Prerequisites

The following prerequisites apply to most data domains:

Annotations#

In order to annotate data (e.g., for object detection, instance segmentation or image segmentation), you should download the ADAMS Annotator application (you only need Java 11 or 17 installed):

Snapshot

Format conversions#

For turning annotations from one format into another, we will utilize the image-dataset-converter Python library:

Since we will use this library through Docker images, there is no need to install anything. If you should decide to install it locally, it is recommended to do so in a virtual environment

Hardware#

For building models, a computer with NVIDIA GPU (8GB+) and Linux operating system is recommended. Only image classification models can be built in a reasonable amount of time on CPU-only machines.

Directory structure#

Create the following directory structure for the examples of this tutorial:

|
+-- applied_deep_learning
    |
    +-- cache
    |   |
    |   +-- torch
    |   |
    |   +-- iopath_cache
    |
    +-- data
    |
    +-- models
    |
    +-- output
    |
    +-- predictions
        |
        +-- in
        |
        +-- out

Use this command-line to create the directories:

mkdir -p \
  applied_deep_learning/cache/torch \
  applied_deep_learning/cache/iopath_cache \
  applied_deep_learning/data \
  applied_deep_learning/models \
  applied_deep_learning/output \
  applied_deep_learning/predictions/in \
  applied_deep_learning/predictions/out

Docker notes#

To make the Docker commands as easy as possible, they will all get issued from the within the applied_deep_learning directory. In order to get access to the data, output and predictions directories, we use the following mapping:

-v `pwd`:/workspace

This will map the applied_deep_learning directory onto the /workspace directory within the container.

If you have spaces in any of the parent directories, you need to use double quotes around the part before the : (otherwise you will get an error like docker: invalid reference format.):

-v "`pwd`":/workspace

Since Docker usually runs as root within the container, we want to make sure that the user group of any files that get generated are being owned by the current user. This can be achieved by adding the following to the Docker command:

-u $(id -u):$(id -g) -e USER=$USER