Feat/stochastic inputs #833
Conversation
hrzn
requested review from
tomasvanpottelbergh,
dennisbader and
brunnedu
as code owners
Codecov Report
@@ Coverage Diff @@
## master #833 +/- ##
==========================================
- Coverage 91.33% 91.32% -0.02%
==========================================
Files 69 69
Lines 6869 6872 +3
==========================================
+ Hits 6274 6276 +2
- Misses 595 596 +1
Continue to review full report at Codecov.
|
| preds = [model.predict(series=stochastic_series, n=10) for _ in range(2)] | ||
|
|
||
| # random samples should differ | ||
| self.assertFalse(np.alltrue(preds[0].values() == preds[1].values())) |
There was a problem hiding this comment.
This is nitpicking but isn't there a slim chance that the predictions will be identical? :D we could add a seed to make sure they are not.
There was a problem hiding this comment.
Yes, normally there is a seed earlier in the file which I believe should apply here :)
| @@ -1176,6 +1176,28 @@ def values(self, copy=True, sample=0) -> np.ndarray: | |||
| else: | |||
| return self._xa.values[:, :, sample] | |||
|
|
|||
| def random_component_values(self, copy=True) -> np.array: | |||
There was a problem hiding this comment.
What do you think about an additional optional to randomly sample a quantile per sample?
I would assume this is similar to quantile regression based on stochastic input (without needing QuantileRegression as a likelihood)
There was a problem hiding this comment.
Oh, you mean returning the exact quantile instead of sampling a sample at random? That's an interesting idea :) But would require a bit of adaptation elsewhere as we would still need to ensure this is applied on the output targets only, and not e.g. on the inputs or covariates. Let's keep the idea though.
Handle stochastic series for
fit()andpredict()in Torch-based models.When fitting: a sample is drawn from the series uniformly at random (instead of taking always the 1st sample, as before). This way the training procedure can naturally be carried over the stochastic samples.
When predicting: one sample is drawn from the series uniformly at random. Unfortunately at the moment it would be quite complicated to batch the forward passes over all of a series' samples, so each call to
predict()will only return the forecast related to this one sample. However, callingsample()many times and concatenating the results allows to get stochastic forecasts even with deterministic models (not using a likelihood) when predicting stochastic series (or using stochastic covariates). Here's an example: