run_resnet_1b.sh
4.91 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
#!/bin/bash
# 1b is as 1a but using more epochs: 100 instead of 60.
# This helps a bit.
#exp/resnet1b_cifar10: num-iters=133 nj=1..2 num-params=1.3M dim=96->10 combine=-0.01->-0.01 loglike:train/valid[87,132,final]=(-0.13,-0.03,-0.01/-0.27,-0.21,-0.16) accuracy:train/valid[87,132,final]=(0.95,0.99,1.00/0.91,0.94,0.95)
#exp/resnet1b_cifar100: num-iters=133 nj=1..2 num-params=1.3M dim=96->100 combine=-0.22->-0.19 loglike:train/valid[87,132,final]=(-0.75,-0.27,-0.16/-1.22,-1.06,-0.89) accuracy:train/valid[87,132,final]=(0.78,0.93,0.96/0.67,0.72,0.76)
# local/nnet3/compare.sh exp/resnet1a_cifar10 exp/resnet1b_cifar10
# System resnet1a_cifar10 resnet1b_cifar10
# final test accuracy: 0.9481 0.9521
# final train accuracy: 0.9992 0.9998
# final test objf: -0.171369 -0.160283
# final train objf: -0.00980603 -0.00672504
# num-parameters: 1322730 1322730
# local/nnet3/compare.sh exp/resnet1a_cifar100 exp/resnet1b_cifar100
# System resnet1a_cifar100 resnet1b_cifar100
# final test accuracy: 0.7478 0.7597
# final train accuracy: 0.9446 0.9638
# final test objf: -0.899789 -0.889707
# final train objf: -0.22468 -0.163996
# num-parameters: 1345860 1345860
# Set -e here so that we catch if any executable fails immediately
set -euo pipefail
# training options
stage=0
train_stage=-10
dataset=cifar10
srand=0
reporting_email=
affix=1b
# End configuration section.
echo "$0 $@" # Print the command line for logging
. ./cmd.sh
. ./path.sh
. ./utils/parse_options.sh
if ! cuda-compiled; then
cat <<EOF && exit 1
This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
If you want to use GPUs (and have them), go to src/, and configure and make on a machine
where "nvcc" is installed.
EOF
fi
dir=exp/resnet${affix}_${dataset}
egs=exp/${dataset}_egs2
if [ ! -d $egs ]; then
echo "$0: expected directory $egs to exist. Run the get_egs.sh commands in the"
echo " run.sh before this script."
exit 1
fi
# check that the expected files are in the egs directory.
for f in $egs/egs.1.ark $egs/train_diagnostic.egs $egs/valid_diagnostic.egs $egs/combine.egs \
$egs/info/feat_dim $egs/info/left_context $egs/info/right_context \
$egs/info/output_dim; do
if [ ! -e $f ]; then
echo "$0: expected file $f to exist."
exit 1;
fi
done
mkdir -p $dir/log
if [ $stage -le 1 ]; then
mkdir -p $dir
echo "$0: creating neural net configs using the xconfig parser";
num_targets=$(cat $egs/info/output_dim)
# Note: we hardcode in the CNN config that we are dealing with 32x3x color
# images.
nf1=48
nf2=96
nf3=256
nb3=128
common="required-time-offsets=0 height-offsets=-1,0,1"
res_opts="bypass-source=batchnorm"
mkdir -p $dir/configs
cat <<EOF > $dir/configs/network.xconfig
input dim=96 name=input
conv-layer name=conv1 height-in=32 height-out=32 time-offsets=-1,0,1 required-time-offsets=0 height-offsets=-1,0,1 num-filters-out=$nf1
res-block name=res2 num-filters=$nf1 height=32 time-period=1 $res_opts
res-block name=res3 num-filters=$nf1 height=32 time-period=1 $res_opts
conv-layer name=conv4 height-in=32 height-out=16 height-subsample-out=2 time-offsets=-1,0,1 $common num-filters-out=$nf2
res-block name=res5 num-filters=$nf2 height=16 time-period=2 $res_opts
res-block name=res6 num-filters=$nf2 height=16 time-period=2 $res_opts
conv-layer name=conv7 height-in=16 height-out=8 height-subsample-out=2 time-offsets=-2,0,2 $common num-filters-out=$nf3
res-block name=res8 num-filters=$nf3 num-bottleneck-filters=$nb3 height=8 time-period=4 $res_opts
res-block name=res9 num-filters=$nf3 num-bottleneck-filters=$nb3 height=8 time-period=4 $res_opts
res-block name=res10 num-filters=$nf3 num-bottleneck-filters=$nb3 height=8 time-period=4 $res_opts
channel-average-layer name=channel-average input=Append(2,6,10,14,18,22,24,28) dim=$nf3
output-layer name=output learning-rate-factor=0.1 dim=$num_targets
EOF
steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
fi
if [ $stage -le 2 ]; then
steps/nnet3/train_raw_dnn.py --stage=$train_stage \
--cmd="$train_cmd" \
--image.augmentation-opts="--horizontal-flip-prob=0.5 --horizontal-shift=0.1 --vertical-shift=0.1 --num-channels=3" \
--trainer.srand=$srand \
--trainer.max-param-change=2.0 \
--trainer.num-epochs=100 \
--egs.frames-per-eg=1 \
--trainer.optimization.num-jobs-initial=1 \
--trainer.optimization.num-jobs-final=2 \
--trainer.optimization.initial-effective-lrate=0.003 \
--trainer.optimization.final-effective-lrate=0.0003 \
--trainer.optimization.minibatch-size=256,128,64 \
--trainer.optimization.proportional-shrink=50.0 \
--trainer.shuffle-buffer-size=2000 \
--egs.dir="$egs" \
--use-gpu=true \
--reporting.email="$reporting_email" \
--dir=$dir || exit 1;
fi
exit 0;