在 ai.google.dev 上查看
试用 Colab 笔记本
在 GitHub 上查看笔记本
下载笔记本在本笔记中，您将学习如何开始使用 PaLM API 调优服务，使用 curl 命令或 Python 请求 API 调用 PaLM REST API。在这里，您将学习如何调整 PaLM API 的文本生成服务背后的文本模型。
设置
身份验证
借助 PaLM API，您可以使用自己的数据对模型进行调参。由于这是您的数据和经过调优的模型，因此需要比 API 密钥提供的更严格的访问权限控制。
在运行本教程之前，您需要为项目设置 OAuth。
如果要在 Colab 中运行此笔记本，请先使用“文件”>“上传”选项上传 client_secret*.json 文件。
cp client_secret*.json client_secret.jsonls
client_secret.json
此 gcloud 命令会将 client_secret.json 文件转换为可用于向服务进行身份验证的凭据。
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'
使用 CURL 调用 REST API
本部分提供了用于调用 REST API 的 curl 语句示例。您将学习如何创建调优作业、检查其状态，以及在完成后进行推理调用。
设置变量
为周期性值设置变量，以便在其余 REST API 调用中使用。该代码使用 Python os 库设置可在所有代码单元格中访问的环境变量。
这仅适用于 Colab 笔记本环境。下一个代码单元中的代码相当于在 bash 终端中运行以下命令。
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"
列出已调参的模型
列出当前可用的经过调整的模型，以验证您的身份验证设置。
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
创建经调参的模型
如需创建经过调优的模型，您需要将数据集传递给 training_data 字段中的模型。
在此示例中，您将调整模型以生成序列中的下一个数字。例如，如果输入为 1，则模型应输出 2。如果输入为 one hundred，则输出应为 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
获取经调参的模型状态
模型的状态在训练期间设为 CREATING，训练完成后会变为 ACTIVE。
下面的一些 Python 代码用于从响应 JSON 中解析生成的模型名称。如果您在终端中运行此命令，可以尝试使用 bash JSON 解析器来解析响应。
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
使用模型名称发出另一个 GET 请求，以获取包含状态字段的模型元数据。
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
运行推断
优化作业完成后，您就可以使用它通过文本服务生成文本了。
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
模型的输出不一定正确。如果经过调整的模型的效果不符合您的要求，您可以尝试添加更多高质量示例、调整超参数或为示例添加前言。您甚至可以根据第一个创建的模型创建另一个经过调优的模型。
如需有关提升性能的更多指导，请参阅调整指南。
使用 Python 请求调用 REST API
您可以使用任何允许您发送 http 请求的库来调用其余 API。下一组示例使用 Python requests 库，并演示了一些更高级的功能。
设置变量
access_token = !gcloud auth application-default print-access-token
access_token = '\n'.join(access_token)
project = 'project-id'
base_url = "https://generativelanguage.googleapis.com"
导入 requests 库。
import requests
import json
列出已调参的模型
列出当前可用的经过调整的模型，以验证您的身份验证设置。
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}]}
创建经调参的模型
与 Curl 示例一样，您可以通过 training_data 字段传入数据集。
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'} }
设置一个变量，并为其赋予经过调优的模型的名称，以便在其余调用中使用。
name=operation.json()["metadata"]["tunedModel"]
name
'tunedModels/number-generator-ncqqnysl74dt'
获取经过调优的模型状态
您可以通过查看状态字段来了解调优作业的进度。CREATING 表示调优作业仍在进行中，ACTIVE 表示训练已完成，调优后的模型可以使用。
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}
以下代码会每 5 秒检查一次状态字段，直到其不再处于 CREATING 状态。
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'}
运行推断
调优作业完成后，您可以像使用基础文本模型一样使用它来生成文本。
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
模型的输出不一定正确。如果经过调整的模型的效果不符合您的要求，您可以尝试添加更多高质量示例、调整超参数或为示例添加前言。
后续步骤
- 如需开始使用优化服务编写代码，请参阅使用 Python 的优化快速入门。
- 如需详细了解如何根据您的使用场景优化模型，请参阅调参指南。