Add draft notebook

[jessegrabowski]

DONT MERGE

---

📚 Documentation preview 📚: https://pytensor--1.org.readthedocs.build/en/1/

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

ricardoV94 commented on 2025-08-17T14:21:08Z
----------------------------------------------------------------

Why no exact solution available?

[ricardoV94]

lost a full comment to nbviewer yay. couple of points:

1. We should mention Theano due to our talk title.
2. Add a short prediction intervention example, before the whole graph manipulation section. Reduce price effect in a period (people less price-sensitive on xmas or whatever?) Show how prediction changes.
3. A bit more blah on compilation, different backends, memory optimzation, fusion.
4. Simplify the constrained optimization by using pt.as_tensor(observed_price)[100:250].set(free_prices), then the optimization is only a function of free_prices, and the code can just use explicit_graph_inputs? Not strong on this one
5. Shouldn't it still see a linear graph even though price is now using set_subtensor? Right, the price root variable is now not such a simple function. Perhaps mention. Mathematically it's still linear in the subset we're optimizing right?
6. Mention who is using pytensor and why at the end?

[ricardoV94]

Also: I suggest showing some numbers when we compile a function. makes it less abstract. Like show 2 loss evals before throwing scipy at it

[ricardoV94]

Missing:

1. Slide with talk title (and conference)
2. Node picture - bipartite graph (got lost at some point)
3. numpy code (what the users want to write) should use sum instead of mean to match the pytensor
4. dprint the graph after the first transformation: stable_loss.dprint(id_type='')