[ENH] EXPERIMENTAL: Example notebook based on the new data pipeline

[phoeenniixx]

Description

This PR adds example notebook for the new v2 data pipeline vignette, having the basic implementation of the tft model using this version. For more info see #1812 , #1811

Colab link: https://colab.research.google.com/drive/148MyhcNfYEh4CZ6vBXLqQNsUBF0n6_0v?usp=sharing

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[phoeenniixx]

Hi @fkiraly, I am getting this error:

I just downloaded the notebook from colab and pasted it in the repo, is there anything else I should do to avoid this? Really have no idea 😅

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Please upload report for BASE (main@cfc7fc6). Learn more about missing BASE report.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1813   +/-   ##
=======================================
  Coverage        ?   83.14%           
=======================================
  Files           ?       61           
  Lines           ?     6153           
  Branches        ?        0           
=======================================
  Hits            ?     5116           
  Misses          ?     1037           
  Partials        ?        0           

Flag	Coverage Δ
cpu	83.14% <ø> (?)
pytest	83.14% <ø> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[xandie985]

Hi @phoeenniixx , the implementation looks good and insightful. Here are some questions that I have about the implementation. Its possible that there is difference between the objectives we had with DSIPTS and SKTIME, still I would like to discuss about your pov wrt sktime.

1. EncoderDecoderTimeSeriesDataModule assumes that the data would fit int he memory. How would your approach scale to datasets that do not fit into RAM? Are there plans to incorporate memory-efficient loading strategies like chunking or on-demand loading from disk?
2. Does your module have any mechanisms to detect or correct irregular time series within its scope?
3. Your _create_windows method checks for sufficient sequence length.. how does your module handle missing values within a window, and what are the potential consequences for model training if windows contain significant NaNs?
4. Adding random seed for reproducibility, especially in setup() where random shuffling takes place.

[phoeenniixx]

Thank you for the review @xandie985!

EncoderDecoderTimeSeriesDataModule assumes that the data would fit int he memory. How would your approach scale to datasets that do not fit into RAM? Are there plans to incorporate memory-efficient loading strategies like chunking or on-demand loading from disk?

So right now we assume that data could fit in the memory, but yes, in future we plan to add features like chunking, on-demand loading etc

2. Does your module have any mechanisms to detect or correct irregular time series within its scope?
3. Your _create_windows method checks for sufficient sequence length.. how does your module handle missing values within a window, and what are the potential consequences for model training if windows contain significant NaNs?
4. Adding random seed for reproducibility, especially in setup() where random shuffling takes place.

These are some open questions, we still need to work on - We will tackle these questions in future iterations once an end-to-end prototype is ready and we get some reviews from the users of the package on this prototype.

[fkiraly]

More detailed review.

• please remove the install from the start of the notebook
• we should test that this is running, while we are working on v2. One way is to move the content to docs/examples/tutorials, the contents of which are automatically run an tested.
• the data generation cell is useful, but not too illustrative. Can you move the code to a function load_toydata or similar, in pytorch_forecasting.data, new module, e.g., toydata? Then we can also use this in testing later!
• can you add basic markdown cells that explain what the notebook is showing, and what each steps are? E.g., a summary at the top of the multiple steps, and then again small headers for the steps with minimal explanations.

[phoeenniixx]

Thanks! I would make the changes accordingly, Just one doubt:

the data generation cell is useful, but not too illustrative. Can you move the code to a function load_toydata or similar, in pytorch_forecasting.data, new module, e.g., toydata? Then we can also use this in testing later!

I think we can add it to pytorch_forecasting.data.examples? Like right now people import get_stallion_data from there, so they can import toydata from there as well. Just think that this would follow an already available mapping of "test data" to examples...

[fkiraly]

Like right now people import get_stallion_data from there, so they can import toydata from there as well.

Makes sense, to add it to the established location with data loaders.

Would it make sense to split the file up and have on loader per file? Need not be done in this PR.

[phoeenniixx]

Would it make sense to split the file up and have on loader per file?

Then I think we need to should create a new folder called loaders or datasets and have these files there, and we can add more loaders to that folder in future