Home

Awesome

Auxiliary Image Regularization for Deep CNNs with Noisy Labels

Copyright (C) {2016} {Samaneh Azadi, Jiashi Feng, Stefanie Jegelka, Trevor Darrell}

This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, version 3 of the License.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program. If not, see http://www.gnu.org/licenses/.

The research is to be cited in the bibliography as:

S. Azadi, J. Feng, S. Jegelka, T. Darrell, “Auxiliary Image Regularization for Deep CNNs with Noisy Labels”, International Conference on Learning Representations (ICLR), 2016.

The code is to run Auxiliary Image Regularizer (AIR) for classification on different datasets Paper at: http://arxiv.org/pdf/1511.07069v2.pdf Code by: Samaneh Azadi

											Help:

OBJECTIVE = lambda_sen * normalized \sum ||.||_2 + lambda_las * ||.||_1 +... (1/num_samples) * Loss + (mu/2*num_classes) * ||w||_2^2

rho : the factor used in the Augmented Lagrangain term if use 'SADMM' as the solver and regularization parameter if use 'svm-sgd'

train : number of training samples per class

num_classes : number of categories

dataType : 'imagenet' / 'cifar10' / 'mnist'

noisy_label : 0/1 want to add noise to the labels?

noise_ratio : with what ratio? e.g. if you want "50% = 1/2" noise, set noise_ratio to "2"

state : 'test'/'val'

solver : 'SADMM'/'svm-sgd'

percent_features : What percent of the feature values to be kept? want to make the feature vector more sparse for the sake of memory/computation cost

num_parts : divide the training and testing data into 'num_parts' batches to save them in a matrix at the first time of reading the features

id : 0-9 which random setting to use?

					Continue Help : Only Valid for 'SADMM' solver

epochs : number of epochs to do training

batch_size : number of samples to use per iteration

max_rho: the max value for adaptive rho used in the Augmented Lagrangian term

mult_rho : {1.1, 1.3, 1.5, 1.7} multiplier factor for rho in each step

ci_first, ci_last, par_finished : can run the one-vs-all classification for different categories in parallel to make it finish faster ===>>

ci_first : the first category in the current parallel run of algorithm

ci_last : the last category in the current parallel run of algorithm

par_finished : 0/1 : Is the parallel training of all categories done and want to start testing? (If parallel training not finished yet, it continues training on the categories not trained yet)

(While using "deep_onevsall_nus_data.m", you should first train the one-vs-others classifiers in parallel cores with "par_finished = 0" and then set "par_finished = 1" to do evaluation. If want to train only one classifier at each time, you can set "par_finished = 1").

									Instruction to run the code

  1. Fill all lines with the correct directory path in "SADMM_normalized.m" as well as "deep_reg_multiclass.m" or "deep_classification_onevsall_nus.m".

    The extracted deep feature vectors should be saved as '.h5' files under their own category directory as 'calDir/train/specific_category/'. We extracted FC_7 features from AlexNet for all images.

    calDir : includes extracted features (.h5) in subdirectories of "calDir/train/class_name_i" and "calDir/test/class_name_i";Read both train and validation data from this directory For CIFAR10 features should be saved in subdirs "calDir/data_batch_i/class_name_i" and "test/class_name_i"

    caffeDir : includes the original png images; only the path name is needed to produce a txt file including images as a list

    dataDir: the path to save the model weights

    outDir: the path to save the final result

    paramDir: the path to save the parameters

    For NUS_WIDE_LITE dataset, the images are read based on the txt file in the NUS_LITE_data directory.

  2. Start MATLAB and run the VLFeat setup command: vlfeat-0.9.19/toolbox/vl_setup

  3. Run deep_reg_multiclass.m or dep_onevsall_nus_data.m (the final hyper-parameters found for the results in the paper were run similar to run_AIR.sh file. For NUS_LITE dataset, one can use the parameters used in run_onevsall_nus.sh to get the same results as in the paper.)