3 things to know when working & debugging Amazon SageMaker
The case of training & deploying an MXNet-based deep learning network via Jupyter Notebook
Recently I had the chance to use Amazon SageMaker and so it happened that I needed to debug a call request from my Jupyter Notebook to the SageMaker API Endpoint. In this short post, I'm going to show you 3 things that might come handy if you find yourself in a similar situation.
1. AWS CLI
I did that using the AWS command line interface from my terminal. There are few things you should make sure you provide, so I'm going to highlight them quickly before showing the command itself.
- Provide the region name
- Provide the name of your SageMaker API endpoint
- Provide the body (which in this case is a pixelized matrix representing handwritten number 4); Note the body depends on the model you are using. I used the AWS example of training and deploying a deep learning network using MXNet, if you would like to follow the example you can do it on this link [Using Apache MXNet].(https://docs.aws.amazon.com/sagemaker/latest/dg/mxnet-example1-prepare.html)
- Tell the endpoint what type of content are you sending (via the --content-type attribute); You can learn more about the possible content types Common Data Formats - Inference
- Tell the endpoint what type of content you are expecting (via the --accept attribute)
- Should you need to debug the cli command then add the --debug attribute
aws sagemaker-runtime --debug --region <region-name>
invoke-endpoint --endpoint-name <sagemaker-endpoint-name>
--body "[[[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]]"
--content-type "application/json" --accept "application/json" output.json
2. Jupyter Notebook
One great thing about Jupyter notebook is that you can add inline code easily. So if you need to increase the verbosity of logging, say for debugging purposes, you can quickly do that via the following 2 lines:
try:
import http.client as http_client
except ImportError:
# Python 2
import httplib as http_client
http_client.HTTPConnection.debuglevel = 1
3. Cost Control
Once you finished with the work on SageMaker, make sure that you stop the notebook or else you will pay hosting fees as per the Amazon SageMaker pricing. If you don't need the Endpoint, then you should delete it or else you will pay for hosting it. Note, that you can re-create the endpoint when you need it, as soon as you don't delete the model (and the endpoint configuration).
Gotchas
Note that if you get the following error:
botocore.errorfactory.ModelError: An error occurred (ModelError)
when calling the InvokeEndpoint operation: Received client error (406)
from model with message ""Requested unsupported ContentType in Accept: */*"".
See https://<region-name>.console.aws.amazon.com/cloudwatch/home?
region=<region-name>#logEventViewer:group=/aws/sagemaker/Endpoints/<endpoint-name>
in account <account-name> for more information.
... and all you have in the CloudWatch log file is:
10.32.0.2 - - [24/Apr/2018:14:27:25 +0000] "POST /invocations HTTP/1.1" 406 50 "-" "AHC/2.0"
then add the --accept attribute to the cli (e.g. --accept "application/json")
