, 0.020755561, 0.0028288597, 0.037333105, 0.0103595415, -0.01942964, 0.033088185, 0.009558319, -0.06524442, -0.07101354, -0.053975347, -0.003952934, -0.11641813, -0.039488368, -0.0033782825, -0.017735159, 0.03198736, 0.014555729, 0.050724585, -0.07849815, -0.0070436746, 0.017992217, -0.003975652, -0.0039650565, 0.08063971, -0.011685766, -0.018323965, 0.007763516, 0.012011537, 0.028457757, -0.099603206, 0.0328822, 0.0063217366, 0.051288057, 0.060445003, -0.007725884, -0.0033487668, -0.02697037, -0.04471915, 0.014793467, 0.0029390613, -0.04365732, -0.036976494, 0.05571355, -0.034228597, 0.05610819, 0.0016565409, 0.06461147, 0.012197695, -0.029221235, 0.015400638, 0.009992722, -0.0126949195, 0.027302667, 0.04309881, 0.013308768, -0.034253325, -0.028620966, 0.0032988666, 0.008901495, 0.0051033413, 0.08693829, -0.035939537, -0.00014025549, -0.0021354076, 0.043875773, -0.057092454, 0.0048032254, 0.04456835, -0.01337361, 0.018620204, -0.0037525205, 0.018113593, -0.0024051766, -0.006519982, 0.043426506, -0.028869089, -0.07003764, -0.027043046, -0.047674373, -0.036566455, -0.029664699, 0.054604772, 0.056459025, 0.016209831, 0.06588335, 0.07294827, -0.07351654, -0.050157, 0.05211485, -0.02302033, 0.022877783, 0.013553745, -0.019406103, -0.0058154585, 0.0373227, 0.0052685454, 0.02164789, -0.019631775, -0.015719362, -0.06862338, 0.021698158, -0.013781832, 0.06955018, -0.023942512, -0.018029014, -0.018007774, -0.0059923544, -0.02771734, -0.0019507131, -0.069619514, 0.054189045, 0.0021985532, -0.01132558, 0.015128105, 0.015424623, -0.038302787, -0.038970694, 0.044268098, 0.015156813, 0.030262465, -0.0010455108, -0.032175235, -0.03357542, -9.529959e-05, 0.062028274, -0.10134925, -0.009874221, 0.051682726, -0.022124732, 0.010147164, -0.012185555, 0.03731382, -0.00059438165, -0.017981028, -0.070909515, 0.02605233, 0.06992509, 0.026033426, -0.023944097, -0.047794044, 0.0204043, 0.025562089, -0.01985736, -0.027300185, 0.029983355, -0.0821883, -0.018791717, -0.004772287, -0.02490102, -0.010111937, 0.050968856, 0.029660473, 3.4716293e-05, -0.017517656, 0.023977743, 0.022549666, 0.04181301, 0.007500569, -0.0019229053, 0.023285722, -0.010899088, -0.004949611, -0.012531907, 0.041027624, -0.004620342, -0.013926477, -0.020054528, 0.026111232, -0.06232942, 0.09978252, -0.044156674, 0.061204664, 0.007044644, -0.0027112814, 0.04620226, 0.006134901, 0.03983195, -0.009853767, 0.0137631735, -0.07085734, 0.009606741, -0.008636412, 0.050337072, 0.045284208, -0.0032710661, -0.016086245, 0.008386805, -0.007903436, 0.0350885, 0.0025110857, 0.04684593, 0.12780859, -0.038998656, -0.029157333, -0.029113598, 0.0074333544, 0.05532698, -0.034412585, -0.00013683736, -0.020530468, 0.06506163, 0.0019480588, 0.0030335467, -0.018495142, -0.054084025, 0.023021378, -0.010500294, -0.007759436, -0.020039978, -0.017755102, 0.0006766737, 0.014525485, -0.026014434, 0.002474586, -0.027173916, 0.0093613025, 0.0058087856, 0.0006998545, 0.04791365, -0.04368597, -0.015235596, 0.0069595333, 0.009612967, -0.0009247106, 0.033619776, -0.00649697, -0.04766721, 0.0391879, -0.010284179, -0.006610166, -0.0020641836, -0.05440346, -0.007050968, -0.015853178, -0.031741284, -0.02172385, 0.03021658, -0.0012069787, 0.050265886, 0.04510601, -0.024716277, -0.05543306, -0.06419837, -0.014273427, -0.023703339, 0.0017521745, -0.056149185, 0.0069642677, 0.0065768356, 0.035255834, 0.039023213, 0.016403731, 0.025051782, 0.00695039, -0.05579997, 0.013183741, 0.08474835, -0.012680079, 0.0041794777, 0.02355896, -0.07197163, 0.024911461, -0.018766653, 0.025204346, 0.0048066434, 0.04904056, 0.016669538, -0.037882168, -0.021643393, 0.0053031743, -0.031009668, -0.016543044, -0.020345997, -0.005761681, -0.0743119, -0.02601627, -0.023271384, -0.07075993, -0.0029876109, 0.0066218525, -0.061091717, 0.032953493, 0.03662513, 0.010290128, 0.05418312, -0.03828874, 0.03312786, -0.014862627, -0.03720938, 0.018570531, -0.020742243, 0.048026983, 0.005438336, 0.020241424, -0.04405181, 0.030792728, 0.033958763, -0.023588262, 0.037658524, 0.010072951, 0.0064869304, 0.019048406, -0.06919818, -0.017083945, -0.016801478, 0.0027415873, 0.008172279, 0.0019755305, -0.057162683, -0.0053946367, 0.0014972482, -0.033361986, -0.0033606717, 0.03242665, 0.072544955, 0.02279949, -0.046871353, -0.06308129, 0.029209439, 0.011341486, 0.032790348, -0.020073028, -0.0044093695, 0.08292041, -0.03140556, 0.009308279, -0.004211382, -0.052444175, 0.0180874, 0.008575959, -0.0013550716, -0.07186043, 0.028372435, 0.024996122, 0.027749002, 0.016944503, -0.014632978, -0.06674174, -0.043031745, -0.044137582, 0.03530514, 0.030504197, 0.060496386, -0.06423886, 0.012235539, -0.05830343, -0.015868725, 0.041861057, 0.027080601, -0.014182999, -0.028095996, 0.0016349283, 0.010679886, 0.048808616, -0.058294244, -0.010633062, -0.056791265, -0.027161647, -0.030019993, -0.010299281, -0.03821823, -0.016588321, -0.0059704296, -0.053497788, 0.05661912, 0.005010262, -0.020186698, -0.03151958, -0.07490499, 0.045715272, -0.03747153, 0.02902543, 0.015007152, -0.01799195, 0.0079564275, -0.028715475, -0.018788284, -0.041037183, 0.012932907, -0.0072463937, -0.0046510296, 0.052094106, 0.047214568, -0.05604256, 0.006124289, -0.06112983, -0.028900363, -0.0033062366, -0.016411366, -0.03985708, -0.005927899, 0.027991273, -0.034023542, 0.0023991684, 0.020010024, 0.014298016, 0.017212953, 0.002652654, -0.08308305, 0.01726592, 0.013845524, 0.0065021385, 0.0364733, 0.020361774, 0.09685079, 0.04039578, 0.016480403, -0.08329836, -0.06590067, 0.00012861127, -0.055775307, 0.0065172235, -0.018937778, -0.021399701, 0.0004559998, -0.0097613875, -0.003239602, 0.0041429265, 0.059930306, -0.01656465, 0.018544743, -0.03232914, 0.006037772, -0.06402926, 0.05761484, -0.02093143, 0.018229362, 0.024098346, 0.025045564, -0.009451666, -0.010259512, 0.006660359, -0.029620942, -0.03495546, -0.06783166, -0.03193859, -0.04261954, 0.027878316, 0.023951625, 0.016354026, -0.0015310713, -0.05785183, -0.04868827, -0.06779814, -0.09212996, 0.04355289, 0.02634198, 0.045933742, -0.012108333, -0.017381534, 0.012251423, 0.035591044, 0.05024221, 0.056855064, 0.0101336455, -0.009532219, -0.054251555, 0.034745548, 0.020292252, 0.033525895, -0.040225316, -0.00015249893, -0.07806101, 0.0075722514, 0.015309747, 0.022623314, 0.06536824, 0.064232446, -0.01557734, -0.04813796, -0.013913105, 0.020742541, 0.060864896, -0.056623433, 0.057601452, -1.6570028e-05, 0.010925783, 0.0036125665, 0.032784764, -0.0801319, -0.048450164, 0.06296668, 0.02989288, -0.011754737, -0.0010066505, -0.05441974, -0.017106231, -0.04285682, -0.005424776, -0.028312048, -0.0022843084, -0.02028908, -0.007416978, 0.016722959, 0.03343588, -0.049168676, 0.003828647, 0.043084797, -0.011436926, -0.017679023, -0.012748326, -0.015104218, 0.008225339, -0.005965197, -0.010827806, -0.015990732, 0.031933613, 0.01862576, -0.013171726, 0.007987761, -0.018449496, 0.041906953, -0.020788714, 0.03404006, -0.00086082605, -0.007771558, 0.023855729, -0.00295711, -0.0085285455, -0.0556957, -0.005321175, -0.018151492, -0.011129989, -0.05183511, 0.0053123147, 0.009127998, -0.011530388, 0.009631709, 0.0041047884, -0.0353711, 0.052883077, -0.01532676, 0.03040235, 0.008731032, -0.00441319, 0.01950203, 0.014064995, 0.03141337, 0.018041868, 0.059427522, 0.048374873, -0.019928444, -0.004559623, 0.021962427, -0.08567552, -0.007796494, 0.033520035, 0.009779213, 0.05753526, 0.010492746, -0.039363436, -0.103733934, -0.024229618, 0.0062162466, -0.017748242, 0.005122951, -0.055344906, -0.010650967, 0.0309389, -0.073542334, -0.014872006, -0.003081951, 0.016437916, -0.0040901243, 0.0018574661, 0.03331834, 0.005815743, 0.022556618, 0.076257, -0.0065593896, -0.026774084, -0.016839791, 0.008689688, -0.015184644, 0.0073800148, -0.018499345, -0.036080927, 0.053406574, 0.015944907, -0.014478417, -0.021485219, -0.018035412, -0.038147416, 0.014293582, -0.021055873, 0.0314314, -0.07782329, 0.015536577, -0.031045694, 0.059434652, -0.020065695, 0.052754566, -0.08380041, 0.06855744, 0.012167185, -0.015827801, 0.04380172, 0.020258602, -0.058169313, -0.04435873, -0.013054301, -0.041333184, -0.02302342, 0.029140746, 0.00812361, 0.033690967, -0.0030892044, 0.052916355, -0.04835076, -0.0101818545, -0.05420185, -0.033779036, 0.02638142, -0.028346056, -0.02331669, -0.005781761, 0.012981267, -0.0055279816, 0.010089179, -0.04489518, -0.024379171, 0.007590703, -0.025511196, -0.06555892, 0.008145539, 0.021736145, -0.033178225, 0.026871512, -0.05637406, -0.030885229, 0.014512168, -0.008024667, 0.026689196, 0.004108927, -0.04103957, 0.0080031715, -0.0030232186, -0.036158007, 0.04256502, -0.0001681743, 0.0117336465, 0.025762333, -0.010921032, -0.0010622365, -0.07185124, 0.029530818, 0.009698986, 0.011916085, 0.0022654524, 0.07175238, 0.029233111, -0.020834876, -0.052442703, 0.011248308, 0.005422925, 0.018166017, 0.0472275, -0.013550265, 0.0350743, -0.010435109, 0.047774173, 0.021216916, -0.0026447468, -0.021085296, 0.013272342, -0.0133805, 0.02943836, -0.032338675, 0.0021435472, -0.016289461, -0.013629232, -0.03840216, 0.06655019, 0.009643849, 0.025085986, -0.018909356, -0.011246176, -0.05254555, -0.06776485, -0.02931862, 0.014850466, 0.029691922, -0.04090594, 0.0544204, 0.01552631, 0.02912549, -0.0020693596, 0.038805272, -0.009980787, 0.031122748, -0.05562063, 0.021108221, 0.0103203785, 0.044171233, 0.009732269, -0.0011330071]}

Maintenant que vous avez créé les représentations vectorielles continues, utilisons le produit scalaire pour voir dans quelle mesure close_to_x et different_from_x sont liés à x. Le produit scalaire renvoie une valeur comprise entre -1 et 1, et représente l'écart entre deux vecteurs selon la direction dans laquelle ils pointent. Plus la valeur est proche de 0, moins les objets sont semblables aux objets (dans ce cas, deux chaînes). Plus la valeur est proche de 1, plus elles sont similaires.

similar_measure = np.dot(embedding_x['embedding'], embedding_close_to_x['embedding'])

print(similar_measure)

0.7314063252924405

different_measure = np.dot(embedding_x['embedding'], embedding_different_from_x['embedding'])

print(different_measure)

0.43560702838194704

Comme indiqué ici, la valeur de produit scalaire la plus élevée entre les représentations vectorielles continues de x et close_to_x montre plus de relation que celles de x et different_from_x.

Que pouvez-vous faire avec les représentations vectorielles continues ?

Vous avez généré votre premier ensemble de représentations vectorielles continues avec l'API PaLM. Mais que pouvez-vous faire avec cette liste de valeurs à virgule flottante ? Les représentations vectorielles continues peuvent être utilisées pour une grande variété de tâches de traitement du langage naturel (TLN), telles que:

Recherche (documents, Web, etc.)
Systèmes de recommandation
Clustering
Analyse des sentiments/Classification de texte

Pour consulter des exemples, cliquez ici.