how to handle application/x-image ?

[Adblu]

Im getting:

"2020-06-05T11:05:57.016:[sagemaker logs]: data/Third_Set_Case_12774_Bilde_18076231_17.jpg: {"error": "Unsupported Media Type: application/x-image"}" 

Its documented in official documentation:

tensorflow_serving_transformer.transform(input_path, content_type='application/x-image')

How to get rid of this error ?

[nadiaya]

It is not a supported by default content type.
Here is an example of how to write a pre- and post-processing script in order for it to work: https://github.com/awslabs/amazon-sagemaker-examples/blob/master/sagemaker_batch_transform/tensorflow_open-images_jpg/tensorflow-serving-jpg-python-sdk.ipynb

Could you share a link to the official documentation you are referring to, so we could clarify it?

[Adblu]

It is here: link 1, link 2

[laurenyu]

sorry for the delayed response here. in the links you provided, I'm seeing "application/x-image" only as an example.

documentation for supported content types: https://github.com/aws/sagemaker-tensorflow-serving-container/blob/master/README.md#prepost-processing

[kafka399]

@laurenyu I would suggest adding an explicit comment, that 'application/x-image' is not supported by default.

[laurenyu]

@kafka399 thanks for the suggestion! I'm no longer a maintainer of this repository, so I probably won't get to this particularly soon - feel free to submit a PR yourself if you'd like.

cc @ChuyangDeng

[asmaier]

What about the example for inference.py given here: https://docs.aws.amazon.com/sagemaker/latest/dg/neo-deployment-hosting-services-prerequisites.html

It says:

The SageMaker client sends the image file as an application/x-image content type to the input_handler function, where it is converted to JSON.

Is this obsolete?

[asmaier]

Here is another example showing that one can send application/x-image to a sagemaker endpoint:

https://sagemaker-examples.readthedocs.io/en/latest/introduction_to_amazon_algorithms/imageclassification_caltech/Image-classification-fulltraining.html#Evaluation

People might really think it should work from all these examples, but it doesn't.

[rbavery]

Can we get clarity on this? if this is not supported why are there examples on aws docs using application x-image? it seems like this should be supported by default since it's easiest to send a put request with curl http://localhost:8080/invocations/ -T img.jpg

[ShoSoejima]

In a transform job, you can use application/x-image as the content type.
Once you've saved your model and created your model.tar.gz, which doesn't have to include inference.py, you can use your custom input/output handlers. Please look at the following code.

(I checked the code worked with boto3==1.26.31 and sagemaker==2.124.0 .)

import os

import boto3
import sagemaker
from sagemaker.tensorflow import TensorFlowModel

sess = boto3.session.Session()
sagemaker_session = sagemaker.Session(sess)

role = "ARN_OF_YOUR_SAGEMAKER_ROLE"
bucket = "YOUR_BUCKET_NAME"
prefix = "YOUR_OBJECT_PREFIX"
s3_path = f"s3://{bucket}/{prefix}"
input_path = os.path.join(s3_path, "input")
output_path = os.path.join(s3_path, "output")

model_data = sagemaker_session.upload_data(
    "path/to/your/model.tar.gz",
    bucket,
    os.path.join(prefix, "models"),
)
tensorflow_serving_model = TensorFlowModel(
    model_data=model_data,
    role=role,
    framework_version="2.8.0",
    sagemaker_session=sagemaker_session,
    entry_point="path/to/your/inference.py",  # specify your inference.py in your local file system
    name="YOUR_MODEL_NAME",
)
transformer = tensorflow_serving_model.transform(
    instance_count=1,
    instance_type="ml.m5.xlarge",
    max_concurrent_transforms=16,  # optional
    max_payload=4,  # optional
    output_path=output_path,
)
transformer.transform(
    data=input_path,
    data_type="S3Prefix",
    job_name="YOUR_JOB_NAME",
    content_type="application/x-image",
    wait=False,
)

You can include your inference.py in your model.tar.gz. In that case, you need to place the module under a directory named code, and you don't need to specify entry_point or source_dir in the constructor of TensorFlowModel.

Please remember that your custom handlers can handle the content type as follows.

def input_handler(data, context):
    if context.request_content_type == "application/x-image":
        ...

def output_hander(data, context):
    ...

Maybe I'm too late, but I hope this will help you.