23uday/multiaspect-matrixfactorization-cnn-analysis

Project1CodeDSAA2022

Paper Data and Paper Results

Link to Data Files

The above link contains 3 Items:-

1. CIFAR100Dataset : Contains zip file for CIFAR100 datasets
2. ExperimentResultsResNet18: Contains the results of an analysis of a ResNet-18 on CIFAR-100
3. A Pretrained ResNet-18 on CIFAR-100

Requirements

1. Python 3.7.6 and above
2. Numpy 1.18.1 and above
3. Pytorch 1.4.0 and above
4. Scipy 1.4.1 and above
5. install gradcam from https://github.com/jacobgil/pytorch-grad-cam

Command Template to Run the experiment

Prototype Command:

python3 dissectKNNResNet18.py FullPathToCifar100SuperClass/train/ FullPathToCifar100SuperClass/valKDD/ PathToPretrainedResnet18/Resnet18Cifar100 OutputFolderDirectory ResNet18Cifar100 20 20 50 1 1 1 1 1 0.001 0.00001 100 sgd NNDSVD False False False

Prototype Command 2 :

python3 dissectKNNResNet18.py /data/CifarLocal/train /data/CifarLocal/val/ None /outputs/DSAA2022Results/Cifar10/ ResNet18Cifar10 10 20 50 1 1 1 1 1 0.001 0.00001 100 sgd NNDSVD False False False

The parameter values shown were used for experimentation.

Use the following command to get help:

python3 dissectKNNResNet18.py -h

usage: dissectKNNResNet18.py [-h] rootDir rootDirTest networkFile outputFolderName NetworkName Rank1 numGroups maxIters lmbdaF lmbdaTV lmbdaOrtho samplingFactor samplingFactorTest lr wd numEpochs {sgd,adam,rmsprop} {random,ID,NNDSVD} {True,False} {True,False} {True,False}

positional arguments:

• rootDir Enter the name of root folder which containts the data subfolders :
• rootDirTest Enter the name of root folder which containts the test data subfolders :
• networkFile Enter the name of root folder which containts the Network :
• outputFolderName Enter the name(Path) of the Output Folder :
• NetworkName Enter the name(Path) of the network file :
• Rank1 Enter Rank 1 :
• numGroups Enter the number of groups:
• maxIters Enter Max no. of iterations:
• lmbdaF Enter lambda F:
• lmbdaTV Enter lambda TV:
• lmbdaOrtho Enter orthogonality penalty:
• samplingFactor Enter the ratio of dataset to be used:
• samplingFactorTest Enter the ratio of dataset to be used:
• lr learning rate for SGD to be used:
• wd Enter the weight decay:
• numEpochs Enter the number of epochs:
• {sgd,adam,rmsprop} Enter the optimization algorithm
• {random,ID,NNDSVD} Enter the initialization for P
• {True,False} Enter if class based training is required
• {True,False} Enter true if Data matrices have to be ignored
• {True,False} Enter true if F matrix is to have group sparsity