Strided dataset feature branch

[Diego-Llanes]

This pull request contains the StridedDataset feature addition. The only file that have additions / changes is src/neuromancer/dataset.py.
All of the class parameters are well documented and have docstrings.

EDIT: There is now also an example documenting the usecase of this new feature.

Contributors:

• @Seth1Briney
• @HarryLTS
• @Diego-Llanes

[madelynshapiro]

@Diego-Llanes would you please upload a small example (can be a modified snippet from one of your full notebooks) demonstrating the usage to accompany the PR?

[Diego-Llanes]

@Diego-Llanes would you please upload a small example (can be a modified snippet from one of your full notebooks) demonstrating the usage to accompany the PR?

Just added!

[madelynshapiro]

Excellent notebook!

[madelynshapiro]

Suggested change

	def remainder(self, n, d):
	@staticmethod
	def remainder(n, d):

[drgona]

The notebook should be modified:

• the first image in the notebook is inconsistent with parameters in the text - create a new image illustrating parameters N, L, and stride
• include training example of simple system ID model demonstrating the use of the new dataset
• include more visualizations of how the data is processed into subsequences - it is not clear what I am looking at in the last plot, add more explanation, always label axes in the plots

[RBirmiwal]

strided_dataset_rahul_test.ipynb.zip

I agree with Jan.
In addition have attempted to play around with the strided dataset to ensure it works. I have created a notebook zipped up to this comment. I am still confused at times:

• In DictDataset, let's say D:= a dictDataset, i can do D['X'] --> 3D tensor
• In StridedDataset S:= StridedDataset, I cannot do S['X'], i have to do S[idx]['X] --> 2D tensors

While the user can figure out how to do proper reshaping and data creation, this is a difference in our API

• what if instead we had S['X'] --> list of subsequence tensors for X
• Also the tensor X is 2D not 3D......

I attempted to do Neural ODE example with StridedDataset. For L >= nsteps looks like it trains, but the performance is not as good as standard DictDataset.

For L < nsteps, it breaks.

I like this functionality, but we need to ensure it works and performs on-par for all the neuromancer use-cases not just Farama DPC. If there are situations where the StridedDataset will fail (as indicated when L > nsteps), then those need to be handled appropriately.

[RBirmiwal]

Also would like an example of a non-trivial update_fn, right now it is

def update_initial_condition(d):
    d['xn_2'] = d["xn"][0:1, :]
    return d

for the case of the neural ode where our keys are "X" and "xn". I don't understand how "xn_2" would play a role/being used. So what are cases where this would be necessary?