Add Bayesian VAR(2) example script #1658 #1915
Conversation
|
This looks interesting! As of #1904, there is a Gaussian state space model distribution. Would that be suitable for the VAR here? |
|
In addition, could you add the example to docs index together with thumbnails (like in #1429)? |
|
Hi @fehiepsi, added the index and thumbnails. |
Hi @tillahoffmann, awesome work on SSM, do you want me to create example using SSM? |
|
@fehiepsi , Lint has failed, when I am fixing with ruff, it is changing some more files other that var2.py |
I think using the current implementation in this PR or SSM could both work. The latter might be a bit more concise because you wouldn't have to implement the transition function. |
|
@aibit0111 Could you merge with master? lint issues are fixed upstream. |
|
@fehiepsi Done |
|
Could you add some docstring like other examples to give the readers a take on the content? |
aibit0111
force-pushed
the
bayesian-var-tutorial
branch
from
ed19c33 to
c0cf4c4
Compare
|
Hi @fehiepsi, i was on holiday, updated the code with proper doc and passed all the test. |
There was a problem hiding this comment.
Thanks, @aibit0111! The implementation looks great to me. Just have minor comments.
Could you also mention [Forecasting](https://num.pyro.ai/en/stable/tutorials/time_series_forecasting.html#Forecasting) in the description so that users have a pointer?
examples/var2.py
Outdated
| L_omega = numpyro.sample( | ||
| "L_omega", dist.LKJCholesky(dimension=K, concentration=1.0) | ||
| ) | ||
| L_Sigma = jnp.matmul(jnp.diag(sigma), L_omega) |
There was a problem hiding this comment.
maybe jnp.einsum("...i,...ij->...ij", sigma, L_omega) or sigma[..., None] * L_omega to improve the performance
examples/var2.py
Outdated
| # Priors for constants and coefficients | ||
| c = numpyro.sample("c", dist.Normal(0, 1).expand([K])) # Constants vector of size K | ||
| Phi1 = numpyro.sample( | ||
| "Phi1", dist.Normal(0, 1).expand([K, K]) |
There was a problem hiding this comment.
could you add .to_event(2) at the end to make sure that no batch dimension appears here?
examples/var2.py
Outdated
| "Phi1", dist.Normal(0, 1).expand([K, K]) | ||
| ) # Coefficients for lag 1 | ||
| Phi2 = numpyro.sample( | ||
| "Phi2", dist.Normal(0, 1).expand([K, K]) |
examples/var2.py
Outdated
|
|
||
| # Priors for error terms | ||
| with numpyro.plate("K", K): | ||
| sigma = numpyro.sample("sigma", dist.HalfNormal(1.0)) # Standard deviations |
There was a problem hiding this comment.
like above, prefer .expand([K]).to_event(1) than using plate
|
Thanks @fehiepsi for generous review, I am new to num pyro, learning a lot. Please review once again |
VAR(2) examples:
Reference - https://otexts.com/fpp2/VAR.html
fix #1658