-----------------------
DeepCNF_Package (v1.02)
date: 2015.05.30
-----------------------

Title:
Training Conditional Random Fields by Maximizing AUC

Author: Sheng Wang 
Contact email: [email protected] 

References:
1. AUC-maximized Deep Convolutional Neural Fields for Protein Sequence Labeling
	Sheng Wang#, Siqi Sun, Jinbo Xu
	ECML/PKDD, 2016
	https://link.springer.com/chapter/10.1007/978-3-319-46227-1_1

2. AUCpreD: Proteome-level Protein Disorder Prediction by AUC-maximized Deep Convolutional Neural Fields
	Sheng Wang#, Jianzhu Ma, Jinbo Xu
	ECCB, 2016
	Bioinformatics, 2016
	https://academic.oup.com/bioinformatics/article-abstract/32/17/i672/2450776

--------------------------------------------------------


=========
Abstract:
=========

1. The training program of Deep Convolutional Neural Filed ( DeepCNF ) 
for maximizing (1) log probability, (2) posterior probability , or (3) AUC value.

2. The predicting program of DeepCNF for a given feature file 
by using a trained model.



========
Install:
========

1. download the package

git clone https://github.com/realbigws/DeepCNF_AUC/
cd DeepCNF_AUC/

+++++++++++++++++++++

2. compile

2.1 compile under single CPU

cd source_code/
        make
cd ../

-------------

2.2 compile under MPI/OpenMP

cd source_code/
	make mpi
cd ../



======
Usage:
======

1. DeepCNF_Train

Usage :

mpirun -np NP ./DeepCNF_Train -r train_file -t test_file
             -w window_str -d node_str -s state_num -l feat_num [-S feat_select]
            [-n finetune_num] [-f finetune_reg] [-m init_model] [-o out_root]
            [-G gate_function] [-W label_weight] [-M method] [-D AUC_degree]
Options:

-np NP :             number of processors.

-r train_file :      training file.

-t test_file :       testing file.

-w window_str :      window string for DeepCNF. e.g., '5,5'

-d node_str :        node string for DeepCNF. e.g., '40,20'

-s state_num :       state number.

-l feat_num :        feature number at each position.

-S feat_select :     feature selection. e.g., '1-7,9' [default uses all features]

-n finetune_num :    fine-tune iteration number. [default is 200]

-f finetune_reg :    fine-tune regularizer. [default is 0.5]

-m init_model :      file for initial model parameters. [optional, and default is NULL]

-o out_root :        output directory for trained models. [optional, and default is 'MODELS/']

-G gate_function :   gate function type: sigmoid (1), tanh (2), and relu (0). [default is 1]

-M method :          maximize log_prob (0), posterior_prob (1), or AUC (2). [default is 0]

-W label_weight :    label weight. e.g., '0.1,0.9'. [default is 1 for each label]

-D AUC_degree :      degree (1-30) of polynomials for AUC [default is 3].

-------------------------------------


2. DeepCNF_Pred

Usage :

./DeepCNF_Pred -i input_file -w window_str -d node_str -s state_num -l feat_num
               -m init_model [-S feat_select]
Options:

-i input_file :      input feature file.

-w window_str :      window string for DeepCNF. e.g., '5,5'

-d node_str :        node string for DeepCNF. e.g., '40,20'

-s state_num :       state number.

-l feat_num :        feature number at each position.

-m init_model :      file for trained model.

-S feat_select :     feature selection. e.g., '1-7,9' [default uses all features]

---------------------------------------



========
Example:
========

1. Maximize log probability:

cd examples/
	./DeepCNF_Train -r 2fzlA.reso -t 2fzlA.reso -w 5 -d 5 -s 3 -l 40 -n 50
cd ../

---------------------

2. Maximize AUC value:

cd examples/
	./DeepCNF_Train -r 1kq1S.feat -t 1kq1S.feat -w 5 -d 5 -s 2 -l 129 -n 50 -M 2 
cd ../

---------------------

3. Predict the labels of a given feature file by using a certain model:

cd examples/
	./DeepCNF_Pred -i 1kq1S.feat -w 5 -d 5 -s 2 -l 129 -m model.10
cd ../


---------------------

4. make feature file for DeepCNF

cd examples/
	./DeepCNF_FeatMake 5fgnA.profile 5fgnA.label > 5fgnA.feat
cd ../



==================
Input file format:
==================

1. Training data format

//-> data format ( length, features, labels [should be authentic] )
//-> example below:

78
feat1 feat2 feat3, ..., featN
....  (with 78 feature lines)
0
1
0
2
...  ( with 78 label lines)
54
feat1 feat2 feat3, ..., featN
....  (with 54 feature lines)
0
1
0
2
...  ( with 54 label lines)

----------------------------------


2. Testing data format

//-> data format ( length, features, labels [can be anything] )
//-> example below:

78
feat1 feat2 feat3, ..., featN
....  (with 78 feature lines)
-1
-1
-1
-1
...  ( with 78 label lines)
54
feat1 feat2 feat3, ..., featN
....  (with 54 feature lines)
-1
-1
-1
-1
...  ( with 54 label lines)