使用Trainer中data_iterator属性参与循环使得可以通过callback修改数据

fastNLP/core/trainer.py

@@ -648,17 +648,16 @@ class Trainer(object):
         with inner_tqdm(total=self.n_steps, postfix='loss:{0:<6.5f}', leave=False, dynamic_ncols=True) as pbar:
             self.pbar = pbar
             avg_loss = 0
-            data_iterator = self.data_iterator
-            self.batch_per_epoch = data_iterator.num_batches
+            self.batch_per_epoch = self.data_iterator.num_batches
             for epoch in range(1, self.n_epochs + 1):
                 self.epoch = epoch
                 pbar.set_description_str(desc="Epoch {}/{}".format(epoch, self.n_epochs))
                 # early stopping
                 self.callback_manager.on_epoch_begin()
-                for batch_x, batch_y in data_iterator:
+                for batch_x, batch_y in self.data_iterator:
                     self.step += 1
                     _move_dict_value_to_device(batch_x, batch_y, device=self._model_device)
-                    indices = data_iterator.get_batch_indices()
+                    indices = self.data_iterator.get_batch_indices()
                     # negative sampling; replace unknown; re-weight batch_y
                     self.callback_manager.on_batch_begin(batch_x, batch_y, indices)
                     prediction = self._data_forward(self.model, batch_x)
@@ -692,7 +691,7 @@ class Trainer(object):
                     self.callback_manager.on_batch_end()
 
                     if ((self.validate_every > 0 and self.step % self.validate_every == 0) or
-                        (self.validate_every < 0 and self.step % len(data_iterator) == 0)) \
+                        (self.validate_every < 0 and self.step % len(self.data_iterator) == 0)) \
                             and self.dev_data is not None:
                         eval_res = self._do_validation(epoch=epoch, step=self.step)
                         eval_str = "Evaluation on dev at Epoch {}/{}. Step:{}/{}: ".format(epoch, self.n_epochs, self.step,