A developed deep learning architecture for classification

For AAALGO Users

This code requires official tensorflow models and picpac to run. After cloning and cd into the repo, run the following to create the necessary links:

ln -s /shared/s2/users/wdong/picpac/build/lib.linux-x86_64-3.5/picpac.cpython-35m-x86_64-linux-gnu.so ./
ln -s /shared/s2/users/wdong/cls/models ./

Requirements

Python
Tensorflow
Picpac

Importing database

picpac-import

Stream data into same image database format
Refer to picpac for more info

Eg:  
picpac-import -f 2 ImageDirectory db  
#ImageDirectory contains N subdirectories named 0, 1, ..., each containg images for one category

Training

cls-train.py

Trainer of images classification, allows evaluation during training

arguments:  
--db  
		Training image database  
--classes  
		Numbers of categories of classification  
--channels  
		Numbers of channels used in training  
optional arguments:  
--opt  
        	Optimizer of network training, choices of adam(default) or gradient  
--test_db  
        	Evaluating image database  
--model  
        	Directory to save models  
--learning_rate  
		Initial learning rate  
--test_steps  
		Number of steps to run evaluation  
--save_steps   
		Number of steps to save model  
--max_steps  
		Number of steps to run trainer  
--split  
		Split train image into this number for cross-validation  
--split_fold  
		Part index for cross-validation  
Eg:   
./cls-train.py --db db.train --test_db db.test --classes 2 --channels 1  
./cls-train.py --db db.train --split 5 --split_fold 3

fcn-cls-train.py

Trainer of segmented images classification

arguments:  
--pos  
		Training image dataset with positive part labelled  
--neg  
		Training image dataset without labelling  
Other arguments have the same usage as cls-train.py  
Eg:
./fcn-cls-train.py --pos db.pos --neg db.neg

Evaluating

cls-predict.py

Evaluation of classification using cls-train.py

arguments:  
--input  
		Input directory of images, can be any directory  
--model  
		Model saved during training  
--channels  
		Channels of images used  
Eg:  
./cls-predict.py --input ImageDirectory --model model/200000 --channels 1

fcn-cls-val.py

Evaluation of classification using fcn-cls-train.py

arguments:  
--db  
		Evaluation image database

aaalgo / cls