Add GDN algorithm

[SebastianSchmidl]

Adds the GDN algorithm from https://github.com/d-ailin/GDN (Paper: https://doi.org/10.1609/aaai.v35i5.16523)

Fixes #15

[SebastianSchmidl]

@2er0 can you take it from here?

• add source code within the given template (transform data and set the parameters)
• check parameter definitions in algorithm.py (name, type, default value, and their completeness)
• check parameter definitions in manifest.json (consistent with algorithm.py and descriptions)
• adapt implementation to correctly save the model
• check if the algorithm successfully trains and executes (I can help with that)
• does the algorithm need reverse windowing? add a note in the README

[2er0]

@CodeLionX please test - I found some time to make the first full working draft.

docker build -t registry.gitlab.hpi.de/akita/i/python3-base:0.2.6 ./0-base-images/python3-base 
docker build -t registry.gitlab.hpi.de/akita/i/python3-torch:0.2.6 ./0-base-images/python3-torch
docker build -t registry.gitlab.hpi.de/akita/i/gdn:0.2.6 ./gdn

docker run --rm \                                                                               
    -v $(pwd)/1-data:/data:ro \
    -v $(pwd)/2-results:/results:rw \
  registry.gitlab.hpi.de/akita/i/gdn:0.2.6 execute-algorithm '{
    "executionType": "train",
    "dataInput": "/data/multi-dataset.csv",
    "dataOutput": "/results/anomaly_scores.ts",
    "modelInput": "/results/model.pkl",
    "modelOutput": "/results/model.pkl",
    "customParameters": {}
  }'

docker run --rm \                                                                               
    -v $(pwd)/1-data:/data:ro \
    -v $(pwd)/2-results:/results:rw \
  registry.gitlab.hpi.de/akita/i/gdn:0.2.6 execute-algorithm '{
    "executionType": "execute",
    "dataInput": "/data/multi-dataset.csv",
    "dataOutput": "/results/test_anomaly_scores.ts",
    "modelInput": "/results/model.pkl",
    "modelOutput": "/results/model.pkl",
    "customParameters": {}
  }'

[SebastianSchmidl]

I took a brief look and can execute the algorithm locally. 👍🏼
However, I found some issues in the current draft:

• See inline comments (changes to base image)
• The input handling is too tight: You assume that the feature columns are labeled value*, but many datasets use other feature names.
• All other algorithms store a single model-file. TimeEval currently does not support multiple files. I would suggest that you create an archive when storing the model. You can take a look at Torsk as an example for that.
• Do you perform train-test-splitting within the code? The anomaly score files are very short … TimeEval requires that the scoring has the same length as the input TS. You could use a postprocessing step to perform this transformation (allows you to use Code from TimeEval).

[SebastianSchmidl]

We deliberately pinned the dependencies to ensure reproducibility!

If the new algorithm actually needs a new Python version and different dependencies, we should create a new base image. Otherwise, we have to check if all the other algorithms still work with the new base image.

[SebastianSchmidl]

Please revert those changes to the README. They only apply to your algorithm and not to the others.

[SebastianSchmidl]

There is no version 0.2.6 of the base images yet.

[SebastianSchmidl]

Do you have a link to the bug report/issue? This looks like a dirty hack.