diff --git a/README.md b/README.md index 5306161..494e2c0 100644 --- a/README.md +++ b/README.md @@ -78,7 +78,31 @@ print(res.choices[0].message.content) ``` -更多功能介绍，请参阅 [PAI Python SDK文档](https://alipai.readthedocs.io/) 。 +- 微调预训练模型 + +通过PAI提供的微调脚本，提交一个模型微调任务 + +```python + +from pai.model import ModelTrainingRecipe + +training_recipe = ModelTrainingRecipe( + model_name="qwen2-0.5b-instruct", + model_provider="pai", + instance_type="ecs.gn6e-c12g1.3xlarge", +) + +training_recipe.train( + inputs={ + # 本地或是阿里云OSS上的数据路径(oss:///path/to/data) + "train": "" + } +) + + +``` + +通过访问PAI提供的示例仓库，可以了解更多使用示例：[pai-examples](https://github.com/aliyun/pai-examples/tree/master/pai-python-sdk) ## 🤝 贡献代码 diff --git a/README_EN.md b/README_EN.md index 833f99d..4268bcb 100644 --- a/README_EN.md +++ b/README_EN.md @@ -81,7 +81,38 @@ print(res.choices[0].message.content) ``` -For more details, please refer to the [PAI Python SDK Documentation](https://alipai.readthedocs.io/). +- Fine-tune the pretrained model +- +Submit a model fine-tuning task using the fine-tuning script provided by PAI. + +```python + +from pai.model import ModelTrainingRecipe + +# Retrieve the Qwen2-0.5b-instruct model training recipe provided by PAI +training_recipe = ModelTrainingRecipe( + model_name="qwen2-0.5b-instruct", + model_provider="pai", + instance_type="ecs.gn6e-c12g1.3xlarge", +) + +# Submit the training job +job = training_recipe.train( + inputs={ + # Data path on local or Alibaba Cloud OSS (oss:///path/to/data) + "train": "" + } +) + +# Get output model path +print(training_recipe.model_data()) + +# Deploy the fine-tuned model +predictor = training_recipe.deploy(service_name="qwen2_finetune") + +``` + +You can learn more usage examples by visiting the PAI example repository: [pai-examples](https://github.com/aliyun/pai-examples/tree/master/pai-python-sdk) ## 🤝 Contributing diff --git a/docs/source/index.rst b/docs/source/index.rst index f426f83..1ba224f 100644 --- a/docs/source/index.rst +++ b/docs/source/index.rst @@ -30,16 +30,6 @@ PAI Python SDK 文档 user-guide/processing-job -.. toctree:: - :maxdepth: 1 - :caption: 示例教程 - - tutorial/framework - tutorial/train - tutorial/predict - tutorial/advance - - .. toctree:: :maxdepth: 1 :caption: Reference diff --git a/docs/source/user-guide/pretrained-model.rst b/docs/source/user-guide/pretrained-model.rst index b32c4b7..cf3f96c 100644 --- a/docs/source/user-guide/pretrained-model.rst +++ b/docs/source/user-guide/pretrained-model.rst @@ -87,113 +87,46 @@ PAI公共仓库中的部分模型，也提供了微调训练算法，支持用 .. code-block:: python - from pai.model import RegisteredModel - from pai.estimator import AlgorithmEstimator + from pai.model import RegisteredModel, ModelTrainingRecipe # 获取PAI提供的Bert模型 m = RegisteredModel("bert-base-uncased", model_provider="pai") - # 获取模型的微调训练算法 - est: AlgorithmEstimator = m.get_estimator() - - # 查看算法的超参数定义描述、输入定义描述，以及输出定义描述。 - print(est.hyperparameter_definitions) - # [{'DefaultValue': '1', - # 'Type': 'Int', - # 'Description': 'Number of epochs to train the model. Each epoch is one complete iteration over the entire training dataset.', - # 'Required': True, - # 'Name': 'max_epochs'}, - # {'DefaultValue': '16', - # 'Type': 'Int', - # 'Description': 'Number of samples that will be propagated through the model. A higher value might consume more memory.', - # 'Required': False, - # 'Name': 'batch_size'}, - # {'DefaultValue': '0.00001', - # 'Type': 'Float', - # 'Description': 'The initial learning rate to be used for training. A higher value usually implies more aggression in gradient updates.', - # 'Required': False, - # 'Name': 'learning_rate'}, - # {'DefaultValue': '2000', - # 'Type': 'Int', - # 'Description': 'Number of updates steps before two checkpoint.', - # 'Required': False, - # 'Name': 'save_steps'} - # ] - print(est.input_channel_definitions) - # [{'Description': 'Input channel for pretrained model to be fine-tuned on.', - # 'Required': True, - # 'SupportedChannelTypes': ['oss'], - # 'Properties': {'ResourceUse': 'Base', 'ResourceType': 'Model'}, - # 'Name': 'model'}, - # {'Description': 'Input channel for training dataset.', - # 'Required': True, - # 'SupportedChannelTypes': ['oss'], - # 'Properties': {'ResourceUse': 'Train', 'ResourceType': 'Dataset'}, - # 'Name': 'train'}, - # {'Description': 'Input channel for validation dataset.', - # 'Required': False, - # 'SupportedChannelTypes': ['oss'], - # 'Properties': {'ResourceUse': 'Validation', 'ResourceType': 'Dataset'}, - # 'Name': 'validation'}] - - - # 查看算法的默认输入，包含了预训练模型，训练数据，验证数据等 - training_inputs = m.get_estimator_inputs() - print(training_inputs) - # { - # 'model': 'oss://pai-quickstart-cn-hangzhou.oss-cn-hangzhou.aliyuncs.com/huggingface/models/bert-base-uncased/main/', - # 'train': 'oss://pai-quickstart-cn-hangzhou.oss-cn-hangzhou.aliyuncs.com/huggingface/datasets/sst2/main/train.json', - # 'validation': 'oss://pai-quickstart-cn-hangzhou.oss-cn-hangzhou.aliyuncs.com/huggingface/datasets/sst2/main/validation.json' - # } - - # 使用默认输入进行微调训练 - est.fit(inputs=training_inputs) - - # 查看训练输出的模型，默认模型存储在OSS URI上 - print(est.model_data()) - - -以上的训练任务中，我们使用了PAI提供的公共数据集，对模型进行微调训练。当用户需要使用自己的数据集进行微调训练时，需要先将数据准备到OSS，或是NAS上，然后将数据的OSS或是NAS路径，作为训练任务的输入。 - - -使用用户训练数据集提交训练任务： - -.. code-block:: python - - from pai.estimator import AlgorithmEstimator - - # 获取模型的微调训练算法 - est: AlgorithmEstimator = m.get_estimator() - # 配置修改提交的训练算法超参，具体的超参用途可以查看 est.hyperparameter_definitions 中的描述. - est.hyperparameters = { - 'max_epochs': 1, - 'batch_size': 8, - 'learning_rate': 2e-05, - 'save_steps': 2000 - } - - # 默认的训练输入 - default_training_inputs = m.get_estimator_inputs() - # 使用用户的数据集进行微调训练 - training_inputs = { - # 使用PAI提供预训练模型作为基础模型输入 - "model": default_training_inputs["model"], - # 使用用户的训练和测试数据集 - "train": "oss:///my-dataset/train.json", - "validation": "oss:///my-dataset/validation.json" - } - - est.fit(inputs=training_inputs) - -用户可以通过模型卡片上的文档，查看模型的微调训练数据格式。同时也可以参考相应的模型微调训练的默认输入数据格式，进行数据的准备。 - -下载PAI数据集到本地目录: - -.. code-block:: python + training_recipe = m.training_recipe() + + training_recipe = ModelTrainingRecipe( + model_name = "bert-base-uncased", + model_provider = "pai", + instance_type = "ecs.c6.xlarge", + # 训练任务的超参数 + hyperparameters={ + "max_epochs": 1, + "learning_rate": 0.00001, + "batch_size": 16, + "save_steps": 2000, + }, + ) - from pai.common.oss_util import download + # 查看模型微调算法输入定义 + print(training_recipe.input_channels) + # 查看模型微调算法超参数定义 + print(training_recipe.hyperparameter_definitions) + # 查看默认训练输入数据 + print(training_recipe.default_inputs) + + # 提交微调训练作业 + job = training_recipe.train( + job_name="train_recipe_example", + # 配置使用用户在OSS Bucket上的数据作为训练数据 + # inputs={ + # "train": "oss:///" + # } + ) + # 获取微调后模型路径 + print(training_recipe.model_data()) - # 默认的训练输入 - default_training_inputs = m.get_estimator_inputs() + # 使用PAI提供的推理服务配置部署模型 + predictor = training_recipe.deploy( + service_name="bert_example", + ) - # 下载PAI提供的公共训练数据到本地 - download(default_training_inputs["train"], "./train/") +用户可以通过PAI ModelGallery提供的模型卡片上的文档，查看具体模型模型的微调训练数据格式。