LmDataset get_total_num_seqs (#1442)

Should fix #1441.

Also:

CachedDataset2 num_seqs 0 if not initialized.
CachedDataset2 is_less_than_num_seqs same.

Dataset have_seqs more correct.

returnn/datasets/basic.py

@@ -956,7 +956,9 @@ def have_seqs(self) -> bool:
             return total_num_seqs > 0
         except NotImplementedError:
             pass
-        return self.is_less_than_num_seqs(0)
+        if self.epoch is not None:
+            return self.is_less_than_num_seqs(0)
+        raise NotImplementedError(f"{self} have_seqs() is not implemented (and neither get_total_num_seqs())")
 
     def len_info(self):
         """

returnn/datasets/cached2.py

@@ -98,6 +98,8 @@ def num_seqs(self):
         """
         if self._num_seqs is not None:
             return self._num_seqs
+        if self.epoch is None:
+            return 0
         raise NotImplementedError
 
     def _load_seqs(self, start, end):
@@ -128,6 +130,8 @@ def is_less_than_num_seqs(self, n):
         """
         if n < self.expected_load_seq_start:
             return True
+        if self.epoch is None:
+            return False
         # noinspection PyBroadException
         try:
             return super(CachedDataset2, self).is_less_than_num_seqs(n)

returnn/datasets/lm.py

@@ -303,6 +303,10 @@ def supports_seq_order_sorting(self) -> bool:
         """supports sorting"""
         return True
 
+    def get_total_num_seqs(self) -> int:
+        """total num seqs"""
+        return len(self.orths)
+
     def _reduce_log_skipped_seqs(self):
         if isinstance(self.log_skipped_seqs, bool):
             return