Merge pull request #4462 from Topdu/release/2.3

charry pick nrtr_postprocess and modify data type to adaptive

ppocr/losses/rec_nrtr_loss.py

@@ -22,7 +22,7 @@ class NRTRLoss(nn.Layer):
             log_prb = F.log_softmax(pred, axis=1)
             non_pad_mask = paddle.not_equal(
                 tgt, paddle.zeros(
-                    tgt.shape, dtype='int64'))
+                    tgt.shape, dtype=tgt.dtype))
             loss = -(one_hot * log_prb).sum(axis=1)
             loss = loss.masked_select(non_pad_mask).mean()
         else:

ppocr/postprocess/rec_postprocess.py

@@ -168,14 +168,6 @@ class NRTRLabelDecode(BaseRecLabelDecode):
                                              character_type, use_space_char)
 
     def __call__(self, preds, label=None, *args, **kwargs):
-        if preds.dtype == paddle.int64:
-            if isinstance(preds, paddle.Tensor):
-                preds = preds.numpy()
-            if preds[0][0]==2:
-                preds_idx = preds[:,1:]
-            else:
-                preds_idx = preds
-
         if len(preds) == 2:
             preds_id = preds[0]
             preds_prob = preds[1]