v0.1.7

What's Changed

• 34 bug fix reshape observable method in bembpy by @TianyuDu in #36
• Other major changes were implemented in torch-choice v1.0.5, please have a look at the change log here: https://github.com/gsbDBI/torch-choice/releases/tag/v1.0.5

Full Changelog: v0.1.6...v0.1.7

BEMB v.0.1.6

What's Changed

Users can now specify the optimizer they want using the model_optimizer argument while initializing the model as following:

bemb = LitBEMBFlex(... model_optimizer="Adam", ...)
bemb = LitBEMBFlex(... model_optimizer="LBFGS", ...)

The optimizer specified needs to be in torch.optim.

We have developed a cleaner mode estimation pipeline blackened by PyTorch-Lightning:

from bemb import run
run(bemb, dataset_train=dataset_train, dataset_val=dataset_val, dataset_test=dataset_test, batch_size=len(dataset_train) // 20, num_epochs=1000, device="cuda")

import the run directly from bemb package to use it.

• fix typo. by @TianyuDu in #20
• commits for supermarket and deterministic vi by @kanodiaayush in #23
• 30 add lbfgs support and cleaner pytorch lightning training loops by @TianyuDu in #32

New Contributors

• @kanodiaayush made their first contribution in #23

Full Changelog: v0.1.5...v0.1.6

BEMB v.0.1.5

We have updated several features of our package.

What's Changed

• We allow for partial prior_variance dictionary, for example, you can specify prior_variance={'alpha': 0.3, 'default': 1.5} to set prior variance of alpha coefficient to 0.3 while all other coefficients will have prior variance 1.5.
  …to have unit variance.
• For label predictions (pred_item=False), we now include more performance metrics.
• We have updated the utility formula parser mentioned in this issue.
• We have added a posterior_distribution() to query posterior/variational distribution of coefficient directly.
• We have fixed known issues on sherlock mentioned in this issue.

Full Changelog: v0.1.4...v0.1.5

BEMB release v0.1.4

• Update: the forward() method of BEMB module has been updated for easier inference. The new forward() method can return both log-probabilities and utility values.
• For use cases of predicting binary labels (pred_item = False), now the returned log-probabilities are always the predicted log-probability of the actual label. To get the predicted log-probabilities of the positive class for all observations, please get the utility first and apply a log-sigmoid transformation on utility values.
• We have also added a helper function predict_proba() that gives (1) if pred_item=True, predicted probability of choosing each item among items in its category, so output shape is (batch_size, num_items) (2) if pred_item=False predicted probabilities of both label = 0 and label = 1 (so output shape is (batch_size, 2).
• Please refer to the Jupyter notebook for more details.

v0.1.3

Add __init__.py to the bemb.utils sub-module as well.

v0.1.2

Basically no change since v0.1.1, with a bit improvements.

v0.1.1

Merge branch 'main' of github.com:gsbDBI/bemb