Error while using `nearest_neigh_prop`

[hasan-sayeed]

I'm trying to predict using trained model and facing the following error:

InvalidParameterError                     Traceback (most recent call last)
Cell In[16], line 13
     11 for i, val_df in enumerate(val_dfs):
     12     jd = jd + 1
---> 13     disc.predict(val_df, umap_random_state=42)
     14     dens_score.append(disc.dens_score_df.sample(frac=1))
     15     peak_score.append(disc.peak_score_df.sample(frac=1))

File c:\Users\hasan\miniconda3\envs\li_ml\lib\site-packages\mat_discover\mat_discover_.py:665, in Discover.predict(self, val_df, plotting, umap_random_state, pred_weight, proxy_weight, dummy_run, count_repeats, return_peak)
    663 # compound-wise scores (i.e. individual compounds)
    664 with self.Timer("nearest-neighbor-properties"):
--> 665     self.rad_neigh_avg_targ, self.k_neigh_avg_targ = nearest_neigh_props(
    666         self.dm, pred, n_neighbors=self.n_peak_neighbors
    667     )
    668     self.val_rad_neigh_avg = self.rad_neigh_avg_targ[val_ids]
    669     self.val_k_neigh_avg = self.k_neigh_avg_targ[val_ids]

File c:\Users\hasan\miniconda3\envs\li_ml\lib\site-packages\mat_discover\utils\nearest_neigh.py:48, in nearest_neigh_props(X, target, r_strength, radius, n_neighbors, metric, **NN_kwargs)
      6 def nearest_neigh_props(
      7     X,
      8     target,
   (...)
     13     **NN_kwargs,
     14 ):
...
     98     f"The {param_name!r} parameter of {caller_name} must be"
     99     f" {constraints_str}. Got {param_val!r} instead."
    100 )

InvalidParameterError: The 'radius' parameter of NearestNeighbors must be a float in the range [0, inf] or None. Got -0.28528690338134766 instead.

[sgbaird]

I can make a new kwarg that allows you to specify the radius directly. Maybe I need to give some more thought to the heuristic that's used to determine the default radius, too. https://mat-discover.readthedocs.io/en/latest/_modules/mat_discover/utils/nearest_neigh.html#nearest_neigh_props

I also wonder if this might have to do with you using so many neighbors. Might not be the issue, but maybe worth reducing the number of umap neighbors to something like 15 or 30 instead of 100's and see if you get the same error.

How many data points are you working with when you get the error?

[hasan-sayeed]

My val_df has ~48000 datapoints

[sgbaird]

Lmk if reducing the number of neighbors works