File system
- data/ (data main dir)
- data/fbank
- data/label
- data/mfcc
- data/phones
- data/merge (create by your self)
- dnn/ (git main dir)
- dnn/hw1 (all hw1 file in here)
- dnn/hw1/out (test prediction csv file will generate to this dir)
- dnn/hw1/save_models (for model saving)
Data preprocessing
- create merge dir in data/
- run read.py
- run fbank.py
- run mfcc.py
- run l48to39.py
Compile main program
- Download Eigen from Eigen main website
- link
path/to/Eigen
to your include path - go to hw1 dir
- type "make main" in hw1
- You could also use
g++ -O2 --std=c++11 main.cpp -o main -Ipath/to/Eigen
if you don't want to include Eigen in your include path.
- You could also use
Run main program
- run main by ./main
- it will output test prediction for every 5000 mini-batchs to out/ in default settings
For concatenated frame feature training
- run new_fbank.py
- run new_l48to39.py
- set the parameters in new_main.cpp
- type
make new_main
ormake omp_new_main
if you want to use cpu parallelism - run program by
./new_main
or./omp_new_main
- default settings have the best result parameters