fix egs bugs (#552)

* fix egs

* fix log

deepspeech/exps/deepspeech2/model.py

@@ -63,8 +63,6 @@ class DeepSpeech2Trainer(Trainer):
 
         losses_np = {
             'train_loss': float(loss),
-            'train_loss_div_batchsize':
-            float(loss) / self.config.data.batch_size
         }
         msg = "Train: Rank: {}, ".format(dist.get_rank())
         msg += "epoch: {}, ".format(self.epoch)
@@ -90,8 +88,6 @@ class DeepSpeech2Trainer(Trainer):
             loss = self.model(*batch)
 
             valid_losses['val_loss'].append(float(loss))
-            valid_losses['val_loss_div_batchsize'].append(
-                float(loss) / self.config.data.batch_size)
 
         # write visual log
         valid_losses = {k: np.mean(v) for k, v in valid_losses.items()}

deepspeech/modules/loss.py

@@ -62,14 +62,15 @@ class CTCLoss(nn.Layer):
         """Compute CTC loss.
 
         Args:
-            logits ([paddle.Tensor]): [description]
-            ys_pad ([paddle.Tensor]): [description]
-            hlens ([paddle.Tensor]): [description]
-            ys_lens ([paddle.Tensor]): [description]
+            logits ([paddle.Tensor]): [B, Tmax, D]
+            ys_pad ([paddle.Tensor]): [B, Tmax]
+            hlens ([paddle.Tensor]): [B]
+            ys_lens ([paddle.Tensor]): [B]
 
         Returns:
             [paddle.Tensor]: scalar. If reduction is 'none', then (N), where N = \text{batch size}.
         """
+        B = paddle.shape(logits)[0]
         # warp-ctc need logits, and do softmax on logits by itself
         # warp-ctc need activation with shape [T, B, V + 1]
         # logits: (B, L, D) -> (L, B, D)
@@ -78,5 +79,5 @@ class CTCLoss(nn.Layer):
 
         # wenet do batch-size average, deepspeech2 not do this
         # Batch-size average
-        # loss = loss / paddle.shape(logits)[1]
+        # loss = loss / B
         return loss

examples/aishell/.gitignore

@@ -2,3 +2,4 @@ data
 ckpt*
 demo_cache
 *.log
+log

examples/aishell/local/infer.sh

 #! /usr/bin/env bash
 
-if [[ $# != 1 ]];
+if [[ $# != 1 ]]; then
     echo "usage: $0 ckpt-path"
     exit -1
 fi

examples/aishell/run.sh

@@ -7,7 +7,7 @@ source path.sh
 bash ./local/data.sh
 
 # train model
-CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./local/train.sh
+CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./local/train.sh baseline
 
 # test model
 CUDA_VISIBLE_DEVICES=0 bash ./local/test.sh
@@ -16,4 +16,4 @@ CUDA_VISIBLE_DEVICES=0 bash ./local/test.sh
 CUDA_VISIBLE_DEVICES=0 bash ./local/infer.sh ckpt/checkpoints/step-3284
 
 # export model
-bash ./local/export.sh ckpt/checkpoints/step-3284 jit.model
\ No newline at end of file
+bash ./local/export.sh ckpt/checkpoints/step-3284 jit.model

examples/librispeech/local/infer.sh

 #! /usr/bin/env bash
 
-if [[ $# != 1 ]];
+if [[ $# != 1 ]];then
     echo "usage: $0 ckpt-path"
     exit -1
 fi

examples/librispeech/local/train.sh

@@ -11,7 +11,7 @@ python3 -u ${BIN_DIR}/train.py \
 --device 'gpu' \
 --nproc ${ngpu} \
 --config conf/deepspeech2.yaml \
---output ckpt
+--output ckpt-${1}
 
 if [ $? -ne 0 ]; then
     echo "Failed in training!"

examples/tiny/local/infer.sh

 #! /usr/bin/env bash
 
-if [[ $# != 1 ]];
+if [[ $# != 1 ]];then
     echo "usage: $0 ckpt-path"
     exit -1
 fi

examples/tiny/local/test.sh

@@ -6,7 +6,6 @@ if [ $? -ne 0 ]; then
     exit 1
 fi
 
-CUDA_VISIBLE_DEVICES=0 \
 python3 -u ${BIN_DIR}/test.py \
 --device 'gpu' \
 --nproc 1 \

examples/tiny/local/train.sh

@@ -2,7 +2,6 @@
 
 export FLAGS_sync_nccl_allreduce=0
 
-CUDA_VISIBLE_DEVICES=0 \
 python3 -u ${BIN_DIR}/train.py \
 --device 'gpu' \
 --nproc 1 \