Changing RGCN/ExpGrad reshape and sampling

[jdiaz4302]

Made reshape and sampling changes discussed in #13 here, to the 03_04 notebook (using Expected Gradients on river-dl data).

Simon and I previously found a difference in compute time that was a result of different river-dl data (different sequence length 360 vs 180 and number of input features 16 vs 7), so I did subset my data to match his for faster compute.

This seems to work well, potentially with a less strong exact convergence (see the cellblock plotting expected_gradiants_ls1 vs expected_gradiants_ls2), but both notebooks are showing consistent/near-identical results (i.e., I'm not concerned about it)

[jdiaz4302]

Since sampling random combinations of seg/years is the giving the same answers as sampling random years (same for all segs), I think we can choose to do the latter since we were leaning that way anyways

[jdiaz4302]

I can make these changes to the xai_utils.py file pending any discussion. The changes would be:

• Fix data reshaping
• Sample the same random year for all segments

These changes can already be seen in the 03_04_02 notebook included in this PR

[jdiaz4302]

One thing to note is that if you look at the compute time testing (%%time) in the notebooks, it will give misleading results and comparison because the notebook displaying initially faster compute time was running exclusively at first, but then the other notebook (03_04_02) was running concurrently (slowing them both down, but appears to show 03_04_02 being slower across the broad). If you look at the last %%time usage in both notebooks, it shows the real comparison - that they are the same. Each timing would be better if I had not been running them together (using a bottlenecked local CPU)