Deep-Learning

##Environment: Linux with BLAS, LAPACK installed

##Package dependency:

  1. Armadillo C++ linear algebra library
  2. BLAS
  3. LAPACK

##How to compile: cd src sh compile.sh

##How to run: Before training, use the following to merge data and label, if label is not 0N, please supply out_map to save the mapping of string label to numeric label(0N).

script/merge_data_label.py train.ark train.lab new_train.ark [out_map]

Each epoch will save the model to output_model, if load_model supplied, will load that model first and start training.

bin/train learning_rate(0.01) batch_size(10) structure(5-4-3) max_epoch(100) new_train.ark output_model [load_model]

if the .ark has answer(new_train.ark) has_answer=1, if not(test.ark) has_answer=0 bin/predict test_file model_name raw_result has_answer(1/0)

##For Kaggle Sumission 48_phone->39_phone

map_phone_label.py raw_result final_result out_map 48-39.map

1942_state->39_phone

map_state_label.py raw_result final_result state-48-39.map

Upload final_result