Fvector

egs/mini_librispeech/s5/cmd.sh

@@ -10,6 +10,7 @@
 # conf/queue.conf in http://kaldi-asr.org/doc/queue.html for more information,
 # or search for the string 'default_config' in utils/queue.pl or utils/slurm.pl.
 
-export train_cmd="queue.pl --mem 2G"
-export decode_cmd="queue.pl --mem 4G"
-export mkgraph_cmd="queue.pl --mem 8G"
+export train_cmd="queue.pl --mem 2G --config conf/queue_no_k20.conf --allow-k10-k20 true"
+export decode_cmd="queue.pl --mem 4G --config conf/queue_no_k20.conf --allow-k10-k20 true"
+export mkgraph_cmd="queue.pl --mem 8G --config conf/queue_no_k20.conf --allow-k10-k20 true"
+export cuda_cmd="queue.pl --gpu 1 --config conf/queue_no_k20.conf --allow-k10-k20 true"

egs/mini_librispeech/s5/conf/queue_no_k20.conf

@@ -0,0 +1,16 @@
+# Default configuration
+command qsub -v PATH -cwd -S /bin/bash -j y -l arch=*64* -l 'hostname=!a08*&!a09*&!a10*&!c04*&!b18*&!b19*&!b20*'
+option mem=* -l mem_free=$0,ram_free=$0
+option mem=0          # Do not add anything to qsub_opts
+option num_threads=* -pe smp $0
+option num_threads=1  # Do not add anything to qsub_opts
+option max_jobs_run=* -tc $0
+default gpu=0
+option gpu=0 -q all.q
+option gpu=* -l gpu=$0 -q g.q
+default allow_k20=true
+option allow_k20=true
+option allow_k20=false -l 'hostname=!g01*&!g02*&!b06*'
+default allow_k10_k20=true
+option allow_k10_k20=true
+option allow_k10_k20=false -l 'hostname=!b0*&!b10*&!g01*&!g02'

egs/mini_librispeech/s5/local/fvector/add_output_node.py

@@ -0,0 +1,68 @@
+#!/usr/bin/env/python
+
+from __future__ import print_function
+import argparse
+import logging
+import os
+import pprint
+import sys
+import shutil
+import traceback
+
+def get_args():
+    parser = argparse.ArgumentParser(description="Add the S and b output node "
+            "which is used in plda object function.",
+            epilog="Called by local/fvector/run_fvector.sh")
+    parser.add_argument("--input-dim", type=int, required=True,
+            help="The input dimension of fvector network.")
+    parser.add_argument("--output-dim", type=int, required=True,
+            help="The output dimension of fvector network which is used to "
+            "compute the dimension of S matrix.")
+    parser.add_argument("--s-scale", type=float, default=0.2,
+            help="Scaling factor on the output 's' (s is a symmetric matrix "
+            "used for scoring).")
+    parser.add_argument("--b-scale", type=float, default=0.2,
+            help="Scaling factor on output 'b' (b is a scalar offset used in scoring).")
+    parser.add_argument("--config-file", type=str, required=True,
+            help="The file is needed to be modified. It's always is configs/final.config")
+
+    print(' '.join(sys.argv), file=sys.stderr)
+    print(sys.argv, file=sys.stderr)
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = get_args()
+
+    f = open(args.config_file, "a")
+    # The s output
+    s_dim = (args.output_dim) * (args.output_dim+1) / 2
+
+    print('component name=x-s type=ConstantFunctionComponent input-dim={0} output-dim={1} '
+          'output-mean=0 output-stddev=0 '.format(
+              args.input_dim, s_dim), file=f)
+    print('component-node name=x-s component=x-s input=IfDefined(input)',           
+          file=f)                                                                   
+    print('component name=x-s-scale type=FixedScaleComponent dim={0} scale={1}'.format(
+                s_dim, args.s_scale), file=f);                                      
+    print('component-node name=x-s-scale component=x-s-scale input=x-s',            
+          file=f)                                                                   
+    print('output-node name=s input=x-s-scale', file=f)                             
+
+    # now the 'b' output, which is just a scalar.                                   
+    b_dim = 1                                                                       
+    print('component name=x-b type=ConstantFunctionComponent input-dim={0} output-dim=1 '
+          'output-mean=0 output-stddev=0 '.format(args.input_dim), file=f)           
+    print('component-node name=x-b component=x-b input=IfDefined(input)', file=f)   
+    print('component name=x-b-scale type=FixedScaleComponent dim=1 scale={0}'.format(
+            args.b_scale), file=f);                                                 
+    print('component-node name=x-b-scale component=x-b-scale input=x-b',            
+          file=f)                                                                   
+    print('output-node name=b input=x-b-scale', file=f)                             
+    f.close()                                                      
+
+
+
+if __name__ == "__main__":
+    main()

egs/mini_librispeech/s5/local/fvector/generate_sin_cos_matrix.py

@@ -0,0 +1,52 @@
+#!/usr/bin/env/python
+
+from __future__ import print_function
+import argparse
+import logging
+import os
+import pprint
+import shutil
+import sys
+import traceback
+
+sys.path.insert(0, 'steps')
+import libs.common as common_lib
+
+def get_args():
+    parser = argparse.ArgumentParser(description="Generate sine_transform.mat "
+            "and cosine_transform.mat for frequency domain raw waveform setup.",
+            epilog="Called by local/fvector/run_fvector.sh")
+    parser.add_argument("--feat-dim", type=int, required=True,
+            help="The dimension of input.")
+    parser.add_argument("--add-bias", type=str,
+            help="If true, add a column for fft matrix.",
+            default=True, choices=["True","False"])
+    parser.add_argument("--half-range", type=str,
+            help="If true, generate half fft matrix.",
+            default=True, choices=["True","False"])
+    parser.add_argument("--dir", type=str, required=True,
+            help="The output directory.")
+
+    print(' '.join(sys.argv), file=sys.stderr)
+    print(sys.argv, file=sys.stderr)
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = get_args()
+
+    feat_dim = args.feat_dim
+    num_fft_bins = (2**(args.feat_dim-1).bit_length())
+    add_bias = args.add_bias
+    half_range = args.half_range
+
+    common_lib.write_sin_cos_transform_matrix(feat_dim, num_fft_bins,
+            "{0}/configs/cos_transform.mat".format(args.dir),                   
+            compute_cosine=True, add_bias=add_bias, half_range=half_range)
+    common_lib.write_sin_cos_transform_matrix(feat_dim, num_fft_bins,
+            "{0}/configs/sin_transform.mat".format(args.dir),
+            compute_cosine=False, add_bias=add_bias, half_range=half_range)
+
+if __name__ == "__main__":
+    main()

egs/mini_librispeech/s5/local/fvector/run_fvector.sh

@@ -0,0 +1,78 @@
+#!/bin/bash
+
+. ./cmd.sh
+set -e
+
+stage=3
+train_stage=-10
+data=data/train_clean_5
+noise_data=data/noise
+egs_dir=exp/fvector/egs
+fvector_dir=exp/fvector
+use_gpu=true
+
+. ./path.sh
+. ./cmd.sh
+. ./utils/parse_options.sh
+
+if [ $stage -le 3 ]; then
+  #dump egs                                         
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $egs_dir/storage ]; then  
+    utils/create_split_dir.pl \
+    /export/b{11,12,13}/$USER/kaldi-data/egs/minilibrispeech-$(date +'%m_%d_%H_%M')/s5/$egs_dir/storage $egs_dir/storage
+  fi
+
+  steps/nnet3/fvector/get_egs.sh --cmd "$train_cmd" \
+    --nj 8 \
+    --stage 0 \
+    --egs-per-iter 12500 \
+    --egs-per-iter-diagnostic 10000 \
+    --num-diagnostic-percent 5 \
+    --frame-length 25 \
+    --left-padding 1 \
+    --right-padding 1 \
+    "$data" "$noise_data" "$egs_dir"
+fi
+
+if [ $stage -le 4 ]; then
+  #prepare configs
+  echo "$0: creating neural net configs using the xconfig parser";
+  #options
+  input_dim=400
+  num_filters=100
+
+  mkdir -p $fvector_dir/configs
+
+  cat <<EOF > $fvector_dir/configs/network.xconfig
+  input dim=$input_dim name=input
+  # Each eg contains 8 frames, do Frequency-domain feature learning, and then
+  # use TDNN model split it into one vector
+  preprocess-fft-abs-lognorm-affine-log-layer name=raw0 cos-transform-file=$fvector_dir/configs/cos_transform.mat sin-transform-file=$fvector_dir/configs/sin_transform.mat num-filters=$num_filters half-fft-range=true
+  conv-relu-batchnorm-layer name=cnn1 height-in=$num_filters height-out=$[$num_filters/2] time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=64 height-subsample-out=2 learning-rate-factor=0.34 max-change=0.25
+
+  relu-batchnorm-layer name=tdnn0 input=cnn1 dim=625
+  relu-batchnorm-layer name=tdnn1 input=Append(0,1,2) dim=625
+  relu-batchnorm-layer name=tdnn2 input=Append(0,1,2) dim=625
+  relu-batchnorm-layer name=tdnn3 input=Append(0,1,2) dim=625
+  relu-batchnorm-layer name=tdnn4 input=Append(0,1) dim=625
+  output-layer name=output input=tdnn4 dim=200 include-log-softmax=False param-stddev=0.04 bias-stddev=1.0
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $fvector_dir/configs/network.xconfig --config-dir $fvector_dir/configs/
+  # Modify the final.config and generate sin.mat/cos.mat manually
+  python local/fvector/add_output_node.py --input-dim 400 --output-dim 200 --config-file $fvector_dir/configs/final.config
+  python local/fvector/generate_sin_cos_matrix.py \
+    --feat-dim 400 --dir $fvector_dir
+fi
+
+if [ $stage -le 5 ]; then
+  #training
+  steps/nnet3/xvector/train.sh --cmd "$train_cmd" \
+    --initial-effective-lrate 0.002 \
+    --final-effective-lrate 0.0002 \
+    --max-param-change 0.2 \
+    --minibatch-size 16 \
+    --num-epochs 8 --use-gpu $use_gpu --stage $train_stage \
+    --num-jobs-initial 1 --num-jobs-final 5 \
+    --egs-dir $egs_dir \
+    $fvector_dir
+fi

egs/mini_librispeech/s5/local/fvector/run_fvector_separate.sh

@@ -0,0 +1,79 @@
+#!/bin/bash
+
+. ./cmd.sh
+set -e
+
+stage=5
+train_stage=-10
+data=data/train_clean_5
+noise_data=data/noise
+egs_dir=exp/fvector/egs
+fvector_dir=exp/fvector
+use_gpu=true
+
+. ./path.sh
+. ./cmd.sh
+. ./utils/parse_options.sh
+
+if [ $stage -le 3 ]; then
+  #dump egs                                         
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $egs_dir/storage ]; then  
+    utils/create_split_dir.pl \
+    /export/b{11,12,13}/$USER/kaldi-data/egs/minilibrispeech-$(date +'%m_%d_%H_%M')/s5/$egs_dir/storage $egs_dir/storage
+  fi
+
+  steps/nnet3/fvector/get_egs_separate.sh --cmd "$train_cmd" \
+    --nj 8 \
+    --stage 3 \
+    --egs-per-iter 100000 \
+    --egs-per-iter-diagnostic 10000 \
+    --num-diagnostic-percent 5 \
+    "$data" "$noise_data" "$egs_dir"
+fi
+
+if [ $stage -le 4 ]; then
+  #prepare configs
+  echo "$0: creating neural net configs using the xconfig parser";
+  #options
+  input_dim=400
+  num_filters=200
+
+  mkdir -p $fvector_dir/configs
+
+  cat <<EOF > $fvector_dir/configs/network.xconfig
+  input dim=$input_dim name=input
+  # Each eg contains 8 frames, do Frequency-domain feature learning, and then
+  # use TDNN model split it into one vector
+  preprocess-fft-abs-lognorm-affine-log-layer name=raw0 cos-transform-file=$fvector_dir/configs/cos_transform.mat sin-transform-file=$fvector_dir/configs/sin_transform.mat num-filters=$num_filters half-fft-range=true
+  conv-relu-batchnorm-layer name=cnn1 height-in=$num_filters height-out=$[$num_filters/2] time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=64 height-subsample-out=2 learning-rate-factor=0.34 max-change=0.25
+
+  relu-batchnorm-layer name=tdnn0 input=cnn1 dim=625
+  relu-batchnorm-layer name=tdnn1 input=Append(0,1,2) dim=625
+  relu-batchnorm-layer name=tdnn2 input=Append(0,1,2) dim=625
+  relu-batchnorm-layer name=tdnn3 input=Append(0,1,2) dim=625
+  relu-batchnorm-layer name=tdnn4 input=Append(0,1) dim=625
+  output-layer name=output input=tdnn4 dim=200 include-log-softmax=False param-stddev=0.04 bias-stddev=1.0
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $fvector_dir/configs/network.xconfig --config-dir $fvector_dir/configs/
+  # Modify the final.config and generate sin.mat/cos.mat manually
+  python local/fvector/add_output_node.py --input-dim 400 --output-dim 200 --config-file $fvector_dir/configs/final.config
+  python local/fvector/generate_sin_cos_matrix.py \
+    --feat-dim 400 --dir $fvector_dir
+fi
+
+if [ $stage -le 5 ]; then
+  #training
+  steps/nnet3/xvector/train_separate.sh --cmd "$train_cmd" \
+    --initial-effective-lrate 0.002 \
+    --final-effective-lrate 0.0002 \
+    --max-param-change 0.2 \
+    --minibatch-size 16 \
+    --left-padding 1 \
+    --right-padding 1 \
+    --max-snr 20 \
+    --min-snr 10 \
+    --num-epochs 8 --use-gpu $use_gpu --stage $train_stage \
+    --num-jobs-initial 1 --num-jobs-final 3 \
+    --egs-dir $egs_dir \
+    $fvector_dir
+fi