modified Makefile for pytorch, and changed document and run.sh in the…

… WSJ recipe accordingly.
elgeish · Dec 25, 2017 · 2e608c1 · 2e608c1
1 parent e72c979
commit 2e608c1
Show file tree

Hide file tree

Showing 12 changed files with 65 additions and 97 deletions.
diff --git a/LICENSE b/LICENSE
@@ -186,7 +186,7 @@
       same "printed page" as the copyright notice for easier
       identification within third-party archives.
 
-   Copyright [yyyy] [name of copyright owner]
+   Copyright 2017 Johns Hopkins University (Shinji Watanabe)
 
    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.

diff --git a/README.md b/README.md
@@ -3,7 +3,7 @@
 [![Build Status](https://travis-ci.org/espnet/espnet.svg?branch=master)](https://travis-ci.org/espnet/espnet)
 
 ESPnet is an end-to-end speech processing toolkit.
-ESPnet uses [chainer](https://chainer.org/) as a main deep learning engine, 
+ESPnet uses [chainer](https://chainer.org/) and [pytorch](http://pytorch.org/) as a main deep learning engine, 
 and also follows [Kaldi](http://kaldi-asr.org/) style data processing, feature extraction/format, and recipes to provide a complete setup for speech recognition and other speech processing experiments.
 
 ## Installation
@@ -31,16 +31,20 @@ The following directory is an example of performing ASR experiment with the VoxF
 ```sh
 $ cd egs/voxforge/asr1
 ```
-Once move to the directory, then, execute the following main script:
+Once move to the directory, then, execute the following main script with a chainer backend:
 ```sh
 $ ./run.sh
 ```
+or execute the following main script with a pytorch backend:
+```sh
+$ ./run.sh --backend pytorch --etype blstmp
+```
 With this main script, you can perform a full procedure of ASR experiments including
 - Data download
 - Data preparation (Kaldi style, see http://kaldi-asr.org/doc/data_prep.html)
 - Feature extraction (Kaldi style, see http://kaldi-asr.org/doc/feat.html)
 - Dictionary and JSON format data preparation
-- Training based on [chainer](https://chainer.org/).
+- Training based on [chainer](https://chainer.org/) or [pytorch](http://pytorch.org/).
 - Recognition and scoring
 
 ### Error due to matplotlib

diff --git a/egs/chime4/asr1_1ch/path.sh b/egs/chime4/asr1_1ch/path.sh
@@ -1,6 +1,5 @@
 MAIN_ROOT=$PWD/../../..
 KALDI_ROOT=$MAIN_ROOT/tools/kaldi
-JQ_ROOT=$MAIN_ROOT/tools/jq
 SPNET_ROOT=$MAIN_ROOT/src
 
 [ -f $KALDI_ROOT/tools/env.sh ] && . $KALDI_ROOT/tools/env.sh
@@ -9,8 +8,11 @@ export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/tools/sctk/bin
 . $KALDI_ROOT/tools/config/common_path.sh
 export LC_ALL=C
 
-export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$JQ_ROOT:$PATH
+export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PATH
 source $MAIN_ROOT/tools/venv/bin/activate
 export PYTHONPATH=$SPNET_ROOT/nets/:$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PYTHONPATH
 
 export OMP_NUM_THREADS=1
+if [[ $(hostname -f) == *.clsp.jhu.edu ]] ; then
+    export CUDA_VISIBLE_DEVICES=$(/usr/local/bin/free-gpu)
+fi
diff --git a/egs/csj/asr1/path.sh b/egs/csj/asr1/path.sh
@@ -1,6 +1,5 @@
 MAIN_ROOT=$PWD/../../..
 KALDI_ROOT=$MAIN_ROOT/tools/kaldi
-JQ_ROOT=$MAIN_ROOT/tools/jq
 SPNET_ROOT=$MAIN_ROOT/src
 
 [ -f $KALDI_ROOT/tools/env.sh ] && . $KALDI_ROOT/tools/env.sh
@@ -9,8 +8,11 @@ export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/tools/sctk/bin
 . $KALDI_ROOT/tools/config/common_path.sh
 export LC_ALL=C
 
-export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$JQ_ROOT:$MAIN_ROOT/tools/nkf/nkf-2.1.4/:$PATH
+export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$MAIN_ROOT/tools/nkf/nkf-2.1.4/:$PATH
 source $MAIN_ROOT/tools/venv/bin/activate
 export PYTHONPATH=$SPNET_ROOT/nets/:$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PYTHONPATH
 
 export OMP_NUM_THREADS=1
+if [[ $(hostname -f) == *.clsp.jhu.edu ]] ; then
+    export CUDA_VISIBLE_DEVICES=$(/usr/local/bin/free-gpu)
+fi
diff --git a/egs/hkust/asr1/path.sh b/egs/hkust/asr1/path.sh
@@ -1,6 +1,5 @@
 MAIN_ROOT=$PWD/../../..
 KALDI_ROOT=$MAIN_ROOT/tools/kaldi
-JQ_ROOT=$MAIN_ROOT/tools/jq
 SPNET_ROOT=$MAIN_ROOT/src
 
 [ -f $KALDI_ROOT/tools/env.sh ] && . $KALDI_ROOT/tools/env.sh
@@ -9,8 +8,11 @@ export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/tools/sctk/bin
 . $KALDI_ROOT/tools/config/common_path.sh
 export LC_ALL=C
 
-export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$JQ_ROOT:$PATH
+export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PATH
 source $MAIN_ROOT/tools/venv/bin/activate
 export PYTHONPATH=$SPNET_ROOT/nets/:$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PYTHONPATH
 
 export OMP_NUM_THREADS=1
+if [[ $(hostname -f) == *.clsp.jhu.edu ]] ; then
+    export CUDA_VISIBLE_DEVICES=$(/usr/local/bin/free-gpu)
+fi
diff --git a/egs/tedlium/asr1/path.sh b/egs/tedlium/asr1/path.sh
@@ -1,6 +1,5 @@
 MAIN_ROOT=$PWD/../../..
 KALDI_ROOT=$MAIN_ROOT/tools/kaldi
-JQ_ROOT=$MAIN_ROOT/tools/jq
 SPNET_ROOT=$MAIN_ROOT/src
 
 [ -f $KALDI_ROOT/tools/env.sh ] && . $KALDI_ROOT/tools/env.sh
@@ -9,8 +8,11 @@ export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/tools/sctk/bin
 . $KALDI_ROOT/tools/config/common_path.sh
 export LC_ALL=C
 
-export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$JQ_ROOT:$PATH
+export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PATH
 source $MAIN_ROOT/tools/venv/bin/activate
 export PYTHONPATH=$SPNET_ROOT/nets/:$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PYTHONPATH
 
 export OMP_NUM_THREADS=1
+if [[ $(hostname -f) == *.clsp.jhu.edu ]] ; then
+    export CUDA_VISIBLE_DEVICES=$(/usr/local/bin/free-gpu)
+fi
diff --git a/egs/voxforge/asr1/path.sh b/egs/voxforge/asr1/path.sh
@@ -1,6 +1,5 @@
 MAIN_ROOT=$PWD/../../..
 KALDI_ROOT=$MAIN_ROOT/tools/kaldi
-JQ_ROOT=$MAIN_ROOT/tools/jq
 SPNET_ROOT=$MAIN_ROOT/src
 
 [ -f $KALDI_ROOT/tools/env.sh ] && . $KALDI_ROOT/tools/env.sh
@@ -9,8 +8,11 @@ export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/tools/sctk/bin
 . $KALDI_ROOT/tools/config/common_path.sh
 export LC_ALL=C
 
-export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$JQ_ROOT:$PATH
+export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PATH
 source $MAIN_ROOT/tools/venv/bin/activate
 export PYTHONPATH=$SPNET_ROOT/nets/:$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PYTHONPATH
 
 export OMP_NUM_THREADS=1
+if [[ $(hostname -f) == *.clsp.jhu.edu ]] ; then
+    export CUDA_VISIBLE_DEVICES=$(/usr/local/bin/free-gpu)
+fi
diff --git a/egs/wsj/asr1/path.sh b/egs/wsj/asr1/path.sh
@@ -1,6 +1,5 @@
 MAIN_ROOT=$PWD/../../..
 KALDI_ROOT=$MAIN_ROOT/tools/kaldi
-JQ_ROOT=$MAIN_ROOT/tools/jq
 SPNET_ROOT=$MAIN_ROOT/src
 
 [ -f $KALDI_ROOT/tools/env.sh ] && . $KALDI_ROOT/tools/env.sh
@@ -9,8 +8,11 @@ export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/tools/sctk/bin
 . $KALDI_ROOT/tools/config/common_path.sh
 export LC_ALL=C
 
-export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$JQ_ROOT:$PATH
+export PATH=$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PATH
 source $MAIN_ROOT/tools/venv/bin/activate
 export PYTHONPATH=$SPNET_ROOT/nets/:$SPNET_ROOT/utils/:$SPNET_ROOT/bin/:$PYTHONPATH
 
 export OMP_NUM_THREADS=1
+if [[ $(hostname -f) == *.clsp.jhu.edu ]] ; then
+    export CUDA_VISIBLE_DEVICES=$(/usr/local/bin/free-gpu)
+fi
diff --git a/egs/wsj/asr1/run.sh b/egs/wsj/asr1/run.sh
@@ -150,13 +150,17 @@ else
 fi
 mkdir -p ${expdir}
 
+# switch backend
+if [[ ${backend} == chainer ]]; then
+    train_script=asr_train.py
+    decode_script=asr_recog.py
+else
+    train_script=asr_train_th.py
+    decode_script=asr_recog_th.py
+fi
+
 if [ ${stage} -le 3 ]; then
     echo "stage 3: Network Training"
-    if [[ ${backend} == chainer ]]; then
-        train_script=asr_train.py
-    else 
-        train_script=asr_train_th.py
-    fi
 
     ${cuda_cmd} ${expdir}/train.log \
 	    ${train_script} \
@@ -190,7 +194,7 @@ if [ ${stage} -le 3 ]; then
 fi
 
 if [ ${stage} -le 4 ]; then
-    echo "stage 4: Decoding data preparation"
+    echo "stage 4: Decoding"
     nj=32
 
     for rtask in ${recog_set}; do
@@ -201,42 +205,17 @@ if [ ${stage} -le 4 ]; then
 	    data=data/${rtask}
 	    split_data.sh --per-utt ${data} ${nj};
 
-	    # make json labels for recognition
-	    data2json.sh ${data} ${dict} > ${data}/data.json
-
-	) &
-    done
-    wait
-    echo "Finished"
-fi
-
-
-if [ ${stage} -le 5 ]; then
-    echo "stage 5: Decoding"
-    nj=32
-
-    for rtask in ${recog_set}; do
-	(
-	    decode_dir=decode_${rtask}_beam${beam_size}_e${recog_model}_p${penalty}_len${minlenratio}-${maxlenratio}
-
-	    # split data
-	    data=data/${rtask}
-	    sdata=${data}/split${nj}utt;
-
 	    # feature extraction
 	    feats="ark,s,cs:apply-cmvn --norm-vars=true data/${train_set}/cmvn.ark scp:${sdata}/JOB/feats.scp ark:- |"
 	    if ${do_delta}; then
-		    feats="$feats add-deltas ark:- ark:- |"
+		feats="$feats add-deltas ark:- ark:- |"
 	    fi
-	    #### use CPU for decoding
-	    gpu=0
 
-      if [[ ${backend} == chainer ]]; then
-          decode_script=asr_recog.py
-      else 
-          decode_script=asr_recog_th.py
-      fi
-      mkdir -p ${expdir}/${decode_dir}/log
+	    # make json labels for recognition
+	    data2json.sh ${data} ${dict} > ${data}/data.json
+
+	    #### use CPU for decoding
+	    gpu=-1
 
 	    ${decode_cmd} JOB=1:${nj} ${expdir}/${decode_dir}/log/decode.JOB.log \
 			${decode_script} \
@@ -250,24 +229,12 @@ if [ ${stage} -le 5 ]; then
 			--penalty ${penalty} \
 			--maxlenratio ${maxlenratio} \
 			--minlenratio ${minlenratio} \
-      --debugmode ${debugmode} \
-      --verbose ${verbose} \
-			&
+			--debugmode ${debugmode} \
+			--verbose ${verbose} &
 	    wait
-	) &
-    done
-    wait
-fi
-
-if [ ${stage} -le 6 ]; then
-    echo "stage 6: Evaluation"
-    nj=32
 
-    for rtask in ${recog_set}; do
-	(
-	    decode_dir=decode_${rtask}_beam${beam_size}_e${recog_model}_p${penalty}_len${minlenratio}-${maxlenratio}
 	    score_sclite.sh --nlsyms ${nlsyms} ${expdir}/${decode_dir} ${dict}
-
+			
 	) &
     done
     wait

diff --git a/src/bin/asr_train.py b/src/bin/asr_train.py
@@ -418,15 +418,8 @@ def main():
     for key in sorted(vars(args).keys()):
         logging.info('ARGS: ' + key + ': ' + str(vars(args)[key]))
 
-    if args.gpu == 'jhu':
-        # TODO make this one controlled at conf/gpu.conf or whatever
-        # this is JHU CLSP cluster setup
-        cmd = '/home/gkumar/scripts/free-gpu'
-        p = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
-        stdout_data, stderr_data = p.communicate()
-        gpu_id = int(stdout_data.rstrip())
-    else:
-        gpu_id = int(args.gpu)
+    # Set gpu
+    gpu_id = int(args.gpu)
     logging.info('gpu id: ' + str(gpu_id))
     if gpu_id >= 0:
         # Make a specified GPU current

diff --git a/src/bin/asr_train_th.py b/src/bin/asr_train_th.py
@@ -427,15 +427,8 @@ def main():
     for key in sorted(vars(args).keys()):
         logging.info('ARGS: ' + key + ': ' + str(vars(args)[key]))
 
-    if args.gpu == 'jhu':
-        # TODO make this one controlled at conf/gpu.conf or whatever
-        # this is JHU CLSP cluster setup
-        cmd = '/home/gkumar/scripts/free-gpu'
-        p = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
-        stdout_data, stderr_data = p.communicate()
-        gpu_id = int(stdout_data.rstrip())
-    else:
-        gpu_id = int(args.gpu)
+    # Set gpu
+    gpu_id = int(args.gpu)
     logging.info('gpu id: ' + str(gpu_id))
     if gpu_id >= 0:
         # Make a specified GPU current

diff --git a/tools/Makefile b/tools/Makefile
@@ -1,7 +1,7 @@
 .PHONY: all clean
 
 
-all: kaldi venv venv/bin/activate nkf kaldi-io-for-python.git pytorch warp-ctc
+all: kaldi venv venv/bin/activate nkf kaldi-io-for-python.git venv/lib/python2.7/site-packages/torch warp-ctc
 
 kaldi-io-for-python.git:
 	git clone https://github.com/vesis84/kaldi-io-for-python.git
@@ -29,17 +29,16 @@ nkf:
 	cd nkf; wget http://gigenet.dl.osdn.jp/nkf/64158/nkf-2.1.4.tar.gz
 	cd nkf; tar zxvf nkf-2.1.4.tar.gz; cd nkf-2.1.4; $(MAKE) prefix=.
 
-clean:
-	rm -fr kaldi_github kaldi kaldi_python venv nkf kaldi-io-for-python ../src/utils/kaldi_io_py.py
-	find -iname "*.pyc" -delete
-
-
-# optional deps for pytorch backends
-pytorch:
-	pip install http://download.pytorch.org/whl/cu80/torch-0.3.0.post4-cp27-cp27mu-linux_x86_64.whl
+venv/lib/python2.7/site-packages/torch: venv/bin/activate
+	. venv/bin/activate; pip install pip --upgrade; pip install http://download.pytorch.org/whl/cu80/torch-0.3.0.post4-cp27-cp27mu-linux_x86_64.whl
 
-warp-ctc:
+warp-ctc: venv/lib/python2.7/site-packages/torch
 	git clone https://github.com/SeanNaren/warp-ctc.git
 	cd warp-ctc && mkdir build && cd build && cmake .. && make -j4
-	pip install cffi
+	. venv/bin/activate; pip install cffi
 	cd warp-ctc/pytorch_binding && python setup.py install # maybe need to: apt-get install python-dev
+
+clean:
+	rm -fr kaldi_github kaldi kaldi_python venv nkf kaldi-io-for-python ../src/utils/kaldi_io_py.py warp-ctc
+	find -iname "*.pyc" -delete
+