Project1_Pro: Feature extraction and Clustering

1.Task

• Extract features(fc7) from a set of images(cifar10)
• Do clustering and clustering

2.Analysis

I find somebody use conv5 for face clustering, conv5 inherently contain spatial information.

In CNN model, fully connected reps will have a great deal of this spatial information removed but will contain more abstractions.

May be it is a trade-off situation.

3.Solutions

Step1. Download PNG dataset and Processing

• create folder Data, Examples,bvlc_reference_caffenet
• $ ./0.prepare.sh to Download the cifar10 dataset
• $ ./1.create_cifar10_imagenet_lmdb.sh to convert png dataset into lmdb format
• $ ./2.make_imagenet_mean.sh to compute the mean file of the dataset
• $ python 2.binaryproto2npy.py to convert mean file from binaryproto format into npy format
• copy deploy.prototxt, train_val.prototxt, solver.prototxt, bvlc_reference_caffenet.caffemodel from caffe-master/model to bvlc_reference_caffenet
• $ ./3.train.sh to finetuning the model from caffenet model on cifar10 dataset
• $ ./4.test.sh to get the accuracy from model finetuned, accuracy is: 0.8752 with caffenet_train_iter_20000.caffemodel
• $ python 5.feature_extract.py to extract fc7 features for every single jpg
• $ python 6.feature_numpy_combine.py to combine all features into one single txt file
• $ python 7.labelListExtract.py to extract origin label into a list file
• $ python 8.extractImageNumber.py to extract filename number into a list file
• $ python 9.Kmeans_on_fc7.py to do kmeans on fc7 features and store the result into txt file
• $ python 10.plotImages.py to plot the samples of images for each class

Step6. Results and analysis

Data results

99.29% y: 0 (4907L,) {0.0: 12, 1.0: 12, 2.0: 1, 3.0: 2, 6.0: 3, 8.0: 4872, 9.0: 5}
67.43% y: 1 (7368L,) {0.0: 42, 1.0: 147, 2.0: 95, 3.0: 4968, 4.0: 77, 5.0: 410, 6.0: 171, 7.0: 1358, 8.0: 59, 9.0: 41}
99.94% y: 2 (4748L,) {1.0: 4745, 9.0: 3}
99.33% y: 3 (4812L,) {1.0: 4, 2.0: 13, 3.0: 2, 4.0: 11, 5.0: 2, 6.0: 4780}
99.46% y: 4 (3512L,) {4.0: 3, 5.0: 16, 7.0: 3493}
96.46% y: 5 (5083L,) {0.0: 1, 2.0: 22, 3.0: 3, 4.0: 4903, 5.0: 27, 6.0: 13, 7.0: 110, 8.0: 2, 9.0: 2}
98.12% y: 6 (4901L,) {0.0: 11, 1.0: 6, 2.0: 4818, 3.0: 8, 4.0: 5, 5.0: 6, 6.0: 28, 7.0: 9, 8.0: 5, 9.0: 5}
96.63% y: 7 (5013L,) {0.0: 1, 1.0: 65, 5.0: 1, 7.0: 1, 8.0: 14, 9.0: 4931}
97.05% y: 8 (5083L,) {0.0: 4933, 1.0: 18, 2.0: 41, 3.0: 2, 4.0: 1, 6.0: 1, 7.0: 28, 8.0: 48, 9.0: 11}
99.23% y: 9 (4573L,) {1.0: 3, 2.0: 10, 3.0: 15, 5.0: 4538, 6.0: 4, 7.0: 1, 9.0: 2}

average accuracy is: 95.254%

Select 20 images randomly of each class classified by Kmeans