CNN

This program implements a convolutional neron network, and is applied to solve digit recoginition task (MNIST).

The best result of recognition accuracy so far, using the NN encoded in Main.hs: 98.43%.

Note

This project was only a initial learning/experiment of neuron network. Its source code is now further absorbed in the neural-netowrk.

The depend library HMatrix requires some efforts to install.
- The version on Hackage of 0.17.0.2 does not work. It will produce linkage errors, complaining that symbols "random" and "rand48" not found. Instead, please download the head version of HMatrix from github. Place it on the upper directory of this project, as specified in the stack.yaml.
- HMatrix itself requires openblas to get build. It can be found here and the binary package of version 0.2.15 works ok.
  - or, if you want to compile latest Openblas your self. You will need get a MSYS2 environment, install the MINGW_W64, etc. HMatrix needs the OpenBlas with LAPACK included, therefore GFORTRAN is also needed.
- Apart from the binray package of openblas, the libgfortran-3.dll is needed to be seen under the GHC_INSTALL_DIR\mingw\bin. I find a easy way by starting a msys2 shell (the stack utility already installs one), run packman -S mingw-w64-x86_64-gcc-libgfortran, after that you will find a libfortran-3.dll under the MSYS2_INSTALL_DIR\mingw64\bin. Then simply copy it over.

Name		Name	Last commit message	Last commit date
Latest commit History 13 Commits
cbits		cbits
.gitignore		.gitignore
CNN.hs		CNN.hs
LICENSE		LICENSE
Main.hs		Main.hs
Parser.hs		Parser.hs
README.md		README.md
Setup.hs		Setup.hs
Utils.hs		Utils.hs
cnn.cabal		cnn.cabal
stack.yaml		stack.yaml