AWS sagemaker endpoint received client (400) error - amazon-web-services

Question

I've deployed a tensorflow multi-label classification model using a sagemaker endpoint as follows:
predictor = sagemaker_model.deploy(initial_instance_count=1, instance_type="ml.m5.2xlarge", endpoint_name='testing-2')
It gets deployed and works fine when I invoke it from the Sagemaker Jupyter instance:
sample = ['this movie was extremely good']
output=predictor.predict(sample)
output:
{'predictions': [[0.00370046496,
4.32942124e-06,
0.00080883503,
9.25126587e-05,
0.00023958087,
0.000130862]]}
However, I am unable to send a request to the deployed endpoint from other notebooks or sagemaker studio. I'm unsure of the request format.
I've tried several variations in the input format and still failed. The error message is as below:
sagemaker error
Request:
{
"body": {
"text": "Testing model's prediction on this text"
},
"contentType": "application/json",
"endpointName": "testing-2",
"customURL": "",
"customHeaders": [
{
"Key": "sm_endpoint_name",
"Value": "testing-2"
}
]
}
Error:
Error invoking endpoint: Received client error (400) from primary with message "{ "error": "Failed to process element:
0 key: text of 'instances' list. Error: INVALID_ARGUMENT: JSON object: does not have named input: text" }".
See https://us-west-2.console.aws.amazon.com/cloudwatch/home?region=us-west-2#logEventViewer:group=/aws/sagemaker/Endpoints/testing-2
in account 793433463428 for more information.
Is there any way to find out exactly how the model expects the request format to be?
Earlier I had the same model on my local system and the way I tested it was using this curl request:
curl -s -H 'Content-Type: application/json' -d '{"text": "what ugly posts"}' http://localhost:7070/sentiment
And it worked fine without any issues.
I've tried different formats and replaced the "text" key inside body with other words like "input", "body", nothing etc.

Answer 1

Based on your description above, I assume you are deploying the TensorFlow model using the SageMaker TensorFlow container.
If you want to view what your model expects as input you can use the saved_model CLI:
1
├── keras_metadata.pb
├── saved_model.pb
└── variables
├── variables.data-00000-of-00001
└── variables.index
!saved_model_cli show --all --dir {"1"}
After you have confirmed the input name above you can invoke the endpoint as follows:
import json
import boto3
client = boto3.client('runtime.sagemaker')
data = {"instances": ['this movie was extremely good']}
response = client.invoke_endpoint(EndpointName=<EndpointName>,
Body=json.dumps(data))
response_body = response['Body']
print(response_body.read())
The same payload can then also be used in Studio when invoking the endpoint.