CIMON.ipynb : This file contains all the code that was written as a part of the project
CIMON_Visualization.ipynb : This notebook contains code for the PCA visualization of the hash codes
CIMON_Pseudo Graph Viz.ipynb: This notebook contains the code for visualizing the pseudo graph post spectral clustering.
CIMON_paper.pdf : PDF of the paper CIMON : Towards higher quality hash codes
Project Report.pdf : Project Report
Requirements.txt : Contains all the dependencies for running the CIMON.ipynb
-
DL_CIMON folder has been shared. If not, you can access it with the following link
-
Add a shortcut of this folder to the "MyDrive" folder in your drive.
-
Now you are good to go to run the notebook.
-
We present the first network that uses learned deep hash codes for classification purposes.
-
We first train the CIMON network in a (modified manner as mentioned above) to learn a network that can learn good hash codes for images. We train a small MLP (as shown in Figure 1) using the training set of the STL10 dataset. This is done as follows:
-
The image is passed through the trained network to obtain the hash code for the image.
-
Using the hash code as the input to the MLP, we train the MLP to classify the image given the hash code.
-
Just like any other network, using the cross entropy loss, we train the MLP
-
We have also added visualizations of the hash codes that the network has learnt. The visualizations support our hypothesis that classification can be performed from the hash code space because for the classes of images taken, the network was able to produce hash codes that form visibly separate clusters in low dimensional space. (Figure 2) There are only 5000 training examples in the STL10 dataset, and we were able to achieve a good accuracy by first training an unsupervised learning model to learn hash codes, and then training a supervised learning model on just 5000 images. These are results that we obtained on the STL10 dataset:
-
Best test accuracy : 86.56 %
-
Best train accuracy : 84.9 %
Training Plot on STL10 Dataset
For a more detailed report, have a look at Project Report.pdf