batch sampling: cut sampled sampling-rate to (0,1]

ndfa/misc/tensors_data_class/batch_flattened_seq.py

@@ -69,16 +69,18 @@ def _collate_first_pass(cls, inputs: List['BatchFlattenedSequencesDataClassMixin
                 for sequences_field in sequences_fields:
                     tensors = getattr(inp, sequences_field.name)
                     nr_tensors = tensors.size(0) if isinstance(tensors, torch.Tensor) else len(tensors)
+                    if nr_tensors < 1:
+                        continue
                     if sequences_per_example_sampling.min_nr_items_to_sample_by_rate is not None and \
                             nr_tensors < sequences_per_example_sampling.min_nr_items_to_sample_by_rate:
                         continue
                     random_state = np.random.RandomState(random_seed_per_example[example_idx])
                     nr_sequences_to_sample_per_example = nr_tensors
                     if sequences_per_example_sampling.distribution_for_rate_to_sample_by is not None:
-                        sampling_rate = _sample_by_distribution_params(
+                        sampling_rate = max(1 / nr_tensors, min(1, _sample_by_distribution_params(
                             params=sequences_per_example_sampling.distribution_for_rate_to_sample_by,
-                            rng=random_state)
-                        nr_sequences_to_sample_per_example = round(sampling_rate * nr_tensors)
+                            rng=random_state)))
+                        nr_sequences_to_sample_per_example = max(1, min(nr_tensors, round(sampling_rate * nr_tensors)))
                     if sequences_per_example_sampling.max_nr_items is not None:
                         nr_sequences_to_sample_per_example = \
                             min(nr_sequences_to_sample_per_example, sequences_per_example_sampling.max_nr_items)