gen_size > 200 leads to NameError: name 'generic_logsumexp' is not defined

[jordycollingwood]

Issue Description:
Running the DOMIAS code with gen_size > 200 results in the following error message with:

Additionally, running BNAF density generator on gen_size <=200 always results in an AUCROC of 0.5. It Appears the p_G_evaluated always resolves to an all-zero array.

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-6-e273f542404d> in <module>
    135                 training_epochs=training_epoch,
    136                 synthetic_sizes=[gen_size],
--> 137                 density_estimator=density_estimator,
    138             )
    139 

~\anaconda3\lib\site-packages\domias\evaluator.py in evaluate_performance(generator, dataset, mem_set_size, reference_set_size, training_epochs, synthetic_sizes, density_estimator, seed, device, shifted_column, zero_quantile, reference_kept_p)
    200 
    201         wd_n = min(len(synth_set), len(reference_set))
--> 202         eval_met_on_reference = compute_wd(synth_set[:wd_n], reference_set[:wd_n])
    203         performance_logger[synthetic_size]["MIA_performance"][
    204             "sample_quality"

~\anaconda3\lib\site-packages\domias\metrics\wd.py in compute_wd(X_syn, X)
     26     OT_solver = SamplesLoss(loss="sinkhorn")
     27 
---> 28     return OT_solver(X_ten, Xsyn_ten).cpu().numpy().item()

~\anaconda3\lib\site-packages\torch\nn\modules\module.py in _call_impl(self, *input, **kwargs)
   1192         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1193                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1194             return forward_call(*input, **kwargs)
   1195         # Do not call functions when jit is used
   1196         full_backward_hooks, non_full_backward_hooks = [], []

~\anaconda3\lib\site-packages\geomloss\samples_loss.py in forward(self, *args)
    281             labels_x=l_x,
    282             labels_y=l_y,
--> 283             verbose=self.verbose,
    284         )
    285 

~\anaconda3\lib\site-packages\geomloss\sinkhorn_samples.py in sinkhorn_online(a, x, b, y, p, blur, reach, diameter, scaling, cost, debias, potentials, **kwargs)
    387             cost = cost_formulas[p]
    388 
--> 389         my_lse = lse_genred(cost, D, dtype=str(x.dtype)[6:])
    390         softmin = partial(softmin_online, log_conv=my_lse)
    391 

~\anaconda3\lib\site-packages\geomloss\sinkhorn_samples.py in lse_genred(cost, D, dtype)
    324     """Legacy "Genred" implementation, with low-level KeOps formulas."""
    325 
--> 326     log_conv = generic_logsumexp(
    327         "( B - (P * " + cost + " ) )",
    328         "A = Vi(1)",

NameError: name 'generic_logsumexp' is not defined

Issue Summary:
Note that this error relates to all density estimation methods, while the domias AUCROC=0.5 issue is unique to BNAF.

I have attached an example notebook that I have been using:
BNAF_error_github.zip

Settings for reproduction:
reference_set_size = 10000 #held out set
training_epochs = [2000]
training_sizes = [10000]
density_estimator = "bnaf"
gen_size = 300

Windows OS

python==3.7.6
torch==1.13.1
domias==0.0.5
geomloss==0.2.6
pykeops==2.1.2

Any help resolving this issues would be greatly appreciated. I am hoping to generate the results for my MSc thesis before mid-August.

Thanks,
Jordan

[holarissun]

Hi Jordan,

Sorry for the problem. We only verified running our code on the Linux OS when published.

Have you tried installing the KeOps package before installing the geomloss package?

Also, please refer to the BNAF repo for more information on that density estimator.

FYI, those two issues are relevant:

• name 'generic_logsumexp' is not defined jeanfeydy/geomloss#52
• 'generic_logsumexp' not defined when running code sample jeanfeydy/geomloss#17

I hope that helps you.

Best,
Hao