AWS Sagemaker custom PyTorch model inference on raw image input - image

Question

I am new to AWS Sagemaker. I have custom CV PyTorch model locally and deployed it to Sagemaker endpoint. I used custom inference.py code to define model_fn, input_fn, output_fn and predict_fn methods. So, I'm able to generate predictions on json input, which contains url to the image, the code is quite straigtforward:
def input_fn(request_body, content_type='application/json'):
logging.info('Deserializing the input data...')
image_transform = transforms.Compose([
transforms.Resize(size=(224, 224)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
if content_type:
if content_type == 'application/json':
input_data = json.loads(request_body)
url = input_data['url']
logging.info(f'Image url: {url}')
image_data = Image.open(requests.get(url, stream=True).raw)
return image_transform(image_data)
raise Exception(f'Requested unsupported ContentType in content_type {content_type}')
Then I am able to invoke endpoint with code:
client = boto3.client('runtime.sagemaker')
inp = {"url":url}
inp = json.loads(json.dumps(inp))
response = client.invoke_endpoint(EndpointName='ENDPOINT_NAME',
Body=json.dumps(inp),
ContentType='application/json')
The problem is, I see, that locally url request return slightly different image array comparing to the one on Sagemaker. Which is why on the same URL I obtain slightly different predictions. To check that at least model weights are the same I want to generate predictions on image itself, downloaded locally and to Sagemaker. But I fail trying to put image as input to endpoint. E.g.:
def input_fn(request_body, content_type='application/json'):
logging.info('Deserializing the input data...')
image_transform = transforms.Compose([
transforms.Resize(size=(224, 224)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
if content_type == 'application/x-image':
image_data = request_body
return image_transform(image_data)
raise Exception(f'Requested unsupported ContentType in content_type {content_type}')
Invoking endpoint I experience the error:
ParamValidationError: Parameter validation failed:
Invalid type for parameter Body, value: {'img': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=630x326 at 0x7F78A61461D0>}, type: <class 'dict'>, valid types: <class 'bytes'>, <class 'bytearray'>, file-like object
Does anybody know how to generate Sagemaker predictions by Pytorch model on images?

Answer 1

As always, after asking I found a solution. Actually, as the error suggested, I had to convert input to bytes or bytearray. For those who may need the solution:
from io import BytesIO
img = Image.open(open(PATH, 'rb'))
img_byte_arr = BytesIO()
img.save(img_byte_arr, format=img.format)
img_byte_arr = img_byte_arr.getvalue()
client = boto3.client('runtime.sagemaker')
response = client.invoke_endpoint(EndpointName='ENDPOINT_NAME
Body=img_byte_arr,
ContentType='application/x-image')
response_body = response['Body']
print(response_body.read())