Merge pull request #4463 from Topdu/dygraph

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: