fix timit scripts; reader filtype case;

deepspeech/io/reader.py

@@ -322,7 +322,7 @@ class LoadInputsAndTargets():
                 "Not supported: loader_type={}".format(filetype))
 
     def file_type(self, filepath):
-        suffix = filepath.split(":")[0].split('.')[-1]
+        suffix = filepath.split(":")[0].split('.')[-1].lower()
         if suffix == 'ark':
             return 'mat'
         elif suffix == 'scp':

deepspeech/utils/error_rate.py

@@ -14,6 +14,7 @@
 """This module provides functions to calculate error rate in different level.
 e.g. wer for word-level, cer for char-level.
 """
+import editdistance
 import numpy as np
 
 __all__ = ['word_errors', 'char_errors', 'wer', 'cer']
@@ -89,6 +90,7 @@ def word_errors(reference, hypothesis, ignore_case=False, delimiter=' '):
     hyp_words = list(filter(None, hypothesis.split(delimiter)))
 
     edit_distance = _levenshtein_distance(ref_words, hyp_words)
+    # edit_distance = editdistance.eval(ref_words, hyp_words)
     return float(edit_distance), len(ref_words)
 
 
@@ -119,6 +121,7 @@ def char_errors(reference, hypothesis, ignore_case=False, remove_space=False):
     hypothesis = join_char.join(list(filter(None, hypothesis.split(' '))))
 
     edit_distance = _levenshtein_distance(reference, hypothesis)
+    # edit_distance = editdistance.eval(reference, hypothesis)
     return float(edit_distance), len(reference)
 
 

deepspeech/utils/tensor_utils.py

@@ -93,20 +93,25 @@ def pad_sequence(sequences: List[paddle.Tensor],
     for i, tensor in enumerate(sequences):
         length = tensor.shape[0]
         # use index notation to prevent duplicate references to the tensor
+        logger.info(
+            f"length {length}, out_tensor {out_tensor.shape}, tensor {tensor.shape}"
+        )
         if batch_first:
             # TODO (Hui Zhang): set_value op not supprot `end==start`
+            # TODO (Hui Zhang): set_value op not support int16
+            # TODO (Hui Zhang): set_varbase 2 rank not support [0,0,...] 
             # out_tensor[i, :length, ...] = tensor
             if length != 0:
-                out_tensor[i, :length, ...] = tensor
+                out_tensor[i, :length] = tensor
             else:
-                out_tensor[i, length, ...] = tensor
+                out_tensor[i, length] = tensor
         else:
             # TODO (Hui Zhang): set_value op not supprot `end==start`
             # out_tensor[:length, i, ...] = tensor
             if length != 0:
-                out_tensor[:length, i, ...] = tensor
+                out_tensor[:length, i] = tensor
             else:
-                out_tensor[length, i, ...] = tensor
+                out_tensor[length, i] = tensor
 
     return out_tensor
 

examples/ted_en_zh/t0/local/data.sh

@@ -1,5 +1,7 @@
 #!/bin/bash
 
+set -e
+
 stage=-1
 stop_stage=100
 

examples/timit/s1/.gitignore

@@ -0,0 +1,3 @@
+data
+exp
+test.profile

examples/timit/s1/README.md

@@ -1,11 +1,9 @@
 # TIMIT
 
-
-
-
 ### Transformer
-| Model | Params | Config | Decode method          | PER    |
+| Model | Params | Config | Decode method | Loss |  PER |
 | --- | --- | --- | --- | --- |
-| transformer | 5.17M | conf/transformer.yaml | attention              | 0.5531 |
-| transformer | 5.17M | conf/transformer.yaml | ctc_greedy_search      | 0.3922 |
-| transformer | 5.17M | conf/transformer.yaml | ctc_prefix_beam_search | 0.3768 |
+| transformer | 5.17M | conf/transformer.yaml | attention              | 49.25688171386719 | 0.510742 |
+| transformer | 5.17M | conf/transformer.yaml | ctc_greedy_search      | 49.25688171386719 | 0.382398 |
+| transformer | 5.17M | conf/transformer.yaml | ctc_prefix_beam_search | 49.25688171386719 | 0.367429 |
+| transformer | 5.17M | conf/transformer.yaml | attention_rescore      | 49.25688171386719 | 0.357173 |  

examples/timit/s1/local/test.sh

@@ -1,10 +1,18 @@
 #!/bin/bash
 
+set -e
+
+stage=0
+stop_stage=50
+
+. ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
+
 if [ $# != 2 ];then
     echo "usage: ${0} config_path ckpt_path_prefix"
     exit -1
 fi
 
+
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
@@ -23,6 +31,7 @@ fi
 #    exit 1
 #fi
 
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     for type in attention ctc_greedy_search; do
         echo "decoding ${type}"
         if [ ${chunk_mode} == true ];then
@@ -44,8 +53,11 @@ for type in attention ctc_greedy_search; do
             exit 1
         fi
     done
+fi
+
 
-for type in ctc_prefix_beam_search attention_rescoring; do
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    for type in ctc_prefix_beam_search; do
         echo "decoding ${type}"
         batch_size=1
         python3 -u ${BIN_DIR}/test.py \
@@ -61,6 +73,25 @@ for type in ctc_prefix_beam_search attention_rescoring; do
             exit 1
         fi
     done
+fi
 
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    for type in attention_rescoring; do
+        echo "decoding ${type}"
+        batch_size=1
+        python3 -u ${BIN_DIR}/test.py \
+        --nproc ${ngpu}  \
+        --config ${config_path} \
+        --result_file ${ckpt_prefix}.${type}.rsl \
+        --checkpoint_path ${ckpt_prefix} \
+        --opts decoding.decoding_method ${type} \
+        --opts decoding.batch_size ${batch_size}
+    
+        if [ $? -ne 0 ]; then
+            echo "Failed in evaluation!"
+            exit 1
+        fi
+    done
+fi
 
 exit 0

examples/timit/s1/run.sh

@@ -1,13 +1,15 @@
 #!/bin/bash
 set -e
-source path.sh
+
+. path.sh || exit 1;
 
 stage=0
 stop_stage=50
 conf_path=conf/transformer.yaml
 avg_num=10
-TIMIT_path= #path of TIMIT (Required, e.g. /export/corpora5/LDC/LDC93S1/timit/TIMIT)
-source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
+TIMIT_path=/workspace/zhanghui/dataset/data/lisa/data/timit/raw/TIMIT
+
+. ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 
 avg_ckpt=avg_${avg_num}
 ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}')

requirements.txt

@@ -19,3 +19,4 @@ tqdm
 typeguard
 visualdl==2.2.0
 yacs
+editdistance