Updated on 2024-06-12 GMT+08:00

Creating a Training Job

Sample Code

In ModelArts notebook, you do not need to enter authentication parameters for session authentication. For details about session authentication of other development environments, see Session Authentication.

ModelArts SDK cannot be used to create training jobs using algorithms subscribed to in AI Gallery.

  • Example 1: Create a training job using a common AI engine.

    If both framework_type and framework_version are specified in estimator, a training job will be created using a common AI engine.

    from modelarts.session import Session
    from modelarts.train_params import TrainingFiles
    from modelarts.train_params import OutputData
    from modelarts.train_params import InputData
    from modelarts.estimatorV2 import Estimator
    session = Session()
    # Parameters received in the training script (set based on the site requirements):
    
    parameters = [{"name": "mod", "value":"gpu"}, 
                  {"name": "epoc_num", "value":2}]
    estimator = Estimator(session=session,
                          training_files=TrainingFiles(code_dir= "obs://bucket_name/code_dir/", boot_file="boot_file.py"),
                          outputs=[OutputData(obs_path="obs://bucket_name/output/", name="output_dir")],
                          parameters=parameters,
                          framework_type='PyTorch',                      # Common AI engine
                          framework_version='PyTorch-1.4.0-python3.6',   # Version of the AI engine
                          train_instance_type="modelarts.p3.large.public",
                          train_instance_count=1,
                          log_url="obs://bucket_name/log/",
                          env_variables={"USER_ENV_VAR": "customize environment variable"},
                          working_dir="/home/ma-user/modelarts/user-job-dir",
                          local_code_dir="/home/ma-user/modelarts/user-job-dir",
                          job_description='This is an image net train job')
    job_instance = estimator.fit(inputs=[InputData(obs_path="obs://bucket_name/input/", name="data_url")],
                                 job_name="job_name_1")
  • Example 2: Create a training job using a custom image.

    If both user_image_url and user_command are specified in estimator, a training job will be created using a custom image and started using a custom boot command.

    from modelarts.session import Session
    from modelarts.train_params import TrainingFiles
    from modelarts.train_params import OutputData
    from modelarts.train_params import InputData
    from modelarts.estimatorV2 import Estimator
    session = Session()
    # Parameters received in the training script (set based on the site requirements):
    
    parameters = [{"name": "mod", "value":"gpu"}, 
                  {"name": "epoc_num", "value":2}]
    estimator = Estimator(session=session,
                          training_files=TrainingFiles(code_dir= "obs://bucket_name/code_dir/", boot_file="boot_file.py"),
                          outputs=[OutputData(obs_path="obs://bucket_name/output/", name="output_dir")],
                          parameters=parameters,
                          user_image_url="sdk-test/pytorch1_4:1.0.1",     # URL of the custom image
                          user_command="/home/ma-user/anaconda3/envs/PyTorch-1.4/bin/python /home/ma-user/modelarts/user-job-dir/train/test-pytorch.py",  # Custom boot command
                          train_instance_type="modelarts.p3.large.public",
                          train_instance_count=1,
                          log_url="obs://bucket_name/log/",
                          env_variables={"USER_ENV_VAR": "customize environment variable"},
                          working_dir="/home/ma-user/modelarts/user-job-dir",
                          local_code_dir="/home/ma-user/modelarts/user-job-dir",
                          job_description='This is an image net train job')
    job_instance = estimator.fit(inputs=[InputData(obs_path="obs://bucket_name/input/", name="data_url")],
                                 job_name="job_name_2")
  • Example 3: Creating a training job in a dedicated resource pool
    from modelarts.session import Session
    from modelarts.train_params import TrainingFiles
    from modelarts.train_params import OutputData
    from modelarts.train_params import InputData
    from modelarts.estimatorV2 import Estimator
    session = Session()
    # Parameters received in the training script (set based on the site requirements):
    
    parameters = [{"name": "mod", "value":"gpu"}, 
                  {"name": "epoc_num", "value":2}]
    estimator = Estimator(session=session,
                          training_files=TrainingFiles(code_dir= "obs://bucket_name/code_dir/", boot_file="boot_file.py"),
                          outputs=[OutputData(obs_path="obs://bucket_name/output/", name="output_dir")],
                          parameters=parameters,
                          framework_type='PyTorch',                       
                          framework_version='PyTorch-1.4.0-python3.6',    
                          pool_id="your pool id",                                 # Dedicated resource pool ID
                          train_instance_type="modelarts.pool.visual.xlarge",     # VM flavor of the dedicated pool
                          train_instance_count=1,
                          log_url="obs://bucket_name/log/",
                          env_variables={"USER_ENV_VAR": "customize environment variable"},
                          working_dir="/home/ma-user/modelarts/user-job-dir",
                          local_code_dir="/home/ma-user/modelarts/user-job-dir",
                          job_description='This is an image net train job')
    job_instance = estimator.fit(inputs=[InputData(obs_path="obs://bucket_name/input/", name="data_url")],
                                 job_name="job_name_3")
  • Example 4: Create a training job using a dataset.
    from modelarts.session import Session
    from modelarts.train_params import TrainingFiles
    from modelarts.train_params import OutputData
    from modelarts.train_params import InputData
    from modelarts.estimatorV2 import Estimator
    session = Session()
    # Parameters received in the training script (set based on the site requirements):
    parameters = [{"name": "model_name", "value":"s"}, 
                  {"name": "batch-size", "value": 32},
                  {"name": "epochs", "value":100},
                  {"name": "img-size", "value":"640,640"} ]
    estimator = Estimator(session=session,
                          training_files=TrainingFiles(code_dir= "obs://bucket_name/code_dir/", boot_file="boot_file.py"),
                          outputs=[OutputData(obs_path="obs://bucket_name/output/", name="output_dir")],
                          parameters=parameters,
                          framework_type='PyTorch',                       # Common AI engine
                          framework_version='PyTorch-1.4.0-python3.6',    # Version of the AI engine
                          train_instance_type="modelarts.p3.large.public",
                          train_instance_count=1,
                          log_url="obs://bucket_name/log/",
                          working_dir="/home/ma-user/modelarts/user-job-dir",
                          local_code_dir="/home/ma-user/modelarts/user-job-dir",
                          job_description='This is an image net train job')
    job_instance = estimator.fit(dataset_id="your dataset id",
                                 dataset_version_id="your dataset version id",
                                 job_name="job_name_5")

Parameters

Table 1 Estimator request parameters

Parameter

Mandatory

Type

Description

session

Yes

Object

Session object. For details about the initialization method, see Session Authentication.

training_files

No

TrainingFiles Object

Path to the training script in OBS. For details, see Table 2.

outputs

No

Array of OutputData objects

Training output path. For details, see Table 3.

parameters

No

JSON Array

Running parameters of a training job. The format is as follows:

[{"name":"your name", "value": "your value"}]. The value can be a string or an integer.

train_instance_type

Yes

String

Resource flavor selected for a training job. For details, see Obtaining Resource Flavors.

train_instance_count

Yes

Int

Number of compute nodes in a training job

framework_type

No

String

Engine type selected for a training job. For details, see Obtaining Engine Types.

framework_version

No

String

Engine version selected for a training job. For details, see Obtaining Engine Types.

user_image_url

No

String

SWR URL of the custom image used by a training job

user_command

No

String

Command for starting a training job created using a custom image

log_url

No

String

OBS path for storing training job logs, for example, obs://xx/yy/zz/

local_code_dir

No

String

Local directory to the training container to which the algorithm code directory is downloaded. Note:

  • The directory must be under /home.
  • In v1 compatibility mode, this parameter does not take effect.
  • When code_dir is prefixed with file://, this parameter does not take effect.

working_dir

No

String

Work directory where an algorithm is executed. Note that this parameter does not take effect in v1 compatibility mode.

job_description

No

String

Description of a training job

volumes

No

JSON Array

Information of the disks attached for a training job in the following example format:

[{

"nfs": {

"local_path": "/xx/yy/zz",

"read_only": False,

"nfs_server_path": "xxx.xxx.xxx.xxx:/"

}

}]

env_variables

No

Dict

Environment variables of a training job

pool_id

No

String

ID of the resource pool for a training job. To obtain the ID, do as follows: Log in to the ModelArts management console, choose Dedicated Resource Pools in the navigation pane on the left, and view the resource pool ID in the dedicated resource pool list.

Table 2 Parameters for initializing TrainingFiles

Parameter

Mandatory

Type

Description

code_dir

Yes

String

Code directory of a training job, which is an OBS path and must start with obs:/, for example, obs://xx/yy/

boot_file

Yes

String

Boot file of a training job, which must be stored in the code directory. You can enter a relative path, for example, boot_file.py, or an absolute path, for example, obs://xx/yy/boot_file.py.

Table 3 Parameters for initializing OutputData

Parameter

Mandatory

Type

Description

obs_path

Yes

String

OBS path to which data is exported

name

Yes

String

Keyword parameter name of the output data, for example, output_dir

Table 4 fit request parameters

Parameter

Mandatory

Type

Description

inputs

No

Array of InputData Object

Input data of a training job stored in OBS Either inputs or dataset_id/dataset_version_id can be configured.

wait

No

Boolean

Whether to wait for the completion of a training job. It defaults to False.

job_name

No

String

Name of a training job

show_log

No

Boolean

Whether to output training job logs after a job is submitted. It defaults to False.

dataset_id

No

String

Dataset ID of a training job. For details, see Data Management. This parameter must be used with dataset_version_id, but cannot be used with inputs.

dataset_version_id

No

String

Dataset version ID of a training job. For details, see Data Management. This parameter must be used with dataset_id, but cannot be used with inputs.

Table 5 Parameters for initializing InputData

Parameter

Mandatory

Type

Description

obs_path

Yes

String

OBS path to the dataset required by a training job, for example, obs://xx/yy/

name

Yes

String

Keyword parameter name of the input data, for example, data_url.

Table 6 Response for creating a training job

Parameter

Type

Description

TrainingJob

Object

Training object, which contains attributes such as job_id. When you perform operations on a training job, for example, obtain information of, update, or delete a training job, you can use job_instance.job_id to obtain the ID of the training job.

Table 7 Response for the failure to call a training API

Parameter

Type

Description

error_msg

String

Error message when calling an API failed. This parameter is unavailable if an API is successfully called.

error_code

String

Error code when calling an API failed. For details, see "Error Codes" in ModelArts API Reference. This parameter is unavailable if an API is successfully called.

error_solution

String

Solution to an API calling failure. This parameter is unavailable if an API is successfully called.