InvalidArgumentError (see above for traceback): indices[0] = 1 is not in [0, 1) about deeplift HOT 3 CLOSED

If you have only one task, then rather than iterating over for task_idx in range(10):, just always have task_idx=0. Let me know if this addresses the issue.

from deeplift.

Thanks @AvantiShri, that worked. Thanks again for not only making this awesome contribution, but providing support.

Before we close this ticket out, can you please elaborate on:

1. In what unit is the attribution scores/value of each algorithm (specifically DeepLIFT, IG-5)? For instance, I know that SHAP attribution values are in log-order. I'm trying to understand how to compare attribution scores across various algorithms. Essentially is a .5 from DeepLIFT equivalent to .5 from IG etc.

2. How does the implementation of IG in DeepLIFT compare with DeepExplain or this or that. Out of curiosity, trying to figure out how you have validated your implementation of IG. I plan to run some benchmarks, just figured I ask.

3. During your talk @NVIDIA GTC, you mentioned an additional benefit of using DeepLIFT besides feature attribution for an instance, is using DeepLIFT to find patterns. Can you point me to some examples in a notebook where you found patterns. I believe you mentioned the genomics example maybe...

from deeplift.

Hi @donigian,

Glad the fix worked, and thanks for your kind words!

(1) The DeepLIFT attribution scores sum up linearly; the sum of attribution scores equals the difference from reference. A 0.5 from DeepLIFT is analogous to a 0.5 from IG.
(2) I tested my IG implementation using this unit test: https://github.com/kundajelab/deeplift/blob/671ee67a03bd5bebf4c405af59eec45d3ca2a288/tests/blobs/test_integrated_grads.py. I haven't looked at other implementations of IG, but I expect they should all produce similar results within numerical precision. The IG algorithm is straightforward to implement - just average the gradient at linearly interpolated points, then multiply by the difference-from-reference.
(3) That was actually a different algorithm - TF-MoDISco. I have released code for it but no preprint since it's still in development. Here is the repo: https://github.com/kundajelab/tfmodisco

from deeplift.