Protect against endless loop.

src/triacontagon/sampling.py

@@ -28,6 +28,11 @@ def fixed_unigram_candidate_sampler(
     if len(num_repeats.shape) != 1:
         raise ValueError('num_repeats must be 1D')
 
+    if torch.any(len(unigrams) - \
+        (true_classes >= 0).sum(dim=1) < \
+            num_repeats):
+        raise ValueError('Not enough classes to choose from')
+
     num_rows = true_classes.shape[0]
     print('true_classes.shape:', true_classes.shape)
     # unigrams = np.array(unigrams)