fix ctc cuda memcopy error

examples/tiny/local/run_train.sh

@@ -17,6 +17,7 @@ export FLAGS_sync_nccl_allreduce=0
 #CUDA_VISIBLE_DEVICES=0,1,2,3 \
 CUDA_VISIBLE_DEVICES=1,2,3 \
 python3 -u ${MAIN_ROOT}/train.py \
+--device 'gpu' \
 --nproc 1 \
 --config conf/deepspeech2.yaml \
 --output ckpt

model_utils/model.py

@@ -149,8 +149,8 @@ class DeepSpeech2Trainer(Trainer):
         self.logger.info("Setup model/optimizer/criterion!")
 
     def compute_losses(self, inputs, outputs):
-        _, texts, logits_len, texts_len = inputs
+        _, texts, _, texts_len = inputs
-        logits = outputs
+        logits, logits_len = outputs
         loss = self.criterion(logits, texts, logits_len, texts_len)
         return loss
 

model_utils/network.py

@@ -517,11 +517,11 @@ class DeepSpeech2(nn.Layer):
         #ctcdecoder need probs, not log_probs
         probs = F.softmax(logits)
 
-        return logits, probs
+        return logits, probs, audio_len
 
     @paddle.no_grad()
     def infer(self, audio, audio_len):
-        _, probs = self.predict(audio, audio_len)
+        _, probs, audio_len = self.predict(audio, audio_len)
         return probs
 
     def forward(self, audio, text, audio_len, text_len):
@@ -531,8 +531,8 @@ class DeepSpeech2(nn.Layer):
         audio_len: shape [B]
         text_len: shape [B]
         """
-        logits, _ = self.predict(audio, audio_len)
+        logits, _, audio_len = self.predict(audio, audio_len)
-        return logits
+        return logits, audio_len
 
 
 class DeepSpeech2Loss(nn.Layer):