Jump to conversation

Unresolved conversations (5)

ricardoV94

Jan 14, 2023

Todo: Check we are not missing performance by not having explicit sequences. Todo: When there are multiple sequences PyTensor defines n_steps as the shortest sequence. JAX should be able to handle this, but if not we could consider not allowing sequences/n_steps with different lengths in the Pytensor scan. Then we could pass a single shape as n_steps after asserting they are the same?

pytensor/link/jax/dispatch/loop.py

ricardoV94

Jan 13, 2023

This gives an annoying Supervisor Feature missing warning... gotta clean that up

pytensor/link/jax/dispatch/loop.py

ferrine

Jan 13, 2023

Why not graph_replace or using memo for FunctionGraph(memo={symbolic_idx: idx}) ([here](https://github.com/pymc-devs/pytensor/pull/191/files#diff-1a63681fc5a000e63de6f9215e28d43c9e8066fc0a976b5da24499a671cadd11R121))?

pytensor/loop/basic.py

Ricardo Vieira

ferrine

Jan 13, 2023

What about subclassing Scan into * `Map(Scan)` * `Reduce(Scan)` * `Filter(Scan)` It will be easier to dispatch into optimized implementations

pytensor/loop/basic.py

Ricardo Vieira

ricardoV94

Jan 10, 2023

TODO: Add mixin `HasInnerGraph` so that we can see the inner graph in debug_print

pytensor/loop/op.py

Resolved conversations (1)

ricardoV94

Jan 11, 2023

Hmm, I can't do this or else it won't be possible to reference the outer graph (e.g., to request gradients wrt non-sequences) as is don't by the jacobian function

Outdated

pytensor/loop/basic.py

Ricardo Vieira and Adrian Seyboldt