fix some bug. (#2825)

paddlespeech/s2t/models/whisper/tokenizer.py

@@ -155,6 +155,10 @@ class Tokenizer:
                 if ids < len(self.tokenizer):
                     ids_list.append(ids)
             token_ids = ids_list
+        elif len(token_ids) == 1:
+            token_ids = token_ids[0]
+        else:
+            raise ValueError(f"token_ids {token_ids} load error.")
 
         return self.tokenizer.decode(token_ids, **kwargs)
 

paddlespeech/s2t/models/whisper/whipser.py

@@ -17,12 +17,11 @@ from typing import Union
 import numpy as np
 import paddle
 import paddle.nn.functional as F
+import paddlespeech.s2t.modules.align as paddlespeech_nn
 import soundfile
 import tqdm
 from paddle import nn
 from paddle.distribution import Categorical
-
-import paddlespeech.s2t.modules.align as paddlespeech_nn
 from paddlespeech.s2t.models.whisper import utils
 from paddlespeech.s2t.models.whisper.tokenizer import get_tokenizer
 from paddlespeech.s2t.models.whisper.tokenizer import LANGUAGES
@@ -771,8 +770,10 @@ class GreedyDecoder(TokenDecoder):
         if temperature == 0:
             next_tokens = paddle.argmax(logits, axis=-1)
         else:
-            next_tokens = Categorical(logits=logits / temperature).sample(
-                shape=logits.shape)
+            next_tokens = Categorical(logits=logits / temperature).sample([1])
+            next_tokens = paddle.reshape(next_tokens, [
+                next_tokens.shape[0] * next_tokens.shape[1],
+            ])
 
         logprobs = F.log_softmax(logits, axis=-1, dtype=paddle.float32)
         current_logprobs = logprobs[paddle.arange(logprobs.shape[0]),
@@ -1205,9 +1206,8 @@ class DecodingTask:
                 DecodingResult(
                     audio_features=features,
                     language=language,
-                    language_probs=probs)
-                for features, language, probs in zip(audio_features, languages,
-                                                     language_probs)
+                    language_probs=probs) for features, language, probs in
+                zip(audio_features, languages, language_probs)
             ]
 
         # repeat the audio & text tensors by the group size, for beam search or best-of-n sampling