# Train your model normally in SageMaker, then compile it
# for the target edge hardware (e.g., NVIDIA Jetson)

import boto3

sagemaker_client = boto3.client('sagemaker')

compilation_job = sagemaker_client.create_compilation_job(
    CompilationJobName='my-model-jetson-agx',
    RoleArn='arn:aws:iam::123456789012:role/SageMakerRole',
    InputConfig={
        'S3Uri': 's3://my-bucket/models/defect-detector.tar.gz',
        'DataInputConfig': '{"input": [1, 3, 224, 224]}',  # image shape
        'Framework': 'PYTORCH'
    },
    OutputConfig={
        'S3OutputLocation': 's3://my-bucket/compiled/',
        'TargetPlatform': {
            'Os': 'LINUX',
            'Arch': 'ARM64',
            'Accelerator': 'NVIDIA'  # targets the Jetson GPU
        }
    },
    StoppingCondition={'MaxRuntimeInSeconds': 900}
)

print(f"Compilation job started: {compilation_job['CompilationJobArn']}")

# On the Greengrass core device, use the SageMaker Edge Agent
# to load and run the compiled model

import edge_agent_pb2 as pb2
import edge_agent_pb2_grpc as pb2_grpc
import grpc
import numpy as np

# Connect to the local SageMaker Edge Agent (runs as a daemon)
channel = grpc.insecure_channel('unix:///tmp/sagemaker_edge_agent_example.sock')
stub = pb2_grpc.AgentStub(channel)

# Load model (one-time setup)
stub.LoadModel(pb2.LoadModelRequest(
    url='/greengrass/v2/models/defect-detector',
    name='defect-detector'
))

# Run inference on an image captured from local camera
def run_inference(image_array):
    tensor = pb2.Tensor(
        tensor_metadata=pb2.TensorMetadata(
            name='input',
            data_type=5,  # FLOAT32
            shape=[1, 3, 224, 224]
        ),
        byte_data=image_array.tobytes()
    )
    response = stub.Predict(pb2.PredictRequest(
        name='defect-detector',
        tensors=[tensor]
    ))
    return np.frombuffer(response.tensors[0].byte_data, dtype=np.float32)