dbvis-ukon / ts-mule Goto Github PK

• Initial setup for sphinx
• Installation
• Quick Start Tutorial
• Segmentation Tutorial
• Evaluation Tutorial
• Rewrite docstring in Modules. Apply restructedText in docstring
• Prepare hooks for publishing to readthedocs.io

Note:

• do not delete branch gh-pages -> this branch is reserved for building docs generated by sphinx

Write it!

The notebook test_segmentation examples should be moved to demo and renamed to -> demo_segmenation.

Examples/demo is a similar term, choose one

add data to project

start = 0
end = 1000

X = dataset_test[0][start :end]
y = dataset_test[1][start :end]

In, e.g., Beijing_Air_Quality_2_5_evaluation_Keras.ipynb Cell 5

@merowech

I push a branch docs-build-test (do not merge this branch), which contains HTML built from docs. I cannot create github-pages from this repo, hence I ignore 'Settings/Pages'

To view it:

git checkout origin/doc-build-test
open docs/build/html/index.html

Currently, only scratch of tutorial, feel free to add or request.

Local Mean Evaluation expects segments

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-43-0a14d03652c1> in <module>
      5                         eval_fn=metrics.mean_squared_error,
      6                         percentile=90,
----> 7                         delta=0.1
      8                         )
      9 scores

/c/Users/Udo/git/ts-mule/tsmule/xai/evaluation.py in analysis_relevance(self, X, y, R, predict_fn, eval_fn, replace_method, percentile, delta)
    138                             percentile=percentile,
    139                             )
--> 140         X_percentile = np.array(list(X_percentile))
    141 
    142         X_random = self.perturb(X, R, 

/c/Users/Udo/git/ts-mule/tsmule/xai/evaluation.py in perturb(self, X, R, replace_method, percentile, shuffle, delta)
    108             else:
    109                 m = self.mask_percentile(r, percentile)
--> 110             yield self._perturb(x, m, replace_method=replace_method)
    111 
    112 

/c/Users/Udo/git/ts-mule/tsmule/xai/evaluation.py in _perturb(x, m, replace_method)
     79         """
     80         repl_fn = getattr(repl, replace_method)
---> 81         r = repl_fn(x)
     82         assert x.shape == m.shape == r.shape
     83         z = x * m + r * (1 - m)

TypeError: local_mean() missing 1 required positional argument: 'segments'

New ones (ours)

• Using matrix profile
• Using SAX

Old ones (related work)

• Frequency band based

After the paper is accepted, we will publish it to PyPi

Reference: https://realpython.com/pypi-publish-python-package/

Required:

• Documents uploaded to readthedocs
• Auto deploy with GitHub actions (if applicable)
• Auto Test
• Auto Doctest (if applicable)

This task is reserved after the repo is published

We decide to use readthedocs as a host for documentation, instead of using GitHub pages (administrating limitation)

• Setup environment.yaml for readthedocs.io
• Prepare pre-commit hook if necessary
• Create and upload to tsmule.readthedocs.io

• Zero
• Inverse-max
• Inverse-mean
• Mean
• Random
• Randomization

@merowech
The structure of src/tsmule should be renamed to tsmule or src only because

• easy to create a package later, and upload it to pypi.org

Another way is to put init.py in src
which way do you prefer?

Currently I have an issue with running tests due to this issue

• Zero
• Global Mean
• Local Mean
• Noise

Cannot merge sub branch from feature/perturbation to -> sampling

• Uniform segmentation and perturbation for it

According to
https://github.com/dbvis-ukon/ts-mule/blob/main/demo/beijing_air_2_5/Beijing_Air_Quality_2_5_CNN_explain_Keras.ipynb

the relevance valued at zeros for cnn and dnn models, that is unexpected result.

@merowech Could you please reupload the models?
I am writing the quick-start documents and would like to reuse our pre-trained models to demo and show the usage.

• Reupload pre-trained models

I suggest a structure as we had before :
examples/<data-set-name>/<name_model>.h5
examples/<data-set-name>/<name_testset>.something

In my docs tutorial-> I will

1. load model from our repo (remote/local).
2. load data set from our repo (remote/local).
3. show how we use tsmule to generate explanations
4. show how we use tsmule to evaluate the explanation