Deliverable 2 - Image Processing and Deep Learning Environment
Summary
Set up an environment with image processing and deep learning functionalities.
It should be sufficiently sequestered from the main machine while also
support easy integration with the GPU and CUDA framework.
It should not install (esp not as root) anything to the host machine directly.
This environment would contain the deep learning framework Pytorch and
the open CV image processing and manipulation framework. Pytorch has been
simple to set up with a virtual environment. But open CV proved to wreak havoc
on the host machine's environment by overriding install paths and installing
other folders in unexpected places.
A solution was needed to set up these dependencies for object detection
that could support all the functionalities of both frameworks, but leave
the host machine untouched.
The solution is to use Docker to contain the dependencies.
Reasons for Docker
Docker containerizes things. This means OS-level virtualization of
isolated applications and services running on a single host instance sharing
the kernel. This is the main difference between Docker and a virtual machine,
which runs a separate OS and kernel on top of a Hypervisor.
By containing the application, the host machine does not have global
configurations, install paths, or other settings hijacked by an application.
Applications and services are free to do what they need in their containers,
and the host is unaffected.
The following are some key advantages of Docker.
- Flexible - Even the most complex applications can be containerized.
- Lightweight - Containers leverage and share the host kernel.
- Interchangeable - You can deploy updates and upgrades on-the-fly.
Setup
Here are instructions on how to set up Pytorch
Dependencies
- Docker
- Docker-compose (orchestrator for multiple Docker containers)
Inside the Environment - Pytorch and Open CV
Pytorch
Pytorch was covered in a previous deliverable that
classified 0-4 and 5-9 digits.
Open CV
Some Open CV functionalities found are covered in
this collection of Open CV slides
such as the following image transformations.
- Rotation
- Translation
- Resize
- Concatenation
- Affine
There is also some logic open CV supports to convert videos into
a series of frame-by-frame images. If there is camera recording input,
this can be converted to a series of images, then fed into the neural net.
Video is the next level of training input for ML models, and Facebook has been
a forerunner in that respect with its massive data stores of video content.
It makes sense to have this functionality handy, especially for testing.