adapt paddle 3.2, remove some hack code to avoid conflict. (#4120)

paddlespeech/s2t/__init__.py

@@ -37,14 +37,6 @@ paddle.long = 'int64'
 paddle.uint16 = 'uint16'
 paddle.cdouble = 'complex128'
 
-if not hasattr(paddle, 'softmax'):
-    logger.debug("register user softmax to paddle, remove this when fixed!")
-    setattr(paddle, 'softmax', paddle.nn.functional.softmax)
-
-if not hasattr(paddle, 'log_softmax'):
-    logger.debug("register user log_softmax to paddle, remove this when fixed!")
-    setattr(paddle, 'log_softmax', paddle.nn.functional.log_softmax)
-
 if not hasattr(paddle, 'sigmoid'):
     logger.debug("register user sigmoid to paddle, remove this when fixed!")
     setattr(paddle, 'sigmoid', paddle.nn.functional.sigmoid)

paddlespeech/s2t/models/wav2vec2/modules/modeling_wav2vec2.py

@@ -891,7 +891,7 @@ class Wav2Vec2GumbelVectorQuantizer(nn.Layer):
                 hard=True).type_as(hidden_states)
 
             # compute perplexity
-            codevector_soft_dist = paddle.softmax(
+            codevector_soft_dist = paddle.nn.functional.softmax(
                 hidden_states.reshape((batch_size * sequence_length,
                                        self.num_groups, -1)).float(),
                 axis=-1)

paddlespeech/s2t/modules/attention.py

@@ -120,11 +120,11 @@ class MultiHeadedAttention(nn.Layer):
             # for last chunk, time2 might be larger than scores.size(-1)
             mask = mask[:, :, :, :scores.shape[-1]]
             scores = scores.masked_fill(mask, -float('inf'))
-            attn = paddle.softmax(
+            attn = paddle.nn.functional.softmax(
                 scores, axis=-1).masked_fill(mask,
                                              0.0)  # (batch, head, time1, time2)
         else:
-            attn = paddle.softmax(
+            attn = paddle.nn.functional.softmax(
                 scores, axis=-1)  # (batch, head, time1, time2)
 
         p_attn = self.dropout(attn)

paddlespeech/s2t/modules/decoder.py

@@ -189,7 +189,7 @@ class TransformerDecoder(BatchScorerInterface, nn.Layer):
         else:
             y = x[:, -1]
         if self.use_output_layer:
-            y = paddle.log_softmax(self.output_layer(y), axis=-1)
+            y = paddle.nn.functional.log_softmax(self.output_layer(y), axis=-1)
         return y, new_cache
 
     # beam search API (see ScorerInterface)