以PyTorch框架创建训练作业(新版训练)
本节通过调用一系列API,以训练模型为例介绍ModelArts API的使用流程。
概述
使用PyTorch框架创建训练作业的流程如下:
- 调用认证鉴权接口获取用户Token,在后续的请求中需要将Token放到请求消息头中作为认证。
- 调用获取训练作业支持的公共规格接口获取训练作业支持的资源规格。
- 调用获取训练作业支持的AI预置框架接口查看训练作业支持的引擎类型和版本。
- 调用创建算法接口创建一个算法,记录算法id。
- 调用创建训练作业接口使用刚创建的算法返回的uuid创建一个训练作业,记录训练作业id。
- 调用查询训练作业详情接口使用刚创建的训练作业返回的id查询训练作业状态。
- 调用查询训练作业指定任务的日志(OBS链接)接口获取训练作业日志的对应的obs路径。
- 调用查询训练作业指定任务的运行指标接口查看训练作业的运行指标详情。
- 当训练作业使用完成或不再需要时,调用删除训练作业接口删除训练作业。
前提条件
- 已获取IAM的EndPoint和ModelArts的EndPoint。
- 确认服务的部署区域,获取项目ID和名称、获取账号名和ID和获取用户名和用户ID。
- 已准备好PyTorch框架的训练代码,例如将启动文件“test-pytorch.py”存放在OBS的“obs://cnnorth4-job-test-v2/pytorch/fast_example/code/cpu”目录下。
- 已经准备好训练作业的数据文件,例如将训练数据集存放在OBS的“obs://cnnorth4-job-test-v2/pytorch/fast_example/data”目录下。
- 已经创建好训练作业的模型输出位置,例如“obs://cnnorth4-job-test-v2/pytorch/fast_example/outputs”。
- 已经创建好训练作业的日志输出位置,例如“obs://cnnorth4-job-test-v2/pytorch/fast_example/log”。
操作步骤
- 调用认证鉴权接口获取用户的Token。
- 请求消息体:
URI格式:POST https://{iam_endpoint}/v3/auth/tokens
请求消息头:Content-Type →application/json
请求Body:{ "auth": { "identity": { "methods": ["password"], "password": { "user": { "name": "user_name", "password": "user_password", "domain": { "name": "domain_name" } } } }, "scope": { "project": { "name": "ap-southeast-1" } } } }
其中,加粗的斜体字段需要根据实际值填写:- iam_endpoint为IAM的终端节点。
- user_name为IAM用户名。
- user_password为用户登录密码。
- domain_name为用户所属的账号名。
- ap-southeast-1为项目名,代表服务的部署区域。
- 返回状态码“201 Created”,在响应Header中获取“X-Subject-Token”的值即为Token,如下所示:
x-subject-token →MIIZmgYJKoZIhvcNAQcCoIIZizCCGYcCAQExDTALBglghkgBZQMEAgEwgXXXXXX...
- 请求消息体:
- 调用获取训练作业支持的公共规格接口获取训练作业支持的资源规格。
- 请求消息体:
URI格式:GET https://{ma_endpoint}/v2/{project_id}/ training-job-flavors? flavor_type=CPU
请求消息头:X-Auth-Token →MIIZmgYJKoZIhvcNAQcCoIIZizCCGYcCAQExDTALBglghkgBZQMEAgEwgXXXXXX...
其中,加粗的斜体字段需要根据实际值填写:
- ma_endpoint为ModelArts的终端节点。
- project_id为用户的项目ID。
- “X-Auth-Token”的值是上一步获取到的Token值。
- 返回状态码“200”,响应Body如下所示:
{ "total_count": 2, "flavors": [ { "flavor_id": "modelarts.vm.cpu.2u", "flavor_name": "Computing CPU(2U) instance", "flavor_type": "CPU", "billing": { "code": "modelarts.vm.cpu.2u", "unit_num": 1 }, "flavor_info": { "max_num": 1, "cpu": { "arch": "x86", "core_num": 2 }, "memory": { "size": 8, "unit": "GB" }, "disk": { "size": 50, "unit": "GB" } } }, { "flavor_id": "modelarts.vm.cpu.8u", "flavor_name": "Computing CPU(8U) instance", "flavor_type": "CPU", "billing": { "code": "modelarts.vm.cpu.8u", "unit_num": 1 }, "flavor_info": { "max_num": 16, "cpu": { "arch": "x86", "core_num": 8 }, "memory": { "size": 32, "unit": "GB" }, "disk": { "size": 50, "unit": "GB" } } } ] }
- 根据“flavor_id”字段选择并记录创建训练作业时需要的规格类型,本章以“modelarts.vm.cpu.8u”为例,并记录“max_num”字段的值为“16”。
- 请求消息体:
- 调用获取训练作业支持的AI预置框架接口查看训练作业的引擎类型和版本。
- 请求消息体:
URI格式:GET https://{ma_endpoint}/v2/{project_id}/job/ training-job-engines
请求消息头:
X-Auth-Token→MIIZmgYJKoZIhvcNAQcCoIIZizCCGYcCAQExDTALBglghkgBZQMEAgEwgXXXXXX...
Content-Type →application/json
其中,加粗的斜体字段需要根据实际值填写。
- 返回状态码“200”,响应Body如下所示(引擎较多,只展示部分):
{ "total": 28, "items": [ ...... { "engine_id": "mindspore_1.6.0-cann_5.0.3.6-py_3.7-euler_2.8.3-aarch64", "engine_name": "Ascend-Powered-Engine", "engine_version": "mindspore_1.6.0-cann_5.0.3.6-py_3.7-euler_2.8.3-aarch64", "v1_compatible": false, "run_user": "1000", "image_info": { "cpu_image_url": "", "gpu_image_url": "atelier/mindspore_1_6_0:train", "image_version": "mindspore_1.6.0-cann_5.0.3.6-py_3.7-euler_2.8.3-aarch64-snt9-roma-20211231193205-33131ee" } }, ...... { "engine_id": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64", "engine_name": "PyTorch", "engine_version": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64", "tags": [ { "key": "auto_search", "value": "True" } ], "v1_compatible": false, "run_user": "1102", "image_info": { "cpu_image_url": "aip/pytorch_1_8:train", "gpu_image_url": "aip/pytorch_1_8:train", "image_version": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64-20210912152543-1e0838d" } }, ...... { "engine_id": "tensorflow_2.1.0-cuda_10.1-py_3.7-ubuntu_18.04-x86_64", "engine_name": "TensorFlow", "engine_version": "tensorflow_2.1.0-cuda_10.1-py_3.7-ubuntu_18.04-x86_64", "tags": [ { "key": "auto_search", "value": "True" } ], "v1_compatible": false, "run_user": "1102", "image_info": { "cpu_image_url": "aip/tensorflow_2_1:train", "gpu_image_url": "aip/tensorflow_2_1:train", "image_version": "tensorflow_2.1.0-cuda_10.1-py_3.7-ubuntu_18.04-x86_64-20210912152543-1e0838d" } }, ...... ] }
根据“engine_name”和“engine_version”字段选择创建训练作业时需要的引擎规格,并记录对应的“engine_name”和“engine_version”,本章以Pytorch引擎为例创建作业,记录“engine_name”为“PyTorch”,“engine_version”为“pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64”。
- 请求消息体:
- 调用创建算法接口创建一个算法,记录算法id。
- 请求消息体:
URI格式:POST https://{ma_endpoint}/v2/{project_id}/ algorithms
请求消息头:
X-Auth-Token→MIIZmgYJKoZIhvcNAQcCoIIZizCCGYcCAQExDTALBglghkgBZQMEAgEwgXXXXXX...
Content-Type →application/json
其中,加粗的斜体字段需要根据实际值填写。
请求body:
{ "metadata": { "name": "test-pytorch-cpu", "description": "test pytorch job in cpu in mode gloo" }, "job_config": { "boot_file": "/cnnorth4-job-test-v2/pytorch/fast_example/code/cpu/test-pytorch.py", "code_dir": "/cnnorth4-job-test-v2/pytorch/fast_example/code/cpu/", "engine": { "engine_name": "PyTorch", "engine_version": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64" }, "inputs": [{ "name": "data_url", "description": "数据来源1" }], "outputs": [{ "name": "train_url", "description": "输出数据1" }], "parameters": [{ "name": "dist", "description": "", "value": "False", "constraint": { "editable": true, "required": false, "sensitive": false, "type": "Boolean", "valid_range": [], "valid_type": "None" } }, { "name": "world_size", "description": "", "value": "1", "constraint": { "editable": true, "required": false, "sensitive": false, "type": "Integer", "valid_range": [], "valid_type": "None" } } ], "parameters_customization": true }, "resource_requirements": [] }
其中,加粗的斜体字段需要根据实际值填写:
- “metadata”字段下的“name”和“description”分别为算法的名称和描述。
- “job_config”字段下的“code_dir”和“boot_file”分别为算法的代码目录和代码启动文件。代码目录为代码启动文件的一级目录。
- “job_config”字段下的“inputs”和“outputs”分别为算法的输入输出管道。可以按照实例指定“data_url”和“train_url”,在代码中解析超参分别指定训练所需要的数据文件本地路径和训练生成的模型输出本地路径。
- “job_config”字段下的“parameters_customization”表示是否支持自定义超参,此处填true。
- “job_config”字段下的“parameters”表示算法本身的超参。“name”填写超参名称(64个以内字符,仅支持大小写字母、数字、下划线和中划线),“value”填写超参的默认值,“constraint”填写超参的约束,例如此处“type”填写“String”(支持String、Integer、Float和Boolean),“editable”填写“true”,“required”填写“false”等。
- “job_config”字段下的“engine”表示算法所依赖的引擎,使用3记录的“engine_name”和“engine_version”。
- 返回状态码“200 OK”,表示创建算法成功,响应Body如下所示:
{ "metadata": { "id": "01c399ae-8593-4ef5-9e4d-085950aacde1", "name": "test-pytorch-cpu", "description": "test pytorch job in cpu in mode gloo", "create_time": 1641890623262, "workspace_id": "0", "ai_project": "default-ai-project", "user_name": "", "domain_id": "0659fbf6de00109b0ff1c01fc037d240", "source": "custom", "api_version": "", "is_valid": true, "state": "", "size": 4790, "tags": null, "attr_list": null, "version_num": 0, "update_time": 0 }, "share_info": {}, "job_config": { "code_dir": "/cnnorth4-job-test-v2/pytorch/fast_example/code/cpu/", "boot_file": "/cnnorth4-job-test-v2/pytorch/fast_example/code/cpu/test-pytorch.py", "parameters": [ { "name": "dist", "description": "", "i18n_description": null, "value": "False", "constraint": { "type": "Boolean", "editable": true, "required": false, "sensitive": false, "valid_type": "None", "valid_range": [] } }, { "name": "world_size", "description": "", "i18n_description": null, "value": "1", "constraint": { "type": "Integer", "editable": true, "required": false, "sensitive": false, "valid_type": "None", "valid_range": [] } } ], "parameters_customization": true, "inputs": [ { "name": "data_url", "description": "数据来源1" } ], "outputs": [ { "name": "train_url", "description": "输出数据1" } ], "engine": { "engine_id": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64", "engine_name": "PyTorch", "engine_version": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64", "tags": [ { "key": "auto_search", "value": "True" } ], "v1_compatible": false, "run_user": "1102", "image_info": { "cpu_image_url": "aip/pytorch_1_8:train", "gpu_image_url": "aip/pytorch_1_8:train", "image_version": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64-20210912152543-1e0838d" } }, "code_tree": { "name": "cpu/", "children": [ { "name": "test-pytorch.py" } ] } }, "resource_requirements": [], "advanced_config": {} }
记录“metadata”字段下的“id”(算法id,32位UUID)字段的值便于后续步骤使用。
- 请求消息体:
- 调用创建训练作业接口使用刚创建的算法返回的uuid创建一个训练作业,记录训练作业id。
- 请求消息体:
URI格式:POST https://{ma_endpoint}/v2/{project_id}/training-jobs
请求消息头:
- X-Auth-Token →MIIZmgYJKoZIhvcNAQcCoIIZizCCGYcCAQExDTALBglghkgBZQMEAgEwgXXXXXX...
- Content-Type →application/json
其中,加粗的斜体字段需要根据实际值填写。
请求Body:
{ "kind": "job", "metadata": { "name": "test-pytorch-cpu01", "description": "test pytorch work cpu in mode gloo" }, "algorithm": { "id": "01c399ae-8593-4ef5-9e4d-085950aacde1", "parameters": [{ "name": "dist", "value": "False" }, { "name": "world_size", "value": "1" } ], "inputs": [{ "name": "data_url", "remote": { "obs": { "obs_url": "/cnnorth4-job-test-v2/pytorch/fast_example/data/" } } }], "outputs": [{ "name": "train_url", "remote": { "obs": { "obs_url": "/cnnorth4-job-test-v2/pytorch/fast_example/outputs/" } } }] }, "spec": { "resource": { "flavor_id": "modelarts.vm.cpu.8u", "node_count": 1 }, "log_export_path": { "obs_url": "/cnnorth4-job-test-v2/pytorch/fast_example/log/" } } }
其中,加粗的斜体字段需要根据实际值填写:
- “kind”填写训练作业的类型,默认为job。
- “metadata”下的“name”和“description”填写训练作业的名称和描述。
- “algorithm”下的“id”填写4获取的算法ID。
- “algorithm”的“inputs”和“outputs”填写训练作业输入输出管道的具体信息。实例中“inputs”中“remote”下的“obs_url”表示从OBS桶中选择训练数据的OBS路径。实例中“outputs”下种“remote”下的“obs_url”表示上传训练输出至指定OBS路径。
- “spec”字段下的“flavor_id”表示训练作业所依赖的规格,使用2记录的flavor_id。“node_count”表示训练是否需要多机训练(分布式训练),此处为单机情况使用默认值“1”。“log_export_path”用于指定用户需要上传日志的obs目录。
- 返回状态码“201 Created”,表示训练作业创建成功,响应Body如下所示:
{ "kind": "job", "metadata": { "id": "66ff6991-fd66-40b6-8101-0829a46d3731", "name": "test-pytorch-cpu01", "description": "test pytorch work cpu in mode gloo", "create_time": 1641892642625, "workspace_id": "0", "ai_project": "default-ai-project", "user_name": "", "annotations": { "job_template": "Template DL", "key_task": "worker" } }, "status": { "phase": "Creating", "secondary_phase": "Creating", "duration": 0, "start_time": 0, "node_count_metrics": null, "tasks": [ "worker-0" ] }, "algorithm": { "id": "01c399ae-8593-4ef5-9e4d-085950aacde1", "name": "test-pytorch-cpu", "code_dir": "/cnnorth4-job-test-v2/pytorch/fast_example/code/cpu/", "boot_file": "/cnnorth4-job-test-v2/pytorch/fast_example/code/cpu/test-pytorch.py", "parameters": [ { "name": "dist", "description": "", "i18n_description": null, "value": "False", "constraint": { "type": "Boolean", "editable": true, "required": false, "sensitive": false, "valid_type": "None", "valid_range": [] } }, { "name": "world_size", "description": "", "i18n_description": null, "value": "1", "constraint": { "type": "Integer", "editable": true, "required": false, "sensitive": false, "valid_type": "None", "valid_range": [] } } ], "parameters_customization": true, "inputs": [ { "name": "data_url", "description": "数据来源1", "local_dir": "/home/ma-user/modelarts/inputs/data_url_0", "remote": { "obs": { "obs_url": "/cnnorth4-job-test-v2/pytorch/fast_example/data/" } } } ], "outputs": [ { "name": "train_url", "description": "输出数据1", "local_dir": "/home/ma-user/modelarts/outputs/train_url_0", "remote": { "obs": { "obs_url": "/cnnorth4-job-test-v2/pytorch/fast_example/outputs/" } }, "mode": "upload_periodically", "period": 30 } ], "engine": { "engine_id": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64", "engine_name": "PyTorch", "engine_version": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64", "usage": "training", "support_groups": "public", "tags": [ { "key": "auto_search", "value": "True" } ], "v1_compatible": false, "run_user": "1102" } }, "spec": { "resource": { "flavor_id": "modelarts.vm.cpu.8u", "flavor_name": "Computing CPU(8U) instance", "node_count": 1, "flavor_detail": { "flavor_type": "CPU", "billing": { "code": "modelarts.vm.cpu.8u", "unit_num": 1 }, "flavor_info": { "cpu": { "arch": "x86", "core_num": 8 }, "memory": { "size": 32, "unit": "GB" }, "disk": { "size": 50, "unit": "GB" } } } }, "log_export_path": { "obs_url": "/cnnorth4-job-test-v2/pytorch/fast_example/log/" }, "is_hosted_log": true } }
- 记录“metadata”下的“id”(训练作业的任务ID)字段的值便于后续步骤使用。
- “Status”下的“phase”和“secondary_phase”为表示训练作业的状态和下一步状态。示例中“Creating”表示训练作业正在创建中。
- 请求消息体:
- 调用查询训练作业详情接口使用刚创建的训练作业返回的uuid查询训练作业状态。
- 请求消息体:
URI格式:GET https://{ma_endpoint}/v2/{project_id}/training-jobs/{training_job_id}
请求消息头:X-Auth-Token →MIIZmgYJKoZIhvcNAQcCoIIZizCCGYcCAQExDTALBglghkgBZQMEAgEwgXXXXXX...
其中,加粗的斜体字段需要根据实际值填写:
“training_job_id”为5记录的训练作业的任务ID。
- 返回状态码“200 OK”,响应Body如下所示:
{ "kind": "job", "metadata": { "id": "66ff6991-fd66-40b6-8101-0829a46d3731", "name": "test-pytorch-cpu01", "description": "test pytorch work cpu in mode gloo", "create_time": 1641892642625, "workspace_id": "0", "ai_project": "default-ai-project", "user_name": "hwstaff_z00424192", "annotations": { "job_template": "Template DL", "key_task": "worker" } }, "status": { "phase": "Running", "secondary_phase": "Running", "duration": 268000, "start_time": 1641892655000, "node_count_metrics": [ [ 1641892645000, 0 ], [ 1641892654000, 0 ], [ 1641892655000, 1 ], [ 1641892922000, 1 ], [ 1641892923000, 1 ] ], "tasks": [ "worker-0" ] }, "algorithm": { "id": "01c399ae-8593-4ef5-9e4d-085950aacde1", "name": "test-pytorch-cpu", "code_dir": "/cnnorth4-job-test-v2/pytorch/fast_example/code/cpu/", "boot_file": "/cnnorth4-job-test-v2/pytorch/fast_example/code/cpu/test-pytorch.py", "parameters": [ { "name": "dist", "description": "", "i18n_description": null, "value": "False", "constraint": { "type": "Boolean", "editable": true, "required": false, "sensitive": false, "valid_type": "None", "valid_range": [] } }, { "name": "world_size", "description": "", "i18n_description": null, "value": "1", "constraint": { "type": "Integer", "editable": true, "required": false, "sensitive": false, "valid_type": "None", "valid_range": [] } } ], "parameters_customization": true, "inputs": [ { "name": "data_url", "description": "数据来源1", "local_dir": "/home/ma-user/modelarts/inputs/data_url_0", "remote": { "obs": { "obs_url": "/cnnorth4-job-test-v2/pytorch/fast_example/data/" } } } ], "outputs": [ { "name": "train_url", "description": "输出数据1", "local_dir": "/home/ma-user/modelarts/outputs/train_url_0", "remote": { "obs": { "obs_url": "/cnnorth4-job-test-v2/pytorch/fast_example/outputs/" } }, "mode": "upload_periodically", "period": 30 } ], "engine": { "engine_id": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64", "engine_name": "PyTorch", "engine_version": "pytorch_1.8.0-cuda_10.2-py_3.7-ubuntu_18.04-x86_64", "usage": "training", "support_groups": "public", "tags": [ { "key": "auto_search", "value": "True" } ], "v1_compatible": false, "run_user": "1102" } }, "spec": { "resource": { "flavor_id": "modelarts.vm.cpu.8u", "flavor_name": "Computing CPU(8U) instance", "node_count": 1, "flavor_detail": { "flavor_type": "CPU", "billing": { "code": "modelarts.vm.cpu.8u", "unit_num": 1 }, "flavor_info": { "cpu": { "arch": "x86", "core_num": 8 }, "memory": { "size": 32, "unit": "GB" }, "disk": { "size": 50, "unit": "GB" } } } }, "log_export_path": { "obs_url": "/cnnorth4-job-test-v2/pytorch/fast_example/log/" }, "is_hosted_log": true } }
根据响应可以了解训练作业的版本详情,其中“status”为“Running”表示训练作业正在运行。
- 请求消息体:
- 调用查询训练作业指定任务的日志(OBS链接)接口获取训练作业日志的对应的obs路径。
- 请求消息体:
URI格式:GET https://{ma_endpoint}/v2/{project_id}/training-jobs/{training_job_id}/tasks/{task_id}/logs/url
请求消息头:
X-Auth-Token→MIIZmgYJKoZIhvcNAQcCoIIZizCCGYcCAQExDTALBglghkgBZQMEAgEwgXXXXXX...
Content-Type→text/plain
其中,加粗的斜体字段需要根据实际值填写:
- “task_id”为训练作业的任务名称,一般使用work-0。
- Content-Type可以设置成不同方式。text/plain,返回OBS临时预览链接。application/octet-stream,返回OBS临时下载链接。
- 返回状态码“200 OK”,响应Body如下所示:
{ "obs_url": "https://modelarts-training-log-cn-north-4.obs.cn-north-4.myhuaweicloud.com:443/66ff6991-fd66-40b6-8101-0829a46d3731/worker-0/modelarts-job-66ff6991-fd66-40b6-8101-0829a46d3731-worker-0.log?AWSAccessKeyId=GFGTBKOZENDD83QEMZMV&Expires=1641896599&Signature=BedFZHEU1oCmqlI912UL9mXlhkg%3D" }
返回字段表示日志的obs路径。复制至浏览器即可看到对应效果。
- 请求消息体:
- 调用查询训练作业指定任务的运行指标接口查看训练作业的运行指标详情。
- 请求消息体:
URI格式:GET https://{ma_endpoint}/v2/{project_id}/training-jobs/{training_job_id}/metrics/{task_id}
请求消息头:X-Auth-Token →MIIZmgYJKoZIhvcNAQcCoIIZizCCGYcCAQExDTALBglghkgBZQMEAgEwgXXXXXX...
其中,加粗的斜体字段需要根据实际值填写。
- 返回状态码“200 OK”,响应Body如下所示:
{ "metrics": [ { "metric": "cpuUsage", "value": [ -1, -1, 28.622, 35.053, 39.988, 40.069, 40.082, 40.094 ] }, { "metric": "memUsage", "value": [ -1, -1, 0.544, 0.641, 0.736, 0.737, 0.738, 0.739 ] }, { "metric": "npuUtil", "value": [ -1, -1, -1, -1, -1, -1, -1, -1 ] }, { "metric": "npuMemUsage", "value": [ -1, -1, -1, -1, -1, -1, -1, -1 ] }, { "metric": "gpuUtil", "value": [ -1, -1, -1, -1, -1, -1, -1, -1 ] }, { "metric": "gpuMemUsage", "value": [ -1, -1, -1, -1, -1, -1, -1, -1 ] } ] }
可以看到CPU等相关的使用率指标。
- 请求消息体:
- 当训练作业使用完成或不再需要时,调用删除训练作业接口删除训练作业。