Define computation in a function in examples

[mhchia]

What is the PR doing?

In #5 , we allow users to define computation in a function, but we haven't changed the examples.

How to fix it?

Define computation in functions instead of model classes in the examples. However, this PR is not ready yet since our current statistics functions don't work correctly with filtered tensors. For example,

        filter = (X>20)
        num_fil_X = torch.sum(filter.double())
        fil_X = torch.where(filter, X, 0)
        return s.mean(fil_X)

if X has 300 elements, elements filtered out might be 200, but fil_X will still be 300 elements and s.mean result will be incorrect since the number of elements should be 200 instead of 300.

To-Dos

So the following examples stay unchanged and should be changed:

• examples/where+geomean
• examples/where+mean
• examples/where+median
• examples/where+mode
• examples/where+regression
• examples/where+stdev

[mhchia]

@JernKunpittaya I've replaced verifier models with computation functions in all examples. Just as the comments above (and places marked with "FIXME") mentioned: results from where+xxx.ipynb are wrong since we need to add State.where and modify the existing statistics operations to make the result correct

[mhchia]

This should be wrong for calculating filtered_x but I'm not sure how to do it since in the example torch.where is done after torch.cat. Do you have any thoughts on this, or I'm thinking we can just leave it for now and change it after State.where is supported?

[JernKunpittaya]

Thx Kevin! Nw, will open a new PR that specifically address where op!

[mhchia]

We need other filters like fil_mean_X and fil_std_X to make it run correctly. Or probably we can fix it later when State.where is supported?

[JernKunpittaya]

Yes, or maybe we can encapsulate the logic in s.mean to detect if it is addressing the filter data or not. Will open a new PR for this