add valgrind

speechx/TODO.md

+# TODO
+
+* DecibelNormalizer: there is a little bit difference between offline and online db norm. The computation of online db norm read feature chunk by chunk, which causes the feature size is different with offline db norm. In normalizer.cc:73, the samples.size() is different, which causes the difference of result.

speechx/examples/decoder/run.sh

@@ -28,5 +28,5 @@ cmvn=./cmvn.ark
 linear_spectrogram_main --wav_rspecifier=scp:$model_dir/wav.scp --feature_wspecifier=ark:$feat_wspecifier --cmvn_write_path=$cmvn
 
 # 4. run decoder
-../../build/examples/decoder/offline_decoder_main --feature_respecifier=ark:$feat_wspecifier --model_path=$model_dir/avg_1.jit.pdmodel --param_path=$model_dir/avg_1.jit.pdparams --dict_file=$model_dir/vocab.txt --lm_path=$model_dir/avg_1.jit.klm
+offline_decoder_main --feature_respecifier=ark:$feat_wspecifier --model_path=$model_dir/avg_1.jit.pdmodel --param_path=$model_dir/avg_1.jit.pdparams --dict_file=$model_dir/vocab.txt --lm_path=$model_dir/avg_1.jit.klm
 

speechx/examples/decoder/valgrind.sh

+#!/bin/bash
+
+# this script is for memory check, so please run ./run.sh first.
+
+set +x
+set -e
+
+if [ ! -d ../../tools/valgrind/install ]; then
+  echo "please install valgrind in the speechx tools dir.\n" 
+  exit 1
+fi
+
+. ../path.sh
+
+model_dir=../paddle_asr_model
+feat_wspecifier=./feats.ark
+cmvn=./cmvn.ark
+
+valgrind --tool=memcheck --track-origins=yes --leak-check=full --show-leak-kinds=all offline_decoder_main --feature_respecifier=ark:$feat_wspecifier --model_path=$model_dir/avg_1.jit.pdmodel --param_path=$model_dir/avg_1.jit.pdparams --dict_file=$model_dir/vocab.txt --lm_path=$model_dir/avg_1.jit.klm
+

speechx/examples/feat/run.sh

@@ -25,4 +25,4 @@ feat_wspecifier=./feats.ark
 cmvn=./cmvn.ark
 
 # 3. run feat
-linear_spectrogram_main --wav_rspecifier=scp:$model_dir/wav.scp --feature_wspecifier=ark,t:$feat_wspecifier --cmvn_write_path=$cmvn
+../../fc_patch/valgrind-build/bin/valgrind --tool=memcheck --track-origins=yes --leak-check=full --show-leak-kinds=all linear_spectrogram_main --wav_rspecifier=scp:$model_dir/wav.scp --feature_wspecifier=ark,t:$feat_wspecifier --cmvn_write_path=$cmvn

speechx/examples/feat/valgrind.sh

+#!/bin/bash
+
+# this script is for memory check, so please run ./run.sh first.
+
+set +x
+set -e
+
+if [ ! -d ../../tools/valgrind/install ]; then
+  echo "please install valgrind in the speechx tools dir.\n" 
+  exit 1
+fi
+
+. ../path.sh
+
+model_dir=../paddle_asr_model
+feat_wspecifier=./feats.ark
+cmvn=./cmvn.ark
+
+valgrind --tool=memcheck --track-origins=yes --leak-check=full --show-leak-kinds=all linear_spectrogram_main --wav_rspecifier=scp:$model_dir/wav.scp --feature_wspecifier=ark,t:$feat_wspecifier --cmvn_write_path=$cmvn
+

speechx/examples/nnet/run.sh

@@ -22,7 +22,6 @@ fi
 
 model_dir=../paddle_asr_model
 
-
 # 4. run decoder
 pp-model-test --model_path=$model_dir/avg_1.jit.pdmodel --param_path=$model_dir/avg_1.jit.pdparams
 

speechx/examples/nnet/valgrind.sh

+#!/bin/bash
+
+# this script is for memory check, so please run ./run.sh first.
+
+set +x
+set -e
+
+if [ ! -d ../../tools/valgrind/install ]; then
+  echo "please install valgrind in the speechx tools dir.\n" 
+  exit 1
+fi
+
+. ../path.sh
+
+model_dir=../paddle_asr_model
+
+valgrind --tool=memcheck --track-origins=yes --leak-check=full --show-leak-kinds=all pp-model-test --model_path=$model_dir/avg_1.jit.pdmodel --param_path=$model_dir/avg_1.jit.pdparams
+

speechx/examples/path.sh

@@ -4,8 +4,11 @@ SPEECHX_ROOT=$PWD/../..
 SPEECHX_EXAMPLES=$SPEECHX_ROOT/build/examples
 SPEECHX_BIN=$SPEECHX_EXAMPLES/nnet:$SPEECHX_EXAMPLES/decoder:$SPEECHX_EXAMPLES/feat
 
+SPEECHX_TOOLS=$SPEECHX_ROOT/tools
+TOOLS_BIN=$SPEECHX_TOOLS/valgrind/install/bin
+
 [ -d $SPEECHX_EXAMPLES ] || { echo "Error: 'build/examples' directory not found. please ensure that the project build successfully"; }
 
 export LC_AL=C
 
-export PATH=$PATH:$SPEECHX_BIN
+export PATH=$PATH:$SPEECHX_BIN:$TOOLS_BIN

speechx/speechx/decoder/ctc_beam_search_decoder.cc

@@ -206,9 +206,9 @@ int32 CTCBeamSearch::SearchOneChar(
     size_t beam_size = opts_.beam_size;
     const auto& c = log_prob_idx.first;
     const auto& log_prob_c = log_prob_idx.second;
-    size_t prefixes__len = std::min(prefixes_.size(), beam_size);
+    size_t prefixes_len = std::min(prefixes_.size(), beam_size);
 
-    for (size_t i = 0; i < prefixes__len; ++i) {
+    for (size_t i = 0; i < prefixes_len; ++i) {
         auto prefix = prefixes_[i];
         if (full_beam && log_prob_c + prefix->score < min_cutoff) {
             break;
@@ -311,4 +311,4 @@ void CTCBeamSearch::LMRescore() {
     }
 }
 
-}  // namespace ppspeech
+}  // namespace ppspeech
\ No newline at end of file

speechx/tools/setup_valgrind.sh

+#!/bin/bash
+
+VALGRIND_VERSION=3.18.1
+
+set -e
+
+tarball=valgrind-3.18.1.tar.bz2
+
+url=https://sourceware.org/pub/valgrind/valgrind-3.18.1.tar.bz2
+
+if [ -f $tarball ]; then
+  echo "use the $tarball have downloaded."
+else
+  wget -t3 --no-check-certificate $url
+fi
+
+tar xjfv $tarball
+
+mv valgrind-3.18.1 valgrind
+
+prefix=$PWD/valgrind/install
+cd ./valgrind/
+  ./configure --prefix=$prefix
+  make
+  make install
+cd -