!12403 [MD] dataset error when train twice with different models

From: @liyong126
Reviewed-by: @heleiwang,@pandoublefeng
Signed-off-by: @pandoublefeng

mindspore/train/model.py

@@ -157,6 +157,12 @@ class Model:
             if arg not in ['loss_scale_manager', 'keep_batchnorm_fp32']:
                 raise ValueError(f"Unsupported arg '{arg}'")
 
+    def _check_reuse_dataset(self, dataset):
+        if not hasattr(dataset, '__model_hash__'):
+            dataset.__model_hash__ = hash(self)
+        if hasattr(dataset, '__model_hash__') and dataset.__model_hash__ != hash(self):
+            raise RuntimeError('The Dataset cannot be bound to different models, please create a new dataset.')
+
     def _build_train_network(self):
         """Build train network"""
         network = self._network
@@ -388,6 +394,7 @@ class Model:
                                "So the training process will be performed with dataset not sink.")
                 self._train_process(epoch, train_dataset, list_callback, cb_params)
             else:
+                self._check_reuse_dataset(train_dataset)
                 self._train_dataset_sink_process(epoch, train_dataset, list_callback, cb_params, sink_size)
 
     @staticmethod