x.shape bug when is -1

deepspeech/modules/embedding.py

@@ -67,10 +67,10 @@ class PositionalEncoding(nn.Layer):
             paddle.Tensor: Encoded tensor. Its shape is (batch, time, ...)
             paddle.Tensor: for compatibility to RelPositionalEncoding, (batch=1, time, ...)
         """
-        T = x.shape[1]
         assert offset + x.shape[1] < self.max_len
-        #TODO(Hui Zhang): using T = x.size(1), __getitem__ not support Tensor
-        pos_emb = self.pe[:, offset:offset + T]
+        # when frozen graph, the x.shape[1] is -1, offset=0
+        # result pos_emb is [1, 4999, D] not [1, 5000, D]
+        pos_emb = self.pe[:, offset:offset + x.shape[1]]
         x = x * self.xscale + pos_emb
         return self.dropout(x), self.dropout(pos_emb)
 
@@ -116,6 +116,7 @@ class RelPositionalEncoding(PositionalEncoding):
         """
         assert offset + x.shape[1] < self.max_len
         x = x * self.xscale
-        #TODO(Hui Zhang): using x.size(1), __getitem__ not support Tensor
+        # when frozen graph, the x.shape[1] is -1, offset=0
+        # result pos_emb is [1, 4999, D] not [1, 5000, D]
         pos_emb = self.pe[:, offset:offset + x.shape[1]]
         return self.dropout(x), self.dropout(pos_emb)