Benjamin Kurt Miller

If only to add a public voice that this is an issue... I have also found that this dependency makes matminer difficult to use alongside other modern packages.

I added a test. it's extremely simple, but it checks that the estimated ratios are the correct size . (like `log_prob`).

I will follow up after Friday.

@tomMoral I think this is done now.

~~There is a wip push, but supporting sample dim will require:~~ ~~- [ ] making warnings that it is not allowed to be used in loss function computations~~

goals: - [x] no abstraction, only one RatioEstimator - ~~[ ] move RatioEstimator out of neural_nets~~ - [x] primarily RatioEstimator keeps track of x and theta shapes - [x] simple...

I will no longer remove `RatioEstimator` from neural_nets because it looks like the conditional density estimators still live there... unless someone tells me that's wrong. Now it has its own...

if this passes the CI, all tests that check anything related to ratios are passing on my computer as well: ``` pytest tests/inference_on_device_test.py tests/linearGaussian_snre_test.py tests/ratio_estimator_test.py ................xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx................xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx.x......................................................................................................... [ 59%] .......................................................xx.................................................................................... [100%]...

Right now `test_lc2st_true_positiv_rate[LC2ST_NF]` is failing, but that shouldn't affect merging this pull request @janfb

@michaeldeistler @janfb @manuelgloeckler @jnsbck I propose we remove the method `loss` from this abstraction. The estimator should be able to be defined without assuming a way to train it. This...