This repository contains the code that was used in the experiments of this paper.
Tested with keras 1.1.0 with theano backend and python 2.7.12.
Requieres the installation of scikit-learn.
We recommend you to use our exemple_script.sh. In this file you can specify the dataset and the options you want to use.
trainingAnnotations.txt must contain the list of the training images followed by the targets:
path_img_name_1.jpg y1 y2 y3
path_img_name_2.jpg y1 y2 y3
...
validationAnnotations.txt and testAnnotations.txt must contain the list of the validation and test images with the same format.
Importantly the images and annotation files must be located at /pathToData/. In other words, 'path_img_name_1.jpg" is the path of the first traininng image relatively to /pathToData/.
- Download the VGG16 weights
- Download the resNet weights
You need to change the weight file paths in VGG16_sequential.py and resnet50.py.
We provide 4 main files:
- resNetStandard.py: resnet in the case where the validation set is automatically extracted from the training set.
- resNetStandardWithVal.py: resnet in the case where the validation set is given.
- VGGStandard.py: VGG16 in the case where the vqlidation set is automatically extracted from the training set.
- VGGStandardWithVal.py: VGG16 in the case where the validation set is given.
The JOB_ID is a job id used to save the network weights. You can give any number. $rootpathData is the path to your dataset folder.
-
BatchNormalization: -bn: with BN -bnba: with BN before the last activation -nbn: no BN
-
Finetunning: '-ft x' with x in {0,1,2,3}, finetune x blocks nbBlock=int(sys.argv[idarg+1])
-
Batch size: '-bs x': use batches of size x
-
Optimizer: '-opt x' with x in {adam, adadelta, rmsprop,adagrad}
-
Regression layer: '-rf x' with x in {cov,fc1}. otherwise the default value is fc2
-
Dropout: '-do x': with x in {-1,0,1,2}
- -1: refered to as 00 in the paper
- 0: 10
- 1: 01
- 2: 11
For any question, please contact Stéphane Lathuilière.