mnist_euslisp

learning mnist with euslisp

Download mnist dataset

• http://yann.lecun.com/exdb/mnist/

Decompress mnist dataset

$ gzip -dc train-images-idx3-ubyte.gz >train-images-idx3-ubyte
$ gzip -dc train-labels-idx1-ubyte.gz >train-labels-idx1-ubyte
$ gzip -dc t10k-images-idx3-ubyte.gz >test-images-idx3-ubyte
$ gzip -dc t10k-labels-idx1-ubyte.gz >test-labels-idx1-ubyte

Convert dataset to text file

$ od -An -v -tu1 -j16 -w784 train-images-idx3-ubyte | sed 's/^ *//' | tr -s ' ' >train-images.txt
$ od -An -v -tu1 -j8 -w1 train-labels-idx1-ubyte | tr -d ' ' >train-labels.txt
$ od -An -v -tu1 -j16 -w784 test-images-idx3-ubyte | sed 's/^ *//' | tr -s ' ' >test-images.txt
$ od -An -v -tu1 -j8 -w1 test-labels-idx1-ubyte | tr -d ' ' >test-labels.txt

$ rm *-ubyte
$ rm *.gz

Build

# for C
$ cd ./MATPROD
$ make

# for cuBLAS
$ cd ./eus_cuda_matrix
$ mkdir build
$ cd build
$ cmake ..
$ make

Accelarate simulation, otherwise simulation won't finish!

1. Dump mnist datasets as euslisp file

$ roseus
irteusgl$ (compile-file "mnist.l")
irteusgl$ (progn (load "mnist.so") (test) (dump-loadable-structure "mnist-datasets.l" *train-images* *train-labels* *test-images* *test-labels*))

2. Compile euslisp file

irteusgl$ (compile-file "nn.l")

Utilize OpenBLAS (Optional)

$ sudo apt-get install libopenblas-base libopenblas-dev
$ sudo update-alternatives --config libblas.so.3
Select 3.

Run simulation

$ roseus
irteusgl$ (load "nn.so")

irteusgl$ (test-mnist-batch 200 :cblas)  ;; train with minibatch size 200, CBLAS/OpenBLAS
irteusgl$ (test-mnist-batch 200 :cublas) ;; train with minibatch size 200, cuBLAS

irteusgl$ (test-mnist-batch-test 200 :cblas)  ;; train and test sequentially with CBLAS/OpenBLAS
irteusgl$ (test-mnist-batch-test 200 :cublas) ;; train and test sequentially with cuBLAS

irteusgl$ (test-mnist-test) ;; test with train images
irteusgl$ (test-mnist-test) ;; test with test images