478 rise for time series

[cwmeijer]

Add RISE for timeseries.

[cwmeijer]

I added RISE and 2 tests for it. I didn't test it with actual data or with an actual model. There is also something not right with the output format. It is inconsistent with the timeseries visualization. RISE is now returning a saliency map as an np array while the visualization is expecting a dict with starts and stops and weights. Therefore, the test_common_usage.py contains some transformation code.

@geek-yang, maybe you would like to have a look what should be a good general format for timeseries saliency. I suppose the dict option is not bad and quite general, but it contains an 'index' key which I don't understand why that's useful. Also, maybe creating a dataclass for this is better than dicts.

[geek-yang]

I added RISE and 2 tests for it. I didn't test it with actual data or with an actual model. There is also something not right with the output format. It is inconsistent with the timeseries visualization. RISE is now returning a saliency map as an np array while the visualization is expecting a dict with starts and stops and weights. Therefore, the test_common_usage.py contains some transformation code.

@geek-yang, maybe you would like to have a look what should be a good general format for timeseries saliency. I suppose the dict option is not bad and quite general, but it contains an 'index' key which I don't understand why that's useful. Also, maybe creating a dataclass for this is better than dicts.

I looked into the implementation (also with dummy data given in the test, not real model) and it worked well. The output shape of saliency map is exactly the same as the one from rise explainer for image, which is [label2explain, signal length, number of variables] in our case. I prefer to keep it consistent among all RISE approaches. I think we can add a transformation function, just like what you did.

On the other hand, the current implementation of visualization #491 doesn't seem to be compatible with multivariate timeseries. We need to update it as well later. I think for now we can test the implementation with weather dataset, to see if we could get something plausible (well... given our native segmentation strategy...or let's say...in this case no segmentation strategy, the results could be very bad...But at least we have a starting point for the implementation of skillful strategies).

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[geek-yang]

Awesome work! I really like the idea about having the hot and cold day dummy data to test if the method works as expected or not. This solves our big puzzle and it can be applied to test other methods as well. Well done!

[geek-yang]

Follow our discussion, I have created a new task on the board related to the notebook #539 . I will polish the notebook and add an ONNX example there.