Turing 0.41. #661
Turing 0.41. #661
Conversation
|
Preview the changes: https://turinglang.org/docs/pr-previews/661 |
| w = [0.5, 0.5] | ||
| μ = [-3.5, 0.5] | ||
| mixturemodel = MixtureModel([MvNormal(Fill(μₖ, 2), I) for μₖ in μ], w) | ||
| μ = [-2.0, 2.0] | ||
| mixturemodel = MixtureModel([MvNormal(Fill(μₖ, 2), 0.2 * I) for μₖ in μ], w) | ||
| # We draw the data points. | ||
| N = 60 | ||
| N = 30 | ||
| x = rand(mixturemodel, N); |
There was a problem hiding this comment.
Having 30 observations just makes life easier because it triggers fewer resampling steps with PG. To compensate for that I made the data more tightly clustered to make sure that the results are still somewhat meaningful.
There was a problem hiding this comment.
locally, this cuts the time taken to run this tutorial to about 12 minutes. I'm not sure what it was before but I think it was certainly over 20 minutes.
| model = linear_regression(train, train_target) | ||
| chain = sample(model, NUTS(), 5_000) | ||
| chain = sample(StableRNG(468), model, NUTS(), 20_000) | ||
| ``` |
There was a problem hiding this comment.
I don't actually know what went wrong here, i struggled to get this to work without bumping the samples. It might just be that the 5000 samples one just happened to be a very lucky seed. It takes very little time because NUTS is quite fast anyway.
| @assert isapprox(bayes_train_loss, ols_train_loss; rtol=0.01) "Difference between Bayesian training loss ($bayes_train_loss) and OLS training loss ($ols_train_loss) unexpectedly large!" | ||
| @assert isapprox(bayes_test_loss, ols_test_loss; rtol=0.05) "Difference between Bayesian test loss ($bayes_test_loss) and OLS test loss ($ols_test_loss) unexpectedly large!" | ||
| @assert bayes_train_loss > ols_train_loss "Bayesian training loss ($bayes_train_loss) <= OLS training loss ($bayes_train_loss)" | ||
| @assert bayes_test_loss < ols_test_loss "Bayesian test loss ($bayes_test_loss) >= OLS test loss ($ols_test_loss)" |
There was a problem hiding this comment.
Do you know if it was just happenstance that these actually pretty much matched each other before?
There was a problem hiding this comment.
Err, not sure to be honest. I think the atols were quite fragile, even changing the way the test/train split was done led to things failing there. So I'd say yes.
I think pretty much all that is needed is fixing
initial_params.Edit: Oh, there was
istrans->is_transformedtoo.Edit 2: Oh no, there's SamplingContext 😭