"No module named PIL" after "RUN pip3 install Pillow" in docker container; neither PIL nor Pillow present in dist-packages directory

Question

I'm following this SageMaker guide and using the 1.12 cpu docker file. https://github.com/aws/sagemaker-tensorflow-serving-container

If I use the requirements.txt file to install Pillow, my container works great locally, but when I deploy to SageMaker, 'pip3 install' fails with an error indicating my container doesn't have internet access.

To work around that issue, I'm trying to install Pillow in my container before deploying to SageMaker.

When I include the lines "RUN pip3 install Pillow" and "RUN pip3 show Pillow" in my docker file, when building, I see output saying "Successfully installed Pillow-6.2.0" and the show command indicates the lib was installed at /usr/local/lib/python3.5/dist-packages. Also running "RUN ls /usr/local/lib/python3.5/dist-packages" in the docker files shows "PIL" and "Pillow-6.2.0.dist-info" in dist-packages, and the PIL directory includes many code files.

However, when I run my container locally, trying to import in python using "from PIL import Image" results in error "No module named PIL". I've tried variations like "import Image", but PIL doesn't seem to be installed in the context in which the code is running when I start the container.

Before the line "from PIL import Image", I added "import subprocess" and 'print(subprocess.check_output("ls /usr/local/lib/python3.5/dist-packages".split()))'

This ls output matches what I get when running it in the docker file, except "PIL" and "Pillow-6.2.0.dist-info" are missing. Why are those two in /usr/local/lib/python3.5/dist-packages when I run the docker file but not when my container is started locally?

Is there a better way to include Pillow in my container? The referenced Github page also shows that I can deploy libraries by including the files (in code/lib of model package), but to get files compatible with Ubuntu 16.04 (which the docker container uses; I'm on a Mac), I'd probably copy them from the docker container after running "RUN pip3 install Pillow" in my docker file, and it seems odd that I would need to get files from the docker container to deploy to the docker container.

My docker file looks like:

ARG TFS_VERSION

FROM tensorflow/serving:${TFS_VERSION} as tfs
FROM ubuntu:16.04
LABEL com.amazonaws.sagemaker.capabilities.accept-bind-to-port=true

COPY --from=tfs /usr/bin/tensorflow_model_server /usr/bin/tensorflow_model_server

# nginx + njs
RUN \
    apt-get update && \
    apt-get -y install --no-install-recommends curl && \
    curl -s http://nginx.org/keys/nginx_signing.key | apt-key add - && \
    echo 'deb http://nginx.org/packages/ubuntu/ xenial nginx' >> /etc/apt/sources.list && \
    apt-get update && \
    apt-get -y install --no-install-recommends nginx nginx-module-njs python3 python3-pip python3-setuptools && \
    apt-get clean

RUN pip3 install Pillow

# cython, falcon, gunicorn, tensorflow-serving
RUN \
    pip3 install --no-cache-dir cython falcon gunicorn gevent requests grpcio protobuf tensorflow && \
    pip3 install --no-dependencies --no-cache-dir tensorflow-serving-api 

COPY ./ /

ARG TFS_SHORT_VERSION
ENV SAGEMAKER_TFS_VERSION "${TFS_SHORT_VERSION}"
ENV PATH "$PATH:/sagemaker"

RUN pip3 show Pillow
RUN ls /usr/local/lib/python3.5/dist-packages

I've tried installing Pillow on the same line as cython and other dependencies, but the result is the same...those dependencies are in /usr/local/lib/python3.5/dist-packages both at the time the container is built and when the container is started locally, while "PIL" and "Pillow-6.2.0.dist-info" are only present when the container is built.

Answer 1

Apologies for the late response.

If I use the requirements.txt file to install Pillow, my container works great locally, but when I deploy to SageMaker, 'pip3 install' fails with an error indicating my container doesn't have internet access.

If restricted internet access isn't a requirement, then you should be able to enable internet access by making enable_network_isolation=False when instantiating your Model class in the SageMaker Python SDK, as shown here: https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/model.py#L85

If restricted internet access is a requirement, this means that you will need to either install your dependencies in your own container beforehand or make use of the packaging as you mentioned in your correspondence.

I have copied your provided Dockerfile and created an image to run as an image in order to reproduce the error you are seeing. I was not able to reproduce the error as quoted below:

However, when I run my container locally, trying to import in python using "from PIL import Image" results in error "No module named PIL". I've tried variations like "import Image", but PIL doesn't seem to be installed in the context in which the code is running when I start the container.

I created a similar Docker image and ran it as a container with the following command:

docker run -it --entrypoint bash <DOCKER_IMAGE>

from within the container I started a Python3 session and ran the following commands locally without error:

root@13eab4c6e8ab:/# python3 -s
Python 3.5.2 (default, Oct  8 2019, 13:06:37)
[GCC 5.4.0 20160609] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from PIL import Image

Can you please provide the code for how you're starting your SageMaker jobs?

Please double check that the Docker image you have created is the one being referenced when starting your SageMaker jobs.

Please let me know if there is anything I can clarify.

Thanks!