API REST: guida rapida dell'ottimizzazione

Visualizza su ai.google.dev Prova un blocco note di Colab Visualizza blocco note su GitHub Scarica blocco note

In questo blocco note imparerai come iniziare a utilizzare il servizio di ottimizzazione dell'API PaLM utilizzando i comandi curl o l'API di richiesta Python per chiamare l'API REST PaLM. Qui imparerai come ottimizzare il modello di testo alla base del servizio di generazione del testo dell'API PaLM.

Configurazione

Autentica

L'API PaLM consente di ottimizzare i modelli in base ai tuoi dati. Poiché si tratta dei tuoi dati i tuoi modelli ottimizzati richiedono controlli di accesso più rigidi di quelli forniti dalle chiavi API.

Prima di poter eseguire questo tutorial, dovrai per configurare OAuth per il progetto.

Se vuoi eseguire questo blocco note in Colab, inizia caricando il tuo client_secret*.json utilizzando "File > Carica" .

Mostra file di Colab > Opzione di caricamento

cp client_secret*.json client_secret.json
ls
client_secret.json

Questo comando gcloud trasforma il file client_secret.json in credenziali che possono essere utilizzate per l'autenticazione con il servizio.

import os
if 'COLAB_RELEASE_TAG' in os.environ:
  # Use `--no-browser` in colab
  !gcloud auth application-default login --no-browser --client-id-file client_secret.json --scopes='https://www.googleapis.com/auth/cloud-platform,https://www.googleapis.com/auth/generative-language.tuning'
else:
  !gcloud auth application-default login --client-id-file client_secret.json --scopes='https://www.googleapis.com/auth/cloud-platform,https://www.googleapis.com/auth/generative-language.tuning'

Chiamata all'API REST con CURL

Questa sezione fornisce istruzioni curl di esempio per chiamare l'API REST. Imparerai come creare un job di ottimizzazione, controllarne lo stato e, al termine, effettuare una chiamata di inferenza.

Imposta variabili

Imposta le variabili per i valori ricorrenti da utilizzare per le altre chiamate API REST. Il codice utilizza la libreria Python os per impostare le variabili di ambiente accessibili in tutte le celle di codice.

È specifico per l'ambiente del blocco note di Colab. Il codice nella cella di codice successiva equivale all'esecuzione dei seguenti comandi in un terminale bash.

export access_token=$(gcloud auth application-default print-access-token)
export project_id=my-project-id
export base_url=https://generativelanguage.googleapis.com
import os

access_token = !gcloud auth application-default print-access-token
access_token = '\n'.join(access_token)

os.environ['access_token'] = access_token
os.environ['project_id'] = "project-id"
os.environ['base_url'] = "https://generativelanguage.googleapis.com"

Elenco modelli ottimizzati

Verifica la configurazione dell'autenticazione elencando i modelli ottimizzati attualmente disponibili.


curl -X GET ${base_url}/v1beta3/tunedModels \
    -H 'Content-Type: application/json' \
    -H "Authorization: Bearer ${access_token}" \
    -H "x-goog-user-project: ${project_id}" | grep name
"name": "tunedModels/testnumbergenerator-fvitocr834l6",
      "name": "tunedModels/my-display-name-81-9wpmc1m920vq",
      "displayName": "my display name 81",
      "name": "tunedModels/number-generator-model-kctlevca1g3q",
      "name": "tunedModels/my-display-name-81-r9wcuda14lyy",
      "displayName": "my display name 81",
      "name": "tunedModels/number-generator-model-w1eabln5adwp",
% Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100 17583    0 17583    0     0  51600      0 --:--:-- --:--:-- --:--:-- 51563

Crea modello ottimizzato

Per creare un modello ottimizzato, devi passare il set di dati al modello nel campo training_data.

Per questo esempio, ottimizzerai un modello per generare il numero successivo nella sequenza. Ad esempio, se l'input è 1, il modello dovrebbe restituire 2. Se l'input è one hundred, l'output dovrebbe essere one hundred one.


curl -X POST ${base_url}/v1beta3/tunedModels \
    -H 'Content-Type: application/json' \
    -H "Authorization: Bearer ${access_token}" \
    -H "x-goog-user-project: ${project_id}" \
    -d '
      {
        "display_name": "number generator model",
        "base_model": "models/text-bison-001",
        "tuning_task": {
          "hyperparameters": {
            "batch_size": 2,
            "learning_rate": 0.001,
            "epoch_count":3,
          },
          "training_data": {
            "examples": {
              "examples": [
                {
                    "text_input": "1",
                    "output": "2",
                },{
                    "text_input": "3",
                    "output": "4",
                },{
                    "text_input": "-3",
                    "output": "-2",
                },{
                    "text_input": "twenty two",
                    "output": "twenty three",
                },{
                    "text_input": "two hundred",
                    "output": "two hundred one",
                },{
                    "text_input": "ninety nine",
                    "output": "one hundred",
                },{
                    "text_input": "8",
                    "output": "9",
                },{
                    "text_input": "-98",
                    "output": "-97",
                },{
                    "text_input": "1,000",
                    "output": "1,001",
                },{
                    "text_input": "10,100,000",
                    "output": "10,100,001",
                },{
                    "text_input": "thirteen",
                    "output": "fourteen",
                },{
                    "text_input": "eighty",
                    "output": "eighty one",
                },{
                    "text_input": "one",
                    "output": "two",
                },{
                    "text_input": "three",
                    "output": "four",
                },{
                    "text_input": "seven",
                    "output": "eight",
                }
              ]
            }
          }
        }
      }' | tee tunemodel.json
{
  "name": "tunedModels/number-generator-model-q2d0uism5ivd/operations/xvyx09sjxlmh",
  "metadata": {
    "@type": "type.googleapis.com/google.ai.generativelanguage.v1beta3.CreateTunedModelMetadata",
    "totalSteps": 23,
    "tunedModel": "tunedModels/number-generator-model-q2d0uism5ivd"
  }
}
% Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  2277    0   297  100  1980    146    975  0:00:02  0:00:02 --:--:--  1121

Ottieni stato del modello ottimizzato

Lo stato del modello è impostato su CREATING durante l'addestramento e passerà in ACTIVE al termine dell'addestramento.

Di seguito è riportato un codice Python per analizzare il nome del modello generato dal codice JSON della risposta. Se esegui questo comando in un terminale, puoi provare a utilizzare un parser JSON bash per analizzare la risposta.

import json

first_page = json.load(open('tunemodel.json'))
os.environ['modelname'] = first_page['metadata']['tunedModel']

print(os.environ['modelname'])
tunedModels/number-generator-model-q2d0uism5ivd

Esegui un'altra richiesta GET con il nome del modello per ottenere i metadati del modello che includono il campo dello stato.


curl -X GET ${base_url}/v1beta3/${modelname} \
    -H 'Content-Type: application/json' \
    -H "Authorization: Bearer ${access_token}" \
    -H "x-goog-user-project: ${project_id}" \ | grep state
"state": "CREATING",
% Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100   494    0   494    0     0    760      0 --:--:-- --:--:-- --:--:--   760
curl: (3) URL using bad/illegal format or missing URL

Esegui inferenza

Una volta terminato il job di ottimizzazione, puoi utilizzarlo per generare testo con il servizio di testo.


curl -X POST ${base_url}/v1beta3/${modelname}:generateText \
    -H 'Content-Type: application/json' \
    -H "Authorization: Bearer ${access_token}" \
    -H "x-goog-user-project: ${project_id}" \
    -d '{
        "prompt": {
              "text": "4"
              },
        "temperature": 1.0,
        "candidate_count": 2}' | grep output
"output": "3 2 1",
      "output": "3 2",
% Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  1569    0  1447  100   122    183     15  0:00:08  0:00:07  0:00:01   310

L'output del modello potrebbe essere corretto. Se le prestazioni del modello ottimizzato non soddisfano gli standard richiesti, puoi provare ad aggiungere altri esempi di alta qualità, a modificare gli iperparametri o ad aggiungere un preambolo agli esempi. Puoi anche creare un altro modello ottimizzato in base al primo che hai creato.

Consulta la guida all'ottimizzazione per ulteriori indicazioni su come migliorare le prestazioni.

Chiama l'API REST con le richieste Python

Puoi chiamare l'API REST con qualsiasi libreria che ti consenta di inviare richieste http. La serie di esempi successiva utilizza la libreria di richieste Python e mostra alcune delle funzionalità più avanzate.

Imposta variabili

access_token = !gcloud auth application-default print-access-token
access_token = '\n'.join(access_token)

project = 'project-id'
base_url = "https://generativelanguage.googleapis.com"

Importa la libreria requests.

import requests
import json

Elenco modelli ottimizzati

Verifica la configurazione dell'autenticazione elencando i modelli ottimizzati attualmente disponibili.

headers={
  'Authorization': 'Bearer ' + access_token,
  'Content-Type': 'application/json',
  'x-goog-user-project': project
}

result = requests.get(
  url=f'{base_url}/v1beta3/tunedModels',
  headers = headers,
)
result.json()
{'tunedModels': [{'name': 'tunedModels/testnumbergenerator-fvitocr834l6',
   'baseModel': 'models/text-bison-001',
   'displayName': 'test_number_generator',
   'description': '{"description":"generates the  next number in the sequence given the input text","exampleInput":"input: 1","exampleOutput":"output: 2","datasourceUrl":"https://drive.google.com/open?id=11Pdm6GNom4vlBMUHwO6yFjGQT3t1yi44WVShXMFnkVA&authuser=0&resourcekey=0-2d17tccbdBoThXMkNDvtag","showedTuningComplete":false}',
   'state': 'ACTIVE',
   'createTime': '2023-09-18T11:06:39.092786Z',
   'updateTime': '2023-09-18T11:07:24.198359Z',
   'tuningTask': {'startTime': '2023-09-18T11:06:39.461814784Z',
    'completeTime': '2023-09-18T11:07:24.198359Z',
    'snapshots': [{'step': 1,
      'meanLoss': 16.613504,
      'computeTime': '2023-09-18T11:06:44.532937624Z'},
     {'step': 2,
      'epoch': 1,
      'meanLoss': 20.299532,
      'computeTime': '2023-09-18T11:06:47.825134421Z'},
     {'step': 3,
      'epoch': 1,
      'meanLoss': 8.169708,
      'computeTime': '2023-09-18T11:06:50.580344344Z'},
     {'step': 4,
      'epoch': 2,
      'meanLoss': 3.7588992,
      'computeTime': '2023-09-18T11:06:53.219133748Z'},
     {'step': 5,
      'epoch': 3,
      'meanLoss': 2.0643115,
      'computeTime': '2023-09-18T11:06:55.828458606Z'},
     {'step': 6,
      'epoch': 3,
      'meanLoss': 1.9765375,
      'computeTime': '2023-09-18T11:06:58.426053772Z'},
     {'step': 7,
      'epoch': 4,
      'meanLoss': 0.9276156,
      'computeTime': '2023-09-18T11:07:01.231832398Z'},
     {'step': 8,
      'epoch': 5,
      'meanLoss': 1.8424839,
      'computeTime': '2023-09-18T11:07:03.822710074Z'},
     {'step': 9,
      'epoch': 5,
      'meanLoss': 1.1747926,
      'computeTime': '2023-09-18T11:07:06.441685551Z'},
     {'step': 10,
      'epoch': 6,
      'meanLoss': 0.3079359,
      'computeTime': '2023-09-18T11:07:08.793491157Z'},
     {'step': 11,
      'epoch': 7,
      'meanLoss': 0.543368,
      'computeTime': '2023-09-18T11:07:11.393264892Z'},
     {'step': 12,
      'epoch': 7,
      'meanLoss': 0.35068464,
      'computeTime': '2023-09-18T11:07:13.808021238Z'},
     {'step': 13,
      'epoch': 8,
      'meanLoss': 0.026032856,
      'computeTime': '2023-09-18T11:07:16.295972078Z'},
     {'step': 14,
      'epoch': 8,
      'meanLoss': 0.108341046,
      'computeTime': '2023-09-18T11:07:18.941247488Z'},
     {'step': 15,
      'epoch': 9,
      'meanLoss': 0.016470395,
      'computeTime': '2023-09-18T11:07:21.607654306Z'},
     {'step': 16,
      'epoch': 10,
      'meanLoss': 0.063049875,
      'computeTime': '2023-09-18T11:07:24.077271307Z'}],
    'hyperparameters': {'epochCount': 10,
     'batchSize': 16,
     'learningRate': 0.02} },
   'temperature': 0.7,
   'topP': 0.95,
   'topK': 40},
  {'name': 'tunedModels/my-display-name-81-9wpmc1m920vq',
   'baseModel': 'models/text-bison-tuning-test',
   'displayName': 'my display name 81',
   'state': 'ACTIVE',
   'createTime': '2023-09-18T22:02:08.690991Z',
   'updateTime': '2023-09-18T22:02:28.806318Z',
   'tuningTask': {'startTime': '2023-09-18T22:02:09.161100369Z',
    'completeTime': '2023-09-18T22:02:28.806318Z',
    'snapshots': [{'step': 1,
      'meanLoss': 7.2774773,
      'computeTime': '2023-09-18T22:02:12.453056368Z'},
     {'step': 2,
      'meanLoss': 6.1902447,
      'computeTime': '2023-09-18T22:02:13.789508217Z'},
     {'step': 3,
      'meanLoss': 5.5545835,
      'computeTime': '2023-09-18T22:02:15.136220505Z'},
     {'step': 4,
      'epoch': 1,
      'meanLoss': 7.9237704,
      'computeTime': '2023-09-18T22:02:16.474358517Z'},
     {'step': 5,
      'epoch': 1,
      'meanLoss': 7.6770706,
      'computeTime': '2023-09-18T22:02:17.758261108Z'},
     {'step': 6,
      'epoch': 1,
      'meanLoss': 7.378622,
      'computeTime': '2023-09-18T22:02:19.114072224Z'},
     {'step': 7,
      'epoch': 1,
      'meanLoss': 4.485537,
      'computeTime': '2023-09-18T22:02:20.927434115Z'},
     {'step': 8,
      'epoch': 2,
      'meanLoss': 6.815181,
      'computeTime': '2023-09-18T22:02:22.267906011Z'},
     {'step': 9,
      'epoch': 2,
      'meanLoss': 6.411363,
      'computeTime': '2023-09-18T22:02:24.078114085Z'},
     {'step': 10,
      'epoch': 2,
      'meanLoss': 8.585093,
      'computeTime': '2023-09-18T22:02:25.441598938Z'},
     {'step': 11,
      'epoch': 2,
      'meanLoss': 4.901249,
      'computeTime': '2023-09-18T22:02:27.108985392Z'},
     {'step': 12,
      'epoch': 3,
      'meanLoss': 7.073003,
      'computeTime': '2023-09-18T22:02:28.441662034Z'}],
    'hyperparameters': {'epochCount': 3,
     'batchSize': 4,
     'learningRate': 0.001} },
   'temperature': 0.7,
   'topP': 0.95,
   'topK': 40},
  {'name': 'tunedModels/number-generator-model-kctlevca1g3q',
   'baseModel': 'models/text-bison-tuning-test',
   'displayName': 'number generator model',
   'state': 'ACTIVE',
   'createTime': '2023-09-18T23:43:21.461545Z',
   'updateTime': '2023-09-18T23:43:49.205493Z',
   'tuningTask': {'startTime': '2023-09-18T23:43:21.542403958Z',
    'completeTime': '2023-09-18T23:43:49.205493Z',
    'snapshots': [{'step': 1,
      'meanLoss': 7.342065,
      'computeTime': '2023-09-18T23:43:23.356271969Z'},
     {'step': 2,
      'meanLoss': 7.255807,
      'computeTime': '2023-09-18T23:43:24.620248223Z'},
     {'step': 3,
      'meanLoss': 5.4591417,
      'computeTime': '2023-09-18T23:43:25.854505395Z'},
     {'step': 4,
      'meanLoss': 6.968665,
      'computeTime': '2023-09-18T23:43:27.138260198Z'},
     {'step': 5,
      'meanLoss': 4.578809,
      'computeTime': '2023-09-18T23:43:28.404943274Z'},
     {'step': 6,
      'meanLoss': 6.4862137,
      'computeTime': '2023-09-18T23:43:29.631624883Z'},
     {'step': 7,
      'meanLoss': 9.781939,
      'computeTime': '2023-09-18T23:43:30.801341449Z'},
     {'step': 8,
      'epoch': 1,
      'meanLoss': 5.990006,
      'computeTime': '2023-09-18T23:43:31.854703315Z'},
     {'step': 9,
      'epoch': 1,
      'meanLoss': 8.846312,
      'computeTime': '2023-09-18T23:43:33.075785103Z'},
     {'step': 10,
      'epoch': 1,
      'meanLoss': 6.1585655,
      'computeTime': '2023-09-18T23:43:34.310432174Z'},
     {'step': 11,
      'epoch': 1,
      'meanLoss': 4.7877502,
      'computeTime': '2023-09-18T23:43:35.381582526Z'},
     {'step': 12,
      'epoch': 1,
      'meanLoss': 9.660514,
      'computeTime': '2023-09-18T23:43:36.445446408Z'},
     {'step': 13,
      'epoch': 1,
      'meanLoss': 5.6482882,
      'computeTime': '2023-09-18T23:43:37.603237821Z'},
     {'step': 14,
      'epoch': 1,
      'meanLoss': 3.162092,
      'computeTime': '2023-09-18T23:43:38.671463397Z'},
     {'step': 15,
      'epoch': 2,
      'meanLoss': 6.322996,
      'computeTime': '2023-09-18T23:43:39.769742201Z'},
     {'step': 16,
      'epoch': 2,
      'meanLoss': 6.781,
      'computeTime': '2023-09-18T23:43:40.985967994Z'},
     {'step': 17,
      'epoch': 2,
      'meanLoss': 5.136773,
      'computeTime': '2023-09-18T23:43:42.235469710Z'},
     {'step': 18,
      'epoch': 2,
      'meanLoss': 7.2091155,
      'computeTime': '2023-09-18T23:43:43.415178581Z'},
     {'step': 19,
      'epoch': 2,
      'meanLoss': 7.7508755,
      'computeTime': '2023-09-18T23:43:44.775221774Z'},
     {'step': 20,
      'epoch': 2,
      'meanLoss': 8.144815,
      'computeTime': '2023-09-18T23:43:45.788824334Z'},
     {'step': 21,
      'epoch': 2,
      'meanLoss': 5.485137,
      'computeTime': '2023-09-18T23:43:46.812663998Z'},
     {'step': 22,
      'epoch': 2,
      'meanLoss': 3.709197,
      'computeTime': '2023-09-18T23:43:47.971764087Z'},
     {'step': 23,
      'epoch': 3,
      'meanLoss': 6.0069466,
      'computeTime': '2023-09-18T23:43:49.004191079Z'}],
    'hyperparameters': {'epochCount': 3,
     'batchSize': 2,
     'learningRate': 0.001} },
   'temperature': 0.7,
   'topP': 0.95,
   'topK': 40},
  {'name': 'tunedModels/my-display-name-81-r9wcuda14lyy',
   'baseModel': 'models/text-bison-tuning-test',
   'displayName': 'my display name 81',
   'state': 'ACTIVE',
   'createTime': '2023-09-18T23:52:06.980185Z',
   'updateTime': '2023-09-18T23:52:26.679601Z',
   'tuningTask': {'startTime': '2023-09-18T23:52:07.616953503Z',
    'completeTime': '2023-09-18T23:52:26.679601Z',
    'snapshots': [{'step': 1,
      'meanLoss': 7.2774773,
      'computeTime': '2023-09-18T23:52:10.278936662Z'},
     {'step': 2,
      'meanLoss': 6.2793097,
      'computeTime': '2023-09-18T23:52:11.630844790Z'},
     {'step': 3,
      'meanLoss': 5.540499,
      'computeTime': '2023-09-18T23:52:13.027840389Z'},
     {'step': 4,
      'epoch': 1,
      'meanLoss': 7.977523,
      'computeTime': '2023-09-18T23:52:14.368199020Z'},
     {'step': 5,
      'epoch': 1,
      'meanLoss': 7.6197805,
      'computeTime': '2023-09-18T23:52:15.872428752Z'},
     {'step': 6,
      'epoch': 1,
      'meanLoss': 7.3851357,
      'computeTime': '2023-09-18T23:52:17.213094182Z'},
     {'step': 7,
      'epoch': 1,
      'meanLoss': 4.5342345,
      'computeTime': '2023-09-18T23:52:19.090698421Z'},
     {'step': 8,
      'epoch': 2,
      'meanLoss': 6.8603754,
      'computeTime': '2023-09-18T23:52:20.494844731Z'},
     {'step': 9,
      'epoch': 2,
      'meanLoss': 6.418575,
      'computeTime': '2023-09-18T23:52:21.815997555Z'},
     {'step': 10,
      'epoch': 2,
      'meanLoss': 8.659064,
      'computeTime': '2023-09-18T23:52:23.524287192Z'},
     {'step': 11,
      'epoch': 2,
      'meanLoss': 4.856765,
      'computeTime': '2023-09-18T23:52:24.864661291Z'},
     {'step': 12,
      'epoch': 3,
      'meanLoss': 7.1078596,
      'computeTime': '2023-09-18T23:52:26.225055381Z'}],
    'hyperparameters': {'epochCount': 3,
     'batchSize': 4,
     'learningRate': 0.001} },
   'temperature': 0.7,
   'topP': 0.95,
   'topK': 40},
  {'name': 'tunedModels/number-generator-model-w1eabln5adwp',
   'baseModel': 'models/text-bison-tuning-test',
   'displayName': 'number generator model',
   'state': 'ACTIVE',
   'createTime': '2023-09-19T19:29:08.622497Z',
   'updateTime': '2023-09-19T19:29:46.063853Z',
   'tuningTask': {'startTime': '2023-09-19T19:29:08.806930486Z',
    'completeTime': '2023-09-19T19:29:46.063853Z',
    'snapshots': [{'step': 1,
      'meanLoss': 7.342065,
      'computeTime': '2023-09-19T19:29:13.023811994Z'},
     {'step': 2,
      'meanLoss': 7.1960244,
      'computeTime': '2023-09-19T19:29:14.844046282Z'},
     {'step': 3,
      'meanLoss': 5.480289,
      'computeTime': '2023-09-19T19:29:16.596884354Z'},
     {'step': 4,
      'meanLoss': 6.851822,
      'computeTime': '2023-09-19T19:29:17.741735378Z'},
     {'step': 5,
      'meanLoss': 4.5535283,
      'computeTime': '2023-09-19T19:29:18.914760812Z'},
     {'step': 6,
      'meanLoss': 6.449012,
      'computeTime': '2023-09-19T19:29:20.053316042Z'},
     {'step': 7,
      'meanLoss': 9.842458,
      'computeTime': '2023-09-19T19:29:21.371286675Z'},
     {'step': 8,
      'epoch': 1,
      'meanLoss': 5.9831877,
      'computeTime': '2023-09-19T19:29:22.915277044Z'},
     {'step': 9,
      'epoch': 1,
      'meanLoss': 8.936815,
      'computeTime': '2023-09-19T19:29:24.666461680Z'},
     {'step': 10,
      'epoch': 1,
      'meanLoss': 6.14651,
      'computeTime': '2023-09-19T19:29:26.793310451Z'},
     {'step': 11,
      'epoch': 1,
      'meanLoss': 4.853589,
      'computeTime': '2023-09-19T19:29:28.328297535Z'},
     {'step': 12,
      'epoch': 1,
      'meanLoss': 9.6831045,
      'computeTime': '2023-09-19T19:29:29.501236840Z'},
     {'step': 13,
      'epoch': 1,
      'meanLoss': 5.706586,
      'computeTime': '2023-09-19T19:29:30.612807978Z'},
     {'step': 14,
      'epoch': 1,
      'meanLoss': 3.276942,
      'computeTime': '2023-09-19T19:29:31.928747103Z'},
     {'step': 15,
      'epoch': 2,
      'meanLoss': 6.1736736,
      'computeTime': '2023-09-19T19:29:33.588699180Z'},
     {'step': 16,
      'epoch': 2,
      'meanLoss': 6.857398,
      'computeTime': '2023-09-19T19:29:35.239083809Z'},
     {'step': 17,
      'epoch': 2,
      'meanLoss': 5.098094,
      'computeTime': '2023-09-19T19:29:37.000705047Z'},
     {'step': 18,
      'epoch': 2,
      'meanLoss': 7.27724,
      'computeTime': '2023-09-19T19:29:38.532313231Z'},
     {'step': 19,
      'epoch': 2,
      'meanLoss': 7.6310735,
      'computeTime': '2023-09-19T19:29:39.696034301Z'},
     {'step': 20,
      'epoch': 2,
      'meanLoss': 8.152623,
      'computeTime': '2023-09-19T19:29:40.803342042Z'},
     {'step': 21,
      'epoch': 2,
      'meanLoss': 5.451577,
      'computeTime': '2023-09-19T19:29:42.445788199Z'},
     {'step': 22,
      'epoch': 2,
      'meanLoss': 3.7990716,
      'computeTime': '2023-09-19T19:29:43.866737307Z'},
     {'step': 23,
      'epoch': 3,
      'meanLoss': 6.120624,
      'computeTime': '2023-09-19T19:29:45.599248553Z'}],
    'hyperparameters': {'epochCount': 3,
     'batchSize': 2,
     'learningRate': 0.001} },
   'temperature': 0.7,
   'topP': 0.95,
   'topK': 40}]}

Crea modello ottimizzato

Come per l'esempio di Curl, passi nel set di dati attraverso il campo training_data.

operation = requests.post(
    url = f'{base_url}/v1beta3/tunedModels',
    headers=headers,
    json= {
        "display_name": "number generator",
        "base_model": "models/text-bison-001",
        "tuning_task": {
          "hyperparameters": {
            "batch_size": 4,
            "learning_rate": 0.001,
            "epoch_count":3,
          },
          "training_data": {
            "examples": {
              "examples": [
                {
                    'text_input': '1',
                    'output': '2',
                },{
                    'text_input': '3',
                    'output': '4',
                },{
                    'text_input': '-3',
                    'output': '-2',
                },{
                    'text_input': 'twenty two',
                    'output': 'twenty three',
                },{
                    'text_input': 'two hundred',
                    'output': 'two hundred one',
                },{
                    'text_input': 'ninety nine',
                    'output': 'one hundred',
                },{
                    'text_input': '8',
                    'output': '9',
                },{
                    'text_input': '-98',
                    'output': '-97',
                },{
                    'text_input': '1,000',
                    'output': '1,001',
                },{
                    'text_input': '10,100,000',
                    'output': '10,100,001',
                },{
                    'text_input': 'thirteen',
                    'output': 'fourteen',
                },{
                    'text_input': 'eighty',
                    'output': 'eighty one',
                },{
                    'text_input': 'one',
                    'output': 'two',
                },{
                    'text_input': 'three',
                    'output': 'four',
                },{
                    'text_input': 'seven',
                    'output': 'eight',
                }
              ]
            }
          }
        }
      }
)
operation
<Response [200]>
operation.json()
{'name': 'tunedModels/number-generator-ncqqnysl74dt/operations/qqlbwzfyzn0k',
 'metadata': {'@type': 'type.googleapis.com/google.ai.generativelanguage.v1beta3.CreateTunedModelMetadata',
  'totalSteps': 12,
  'tunedModel': 'tunedModels/number-generator-ncqqnysl74dt'} }

Imposta una variabile con il nome del modello ottimizzato da utilizzare per le altre chiamate.

name=operation.json()["metadata"]["tunedModel"]
name
'tunedModels/number-generator-ncqqnysl74dt'

Ottieni stato del modello ottimizzato

Puoi controllare l'avanzamento del job di ottimizzazione controllando il campo dello stato. CREATING significa che il job di ottimizzazione è ancora in corso, mentre ACTIVE significa che i addestramento sono stati completati e il modello ottimizzato è pronto per l'uso.

tuned_model = requests.get(
    url = f'{base_url}/v1beta3/{name}',
    headers=headers,
)
tuned_model.json()
{'name': 'tunedModels/number-generator-ncqqnysl74dt',
 'baseModel': 'models/text-bison-001',
 'displayName': 'number generator',
 'state': 'CREATING',
 'createTime': '2023-09-19T19:56:25.999303Z',
 'updateTime': '2023-09-19T19:56:25.999303Z',
 'tuningTask': {'startTime': '2023-09-19T19:56:26.297862545Z',
  'hyperparameters': {'epochCount': 3, 'batchSize': 4, 'learningRate': 0.001} },
 'temperature': 0.7,
 'topP': 0.95,
 'topK': 40}

Il codice seguente controlla il campo dello stato ogni 5 secondi finché non è più nello stato CREATING.

import time
import pprint

op_json = operation.json()
response = op_json.get('response')
error = op_json.get('error')

while response is None and error is None:
    time.sleep(31)

    operation = requests.get(
        url = f'{base_url}/v1/{op_json["name"]}',
        headers=headers,
    )

    op_json = operation.json()
    response = op_json.get('response')
    error = op_json.get('error')

    percent = op_json['metadata'].get('completedPercent')
    if percent is not None:
      print(f"{percent:.2f}% - {op_json['metadata']['snapshots'][-1]}")
      print()

if error is not None:
    raise Exception(error)
21.28% - {'step': 40, 'epoch': 10, 'meanLoss': 2.4871845, 'computeTime': '2023-09-20T00:23:55.255785843Z'}

21.28% - {'step': 40, 'epoch': 10, 'meanLoss': 2.4871845, 'computeTime': '2023-09-20T00:23:55.255785843Z'}

43.09% - {'step': 81, 'epoch': 21, 'meanLoss': 0.032220088, 'computeTime': '2023-09-20T00:24:56.302837803Z'}

43.09% - {'step': 81, 'epoch': 21, 'meanLoss': 0.032220088, 'computeTime': '2023-09-20T00:24:56.302837803Z'}

63.83% - {'step': 120, 'epoch': 32, 'meanLoss': 0.0030430648, 'computeTime': '2023-09-20T00:25:57.228615435Z'}

63.83% - {'step': 120, 'epoch': 32, 'meanLoss': 0.0030430648, 'computeTime': '2023-09-20T00:25:57.228615435Z'}

85.11% - {'step': 160, 'epoch': 42, 'meanLoss': -1.1145603e-06, 'computeTime': '2023-09-20T00:26:57.819011896Z'}

100.00% - {'step': 188, 'epoch': 50, 'meanLoss': 0.00040101097, 'computeTime': '2023-09-20T00:27:40.024132813Z'}

Esegui inferenza

Una volta terminato il job di ottimizzazione, puoi utilizzarlo per generare testo nello stesso modo in cui useresti il modello di testo di base.

import time

m = requests.post(
    url = f'{base_url}/v1beta3/{name}:generateText',
    headers=headers,
    json= {
         "prompt": {
              "text": "9"
              },
    })
import pprint
print(m.json()['candidates'][0]['output'])
9

L'output del modello potrebbe essere corretto. Se le prestazioni del modello ottimizzato non soddisfano gli standard richiesti, puoi provare ad aggiungere altri esempi di alta qualità, a modificare gli iperparametri o ad aggiungere un preambolo agli esempi.

Passaggi successivi