Debugging a Training Job

Before creating a real-time training job, create a local training job for debugging.

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.

Step 1: Create a local training job. If train_instance_type is set to local, a local training job is created, which can be used to debug code and parameters.

      
           from modelarts.session import Session
from modelarts.estimator import Estimator
from modelarts.environment import Environment
from modelarts.environment.conda_env import CondaDependencies

session = Session()
env = Environment("tensorflow_mlp_mnist")
cd = CondaDependencies.create(pip_packages=["tensorflow==1.13.1", "requests"], conda_packages=["python=3.6.2"])
env.conda = cd
src_local_path = "/home/ma-user/work/tensorflow_mlp_mnist_local_mode/train/"
train_file = "tensorflow_mlp_mnist.py"
estimator = Estimator(modelarts_session=session,
                      code_dir=src_local_path,            # Path of the local training script
                      boot_file=train_file,               # Path of the local training boot script
                      train_instance_type='local',        # Local training
                      train_instance_count=1,             # Number of training nodes
                      environment=env)                    # Environment for running the training script
job_instance = estimator.fit(wait=False, job_name='my_training_job')

Step 2: After the local training job is complete, create a real-time training job. If train_instance_type is set to a training environment flavor, a real-time training job is created.

      
           from modelarts.session import Session
from modelarts.estimator import Estimator
from modelarts.environment import Environment
from modelarts.environment.conda_env import CondaDependencies

session = Session()
env = Environment("tensorflow_mlp_mnist")
cd = CondaDependencies.create(pip_packages=["tensorflow==1.13.1", "requests"], conda_packages=["python=3.6.2"])
env.conda = cd
src_local_path = "/home/ma-user/work/tensorflow_mlp_mnist_local_mode/train/"
train_file = "tensorflow_mlp_mnist.py"
estimator = Estimator(modelarts_session=session,
                      code_dir=src_local_path,                          # Path of the training script
                       boot_file=train_file,                             # Path of the training boot script
                      train_instance_type='modelarts.vm.cpu.2u',        # Real-time training
                      train_instance_count=1,                           # Number of training nodes
                      environment=env)                                  # Environment for running the training script
job_instance = estimator.fit(wait=False, job_name='my_training_job')

Parameters

**Table 1** Environment parameters
Parameter	Mandatory	Type	Description
name	Yes	String	Environment name
conda	No	CondaDependencies	Conda environment. For details, see Table 2.

**Table 2** CondaDependencies parameters
Parameter	Mandatory	Type	Description
channels	No	List	Source for downloading the Python package
pip_packages	No	List	Python package required by the Conda virtual environment, such as TensorFlow and Pillow
conda_packages	No	List	Conda package required by the Conda virtual environment, for example, a specified Python version

**Table 3** Estimator request parameters
Parameter	Mandatory	Type	Description
modelarts_session	Yes	Object	Session object. For details about the initialization method, see Session Authentication.
train_instance_count	Yes	Int	Number of compute nodes in a training job.
code_dir	No	String	Code directory of a training job, for example, /bucket/src/. Leave this parameter blank if model_name is set.
boot_file	No	String	Boot file of a training job, which needs to be stored in the code directory, for example, /bucket/src/boot.py. Leave this parameter blank if model_name is set.
model_name	No	String	Name of the built-in algorithm used by a training job. If you have configured model_name, you do not need to configure app_url, boot_file_url, framework_type, and framework_version. You can obtain this value by calling the API described in Querying a Built-in Algorithm.
output_path	Yes	String	Output path of a training job.
hyperparameters	No	JSON array	Running parameters of label-value pairs of the string type for a training job. This parameter is a container environment variable if a job uses a custom image.
log_url	No	String	OBS URL of training job logs. By default, this parameter is left blank. An example value is /usr/log/.
train_instance_type	Yes	String	Resource flavor selected for a training job. If you choose to train on the training platform, obtain the value by calling the API described in Querying the List of Resource Flavors.
framework_type	No	String	Engine selected for a training job. Obtain the value by calling the API described in Querying the List of Engine Types. Leave this parameter blank if model_name is set.
framework_version	No	String	Engine version selected for a training job. Obtain the value by calling the API described in Querying the List of Engine Types. Leave this parameter blank if model_name is set.
job_description	No	String	Description of a training job.
user_image_url	No	String	SWR URL of the custom image used by a training job. An example value is 100.125.5.235:20202/jobmng/custom-cpu-base:1.0.
user_command	No	String	Boot command used to start the container of the custom image used by a training job. The format is bash /home/work/run_train.sh python /home/work/user-job-dir/app/train.py {python_file_parameter}.
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 console, choose Dedicated Resource Pools in the navigation pane, and view the resource pool ID in the dedicated resource pool list.

**Table 4** **fit** request parameters
Parameter	Mandatory	Type	Description
inputs	Yes	String	Data storage location of a training job. inputs cannot be used with dataset_id and dataset_version_id, or with data_source at the same time. However, one of these parameters must be set. Only this parameter is supported in local training.
dataset_id	No	String	Dataset ID of a training job. 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. This parameter must be used with dataset_id, but cannot be used with inputs.
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. Enter 1 to 64 characters. Only the following characters are allowed: a-z, A-Z, 0-9, hyphens (-), and underscores (_). If this parameter is left blank, a job name is generated randomly.