fix bug

README.md

@@ -80,6 +80,19 @@ yum安装出问题残留清理
           # 临时启用GCC-6.3（仅当前终端生效）
           scl enable devtoolset-6 bash 或 source /opt/rh/devtoolset-6/enable
 
+     # 若上述方法安装后 import horovod.torch时，出现 "/lib64/libstdc++.so.6: version `GLIBCXX_3.4.20' not found" 问题 < 方案 = 编译安装: https://blog.csdn.net/Yanci_/article/details/80016097>
+          # 下载源码包
+          wget http://mirrors.concertpass.com/gcc/releases/gcc-6.3.0/gcc-6.3.0.tar.gz
+          tar xzf gcc-6.3.0.tar.gz
+          cd gcc-6.3.0
+          ./contrib/download_prerequisites
+          mkdir gcc-build-6.3.0
+          cd gcc-build-6.3.0
+          ../configure --enable-checking=release --enable-languages=c,c++ --disable-multilib
+          make -j 4
+          make install
+
+          在 /root/.bashrc 中添加环境变量 export LD_LIBRARY_PATH=/usr/local/lib64:$LD_LIBRARY_PATH
 
       # 安装GPU支持版本（基于NCCL依赖）
       HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_GPU_BROADCAST=NCCL pip3 install horovod

makeFeatures.sh

@@ -3,7 +3,7 @@
 # Copyright xmuspeech (Author:Snowdar 2018-7-25)
 
 pitch=false
-pitch_config=conf/pitch.conf
+pitch_config=subtools/conf/pitch.conf
 cmvn=false
 use_gpu=false
 nj=20 #num-jobs
@@ -13,7 +13,7 @@ exp=exp/features
 
 if [[ $# != 3 ]];then
 echo "[exit] Num of parameters is not equal to 3"
-echo "usage:$0 [--pitch true|false] [--pitch-config "conf/pitch.conf"] [--nj 20|int] <data-dir> <feature-type> <feature-config>"
+echo "usage:$0 [--pitch false|true] [--pitch-config subtools/conf/pitch.conf] [--nj 20|int] <data-dir> <feature-type> <feature-config>"
 echo "[note] Base <feature-type> could be fbank/mfcc/plp/spectrogram and the option --pitch defaults false"
 exit 1
 fi

newCopyData.sh

@@ -2,6 +2,7 @@
 
 # Copyright xmuspeech (Author:Snowdar 2018-7-25)
 
+topdir=data
 force=false # for overwrite
 
 . subtools/parse_options.sh
@@ -18,15 +19,15 @@ srcs=$2
 for src in $srcs;do
 [ ! -d "$src" ] && tmp=$src && src=data/$src && [ ! -d "$src" ] && echo "[exit] No such dir $tmp or $src" && exit 1
 name=`basename $src`
-target=data/${pre}/$name
+target=$topdir/${pre}/$name
 
 [ "$src" == "$target" ] && echo "[Warning] data-dir $src is same to target-data-dir, so skip it" && continue
 [ -d $target ] && [ "$force" == "false" ] && echo "[exit] $target is exist, please delete it carefully by yourself" && exit 1
 
 rm -rf $target
 mkdir -p $target
 
-for x in wav.scp utt2spk spk2utt feats.scp vad.scp utt2num_frames utt2dur reco2dur text;do
+for x in wav.scp utt2spk spk2utt feats.scp vad.scp utt2num_frames utt2dur reco2dur text utt2gender spk2gender trials;do
 [ -f $src/$x ] && cp $src/$x $target/ && echo "[ $x ] copy done"
 done
 echo "[note] Your new datadir is $target"

pytorch/launcher/runSnowdarXvector-multi-GPU.py

@@ -195,7 +195,7 @@
 egs_dir="paper/egs/mfcc_23_pitch_voxceleb1_train_aug" + "_" + sample_type
 
 model_blueprint="subtools/pytorch/model/snowdar-xvector.py"
-model_dir="exp/standard_xv_multi_gpu_r1"
+model_dir="exp/standard_xv_multi_gpu_r2"
 ##--------------------------------------------------##
 ##
 #### Set seed

recipe/voxceleb/runVoxceleb.sh

@@ -3,11 +3,11 @@
 # Copyright xmuspeech (Author: Snowdar 2020-02-23)
 
 ### A record of my experiments about voxceleb1 and voxceleb2 and suggest to execute every script one by one.
-
+### Test is based on voxceleb1-O.
 
 ### Start 
 # Please prepare the data/voxceleb1_train, data/voxceleb2_train and data/voxceleb1_test by yourself (or official make*.pl).
-# voxceleb_train = voxceleb1_train + voxceleb2_train (without test both)
+# voxceleb_train = voxceleb1_train + voxceleb2_train (without test part both)
 subtools/kaldi/utils/combine_data.sh data/voxceleb_train data/voxceleb1_train data/voxceleb2_train
 
 # Get the copies of dataset which is labeled by a prefix like mfcc_23_pitch or fbank_40 etc.