aws_sagemaker_compilationjob

account_id

Type: STRING

compilation_end_time

Type: TIMESTAMP
Provider name: CompilationEndTime
Description: The time when the model compilation job on a compilation job instance ended. For a successful or stopped job, this is when the job’s model artifacts have finished uploading. For a failed job, this is when Amazon SageMaker AI detected that the job failed.

compilation_job_arn

Type: STRING
Provider name: CompilationJobArn
Description: The Amazon Resource Name (ARN) of the model compilation job.

compilation_job_name

Type: STRING
Provider name: CompilationJobName
Description: The name of the model compilation job.

compilation_job_status

Type: STRING
Provider name: CompilationJobStatus
Description: The status of the model compilation job.

compilation_start_time

Type: TIMESTAMP
Provider name: CompilationStartTime
Description: The time when the model compilation job started the CompilationJob instances. You are billed for the time between this timestamp and the timestamp in the CompilationEndTime field. In Amazon CloudWatch Logs, the start time might be later than this time. That’s because it takes time to download the compilation job, which depends on the size of the compilation job container.

creation_time

Type: TIMESTAMP
Provider name: CreationTime
Description: The time that the model compilation job was created.

derived_information

Type: STRUCT
Provider name: DerivedInformation
Description: Information that SageMaker Neo automatically derived about the model.

• derived_data_input_config
  Type: STRING
  Provider name: DerivedDataInputConfig
  Description: The data input configuration that SageMaker Neo automatically derived for the model. When SageMaker Neo derives this information, you don’t need to specify the data input configuration when you create a compilation job.
  

failure_reason

Type: STRING
Provider name: FailureReason
Description: If a model compilation job failed, the reason it failed.

inference_image

Type: STRING
Provider name: InferenceImage
Description: The inference image to use when compiling a model. Specify an image only if the target device is a cloud instance.

input_config

Type: STRUCT
Provider name: InputConfig
Description: Information about the location in Amazon S3 of the input model artifacts, the name and shape of the expected data inputs, and the framework in which the model was trained.

• data_input_config
  Type: STRING
  Provider name: DataInputConfig
  Description: Specifies the name and shape of the expected data inputs for your trained model with a JSON dictionary form. The data inputs are Framework specific.
  • TensorFlow: You must specify the name and shape (NHWC format) of the expected data inputs using a dictionary format for your trained model. The dictionary formats required for the console and CLI are different.
    • Examples for one input:
      • If using the console, {“input”:[1,1024,1024,3]}
      • If using the CLI, {"input":[1,1024,1024,3]}
    • Examples for two inputs:
      • If using the console, {“data1”: [1,28,28,1], “data2”:[1,28,28,1]}
      • If using the CLI, {"data1": [1,28,28,1], "data2":[1,28,28,1]}
  • KERAS: You must specify the name and shape (NCHW format) of expected data inputs using a dictionary format for your trained model. Note that while Keras model artifacts should be uploaded in NHWC (channel-last) format, DataInputConfig should be specified in NCHW (channel-first) format. The dictionary formats required for the console and CLI are different.
    • Examples for one input:
      • If using the console, {“input_1”:[1,3,224,224]}
      • If using the CLI, {"input_1":[1,3,224,224]}
    • Examples for two inputs:
      • If using the console, {“input_1”: [1,3,224,224], “input_2”:[1,3,224,224]}
      • If using the CLI, {"input_1": [1,3,224,224], "input_2":[1,3,224,224]}
  • MXNET/ONNX/DARKNET: You must specify the name and shape (NCHW format) of the expected data inputs in order using a dictionary format for your trained model. The dictionary formats required for the console and CLI are different.
    • Examples for one input:
      • If using the console, {“data”:[1,3,1024,1024]}
      • If using the CLI, {"data":[1,3,1024,1024]}
    • Examples for two inputs:
      • If using the console, {“var1”: [1,1,28,28], “var2”:[1,1,28,28]}
      • If using the CLI, {"var1": [1,1,28,28], "var2":[1,1,28,28]}
  • PyTorch: You can either specify the name and shape (NCHW format) of expected data inputs in order using a dictionary format for your trained model or you can specify the shape only using a list format. The dictionary formats required for the console and CLI are different. The list formats for the console and CLI are the same.
    • Examples for one input in dictionary format:
      • If using the console, {“input0”:[1,3,224,224]}
      • If using the CLI, {"input0":[1,3,224,224]}
    • Example for one input in list format: [[1,3,224,224]]
    • Examples for two inputs in dictionary format:
      • If using the console, {“input0”:[1,3,224,224], “input1”:[1,3,224,224]}
      • If using the CLI, {"input0":[1,3,224,224], "input1":[1,3,224,224]}
    • Example for two inputs in list format: [[1,3,224,224], [1,3,224,224]]
  • XGBOOST: input data name and shape are not needed.
  DataInputConfig supports the following parameters for CoreML TargetDevice (ML Model format):
  • shape: Input shape, for example {“input_1”: {“shape”: [1,224,224,3]}}. In addition to static input shapes, CoreML converter supports Flexible input shapes:
    • Range Dimension. You can use the Range Dimension feature if you know the input shape will be within some specific interval in that dimension, for example: {“input_1”: {“shape”: [“1..10”, 224, 224, 3]}}
    • Enumerated shapes. Sometimes, the models are trained to work only on a select set of inputs. You can enumerate all supported input shapes, for example: {“input_1”: {“shape”: [[1, 224, 224, 3], [1, 160, 160, 3]]}}
  • default_shape: Default input shape. You can set a default shape during conversion for both Range Dimension and Enumerated Shapes. For example {“input_1”: {“shape”: [“1..10”, 224, 224, 3], “default_shape”: [1, 224, 224, 3]}}
  • type: Input type. Allowed values: Image and Tensor. By default, the converter generates an ML Model with inputs of type Tensor (MultiArray). User can set input type to be Image. Image input type requires additional input parameters such as bias and scale.
  • bias: If the input type is an Image, you need to provide the bias vector.
  • scale: If the input type is an Image, you need to provide a scale factor.
  CoreML ClassifierConfig parameters can be specified using OutputConfig CompilerOptions. CoreML converter supports Tensorflow and PyTorch models. CoreML conversion examples:
  • Tensor type input:
    • “DataInputConfig”: {“input_1”: {“shape”: [[1,224,224,3], [1,160,160,3]], “default_shape”: [1,224,224,3]}}
  • Tensor type input without input name (PyTorch):
    • “DataInputConfig”: [{“shape”: [[1,3,224,224], [1,3,160,160]], “default_shape”: [1,3,224,224]}]
  • Image type input:
    • “DataInputConfig”: {“input_1”: {“shape”: [[1,224,224,3], [1,160,160,3]], “default_shape”: [1,224,224,3], “type”: “Image”, “bias”: [-1,-1,-1], “scale”: 0.007843137255}}
    • “CompilerOptions”: {“class_labels”: “imagenet_labels_1000.txt”}
  • Image type input without input name (PyTorch):
    • “DataInputConfig”: [{“shape”: [[1,3,224,224], [1,3,160,160]], “default_shape”: [1,3,224,224], “type”: “Image”, “bias”: [-1,-1,-1], “scale”: 0.007843137255}]
    • “CompilerOptions”: {“class_labels”: “imagenet_labels_1000.txt”}
  Depending on the model format, DataInputConfig requires the following parameters for ml_eia2 OutputConfig:TargetDevice.
  • For TensorFlow models saved in the SavedModel format, specify the input names from signature_def_key and the input model shapes for DataInputConfig. Specify the signature_def_key in OutputConfig:CompilerOptions if the model does not use TensorFlow’s default signature def key. For example:
    • “DataInputConfig”: {“inputs”: [1, 224, 224, 3]}
    • “CompilerOptions”: {“signature_def_key”: “serving_custom”}
  • For TensorFlow models saved as a frozen graph, specify the input tensor names and shapes in DataInputConfig and the output tensor names for output_names in OutputConfig:CompilerOptions . For example:
    • “DataInputConfig”: {“input_tensor:0”: [1, 224, 224, 3]}
    • “CompilerOptions”: {“output_names”: [“output_tensor:0”]}
• framework
  Type: STRING
  Provider name: Framework
  Description: Identifies the framework in which the model was trained. For example: TENSORFLOW.
  
• framework_version
  Type: STRING
  Provider name: FrameworkVersion
  Description: Specifies the framework version to use. This API field is only supported for the MXNet, PyTorch, TensorFlow and TensorFlow Lite frameworks. For information about framework versions supported for cloud targets and edge devices, see Cloud Supported Instance Types and Frameworks and Edge Supported Frameworks.
  
• s3_uri
  Type: STRING
  Provider name: S3Uri
  Description: The S3 path where the model artifacts, which result from model training, are stored. This path must point to a single gzip compressed tar archive (.tar.gz suffix).
  

last_modified_time

Type: TIMESTAMP
Provider name: LastModifiedTime
Description: The time that the status of the model compilation job was last modified.

model_artifacts

Type: STRUCT
Provider name: ModelArtifacts
Description: Information about the location in Amazon S3 that has been configured for storing the model artifacts used in the compilation job.

• s3_model_artifacts
  Type: STRING
  Provider name: S3ModelArtifacts
  Description: The path of the S3 object that contains the model artifacts. For example, s3://bucket-name/keynameprefix/model.tar.gz.
  

model_digests

Type: STRUCT
Provider name: ModelDigests
Description: Provides a BLAKE2 hash value that identifies the compiled model artifacts in Amazon S3.

• artifact_digest
  Type: STRING
  Provider name: ArtifactDigest
  Description: Provides a hash value that uniquely identifies the stored model artifacts.
  

model_package_version_arn

Type: STRING
Provider name: ModelPackageVersionArn
Description: The Amazon Resource Name (ARN) of the versioned model package that was provided to SageMaker Neo when you initiated a compilation job.

output_config

Type: STRUCT
Provider name: OutputConfig
Description: Information about the output location for the compiled model and the target device that the model runs on.

• compiler_options
  Type: STRING
  Provider name: CompilerOptions
  Description: Specifies additional parameters for compiler options in JSON format. The compiler options are TargetPlatform specific. It is required for NVIDIA accelerators and highly recommended for CPU compilations. For any other cases, it is optional to specify CompilerOptions.
  • DTYPE: Specifies the data type for the input. When compiling for ml_* (except for ml_inf) instances using PyTorch framework, provide the data type (dtype) of the model’s input. “float32” is used if “DTYPE” is not specified. Options for data type are:
    • float32: Use either “float” or “float32”.
    • int64: Use either “int64” or “long”.
    For example, {“dtype” : “float32”}.
  • CPU: Compilation for CPU supports the following compiler options.
    • mcpu: CPU micro-architecture. For example, {‘mcpu’: ‘skylake-avx512’}
    • mattr: CPU flags. For example, {‘mattr’: [’+neon’, ‘+vfpv4’]}
  • ARM: Details of ARM CPU compilations.
    • NEON: NEON is an implementation of the Advanced SIMD extension used in ARMv7 processors. For example, add {‘mattr’: [’+neon’]} to the compiler options if compiling for ARM 32-bit platform with the NEON support.
  • NVIDIA: Compilation for NVIDIA GPU supports the following compiler options.
    • gpu_code: Specifies the targeted architecture.
    • trt-ver: Specifies the TensorRT versions in x.y.z. format.
    • cuda-ver: Specifies the CUDA version in x.y format.
    For example, {‘gpu-code’: ‘sm_72’, ’trt-ver’: ‘6.0.1’, ‘cuda-ver’: ‘10.1’}
  • ANDROID: Compilation for the Android OS supports the following compiler options:
    • ANDROID_PLATFORM: Specifies the Android API levels. Available levels range from 21 to 29. For example, {‘ANDROID_PLATFORM’: 28}.
    • mattr: Add {‘mattr’: [’+neon’]} to compiler options if compiling for ARM 32-bit platform with NEON support.
  • INFERENTIA: Compilation for target ml_inf1 uses compiler options passed in as a JSON string. For example, “CompilerOptions”: “"–verbose 1 –num-neuroncores 2 -O2"". For information about supported compiler options, see Neuron Compiler CLI Reference Guide.
  • CoreML: Compilation for the CoreML OutputConfig TargetDevice supports the following compiler options:
    • class_labels: Specifies the classification labels file name inside input tar.gz file. For example, {“class_labels”: “imagenet_labels_1000.txt”}. Labels inside the txt file should be separated by newlines.
• kms_key_id
  Type: STRING
  Provider name: KmsKeyId
  Description: The Amazon Web Services Key Management Service key (Amazon Web Services KMS) that Amazon SageMaker AI uses to encrypt your output models with Amazon S3 server-side encryption after compilation job. If you don’t provide a KMS key ID, Amazon SageMaker AI uses the default KMS key for Amazon S3 for your role’s account. For more information, see KMS-Managed Encryption Keys in the Amazon Simple Storage Service Developer Guide. The KmsKeyId can be any of the following formats:
  • Key ID: 1234abcd-12ab-34cd-56ef-1234567890ab
  • Key ARN: arn:aws:kms:us-west-2:111122223333:key/1234abcd-12ab-34cd-56ef-1234567890ab
  • Alias name: alias/ExampleAlias
  • Alias name ARN: arn:aws:kms:us-west-2:111122223333:alias/ExampleAlias
• s3_output_location
  Type: STRING
  Provider name: S3OutputLocation
  Description: Identifies the S3 bucket where you want Amazon SageMaker AI to store the model artifacts. For example, s3://bucket-name/key-name-prefix.
  
• target_device
  Type: STRING
  Provider name: TargetDevice
  Description: Identifies the target device or the machine learning instance that you want to run your model on after the compilation has completed. Alternatively, you can specify OS, architecture, and accelerator using TargetPlatform fields. It can be used instead of TargetPlatform. Currently ml_trn1 is available only in US East (N. Virginia) Region, and ml_inf2 is available only in US East (Ohio) Region.
  
• target_platform
  Type: STRUCT
  Provider name: TargetPlatform
  Description: Contains information about a target platform that you want your model to run on, such as OS, architecture, and accelerators. It is an alternative of TargetDevice. The following examples show how to configure the TargetPlatform and CompilerOptions JSON strings for popular target platforms:
  • Raspberry Pi 3 Model B+ “TargetPlatform”: {“Os”: “LINUX”, “Arch”: “ARM_EABIHF”}, “CompilerOptions”: {‘mattr’: [’+neon’]}
  • Jetson TX2 “TargetPlatform”: {“Os”: “LINUX”, “Arch”: “ARM64”, “Accelerator”: “NVIDIA”}, “CompilerOptions”: {‘gpu-code’: ‘sm_62’, ’trt-ver’: ‘6.0.1’, ‘cuda-ver’: ‘10.0’}
  • EC2 m5.2xlarge instance OS “TargetPlatform”: {“Os”: “LINUX”, “Arch”: “X86_64”, “Accelerator”: “NVIDIA”}, “CompilerOptions”: {‘mcpu’: ‘skylake-avx512’}
  • RK3399 “TargetPlatform”: {“Os”: “LINUX”, “Arch”: “ARM64”, “Accelerator”: “MALI”}
  • ARMv7 phone (CPU) “TargetPlatform”: {“Os”: “ANDROID”, “Arch”: “ARM_EABI”}, “CompilerOptions”: {‘ANDROID_PLATFORM’: 25, ‘mattr’: [’+neon’]}
  • ARMv8 phone (CPU) “TargetPlatform”: {“Os”: “ANDROID”, “Arch”: “ARM64”}, “CompilerOptions”: {‘ANDROID_PLATFORM’: 29}
  • accelerator
    Type: STRING
    Provider name: Accelerator
    Description: Specifies a target platform accelerator (optional).
    • NVIDIA: Nvidia graphics processing unit. It also requires gpu-code, trt-ver, cuda-ver compiler options
    • MALI: ARM Mali graphics processor
    • INTEL_GRAPHICS: Integrated Intel graphics
  • arch
    Type: STRING
    Provider name: Arch
    Description: Specifies a target platform architecture.
    • X86_64: 64-bit version of the x86 instruction set.
    • X86: 32-bit version of the x86 instruction set.
    • ARM64: ARMv8 64-bit CPU.
    • ARM_EABIHF: ARMv7 32-bit, Hard Float.
    • ARM_EABI: ARMv7 32-bit, Soft Float. Used by Android 32-bit ARM platform.
  • os
    Type: STRING
    Provider name: Os
    Description: Specifies a target platform OS.
    • LINUX: Linux-based operating systems.
    • ANDROID: Android operating systems. Android API level can be specified using the ANDROID_PLATFORM compiler option. For example, “CompilerOptions”: {‘ANDROID_PLATFORM’: 28}

role_arn

Type: STRING
Provider name: RoleArn
Description: The Amazon Resource Name (ARN) of an IAM role that Amazon SageMaker AI assumes to perform the model compilation job.

stopping_condition

Type: STRUCT
Provider name: StoppingCondition
Description: Specifies a limit to how long a model compilation job can run. When the job reaches the time limit, Amazon SageMaker AI ends the compilation job. Use this API to cap model training costs.

• max_pending_time_in_seconds
  Type: INT32
  Provider name: MaxPendingTimeInSeconds
  Description: The maximum length of time, in seconds, that a training or compilation job can be pending before it is stopped. When working with training jobs that use capacity from training plans, not all Pending job states count against the MaxPendingTimeInSeconds limit. The following scenarios do not increment the MaxPendingTimeInSeconds counter:
  • The plan is in a Scheduled state: Jobs queued (in Pending status) before a plan’s start date (waiting for scheduled start time)
  • Between capacity reservations: Jobs temporarily back to Pending status between two capacity reservation periods
  MaxPendingTimeInSeconds only increments when jobs are actively waiting for capacity in an Active plan.
  
• max_runtime_in_seconds
  Type: INT32
  Provider name: MaxRuntimeInSeconds
  Description: The maximum length of time, in seconds, that a training or compilation job can run before it is stopped. For compilation jobs, if the job does not complete during this time, a TimeOut error is generated. We recommend starting with 900 seconds and increasing as necessary based on your model. For all other jobs, if the job does not complete during this time, SageMaker ends the job. When RetryStrategy is specified in the job request, MaxRuntimeInSeconds specifies the maximum time for all of the attempts in total, not each individual attempt. The default value is 1 day. The maximum value is 28 days. The maximum time that a TrainingJob can run in total, including any time spent publishing metrics or archiving and uploading models after it has been stopped, is 30 days.
  
• max_wait_time_in_seconds
  Type: INT32
  Provider name: MaxWaitTimeInSeconds
  Description: The maximum length of time, in seconds, that a managed Spot training job has to complete. It is the amount of time spent waiting for Spot capacity plus the amount of time the job can run. It must be equal to or greater than MaxRuntimeInSeconds. If the job does not complete during this time, SageMaker ends the job. When RetryStrategy is specified in the job request, MaxWaitTimeInSeconds specifies the maximum time for all of the attempts in total, not each individual attempt.
  

tags

Type: UNORDERED_LIST_STRING

vpc_config

Type: STRUCT
Provider name: VpcConfig
Description: A VpcConfig object that specifies the VPC that you want your compilation job to connect to. Control access to your models by configuring the VPC. For more information, see Protect Compilation Jobs by Using an Amazon Virtual Private Cloud.

• security_group_ids
  Type: UNORDERED_LIST_STRING
  Provider name: SecurityGroupIds
  Description: The VPC security group IDs. IDs have the form of sg-xxxxxxxx. Specify the security groups for the VPC that is specified in the Subnets field.
  
• subnets
  Type: UNORDERED_LIST_STRING
  Provider name: Subnets
  Description: The ID of the subnets in the VPC that you want to connect the compilation job to for accessing the model in Amazon S3.