luvaihassanali/tf_sklearn_facialrecog

Facial recognition systems

tf_sklearn_facialrecog

Evaluation of Eigenfaces System Luvai Hassanali
*sklearn, tensorflow, and matplotlib required (installed using pip3)

Notes:

• /.cache folder contains LFW dataset
• /experiments folder contains four .txt files with results from experimentation
• /figures folder contains images used in paper and additional ones which were excluded There are 6 models discussed in my paper and 5 .py files (#2 contains two models). Each .py file can be run as is to display accuracy and time measurements discussed about in my paper:

1. eigenfaces.py has the basic eigenface system:

• running the file will load LFW set, split into test/training, calculate weights of training set, then classify the test set
• the file will also display the first 12 eigenvectors computed by PCA
• line 61 - 67 can be uncommented and the last 52 predictions will be visually displayed, examples of these figures are used in Figure 2 of my paper

2. eigenfaces_v2.py contains implementation for Scikit-Learn eigenface system AND Scikit-Learn mlp model:

• the .py file is setup to run scikit-learn eigenface system which uses KNN algorithm
• it will load LFW set, split into test/training, project into eigenspace, then classify the test set
• this file also contains implementation of mlp fully dependent on Scikit-Learn which is an experiment discussed in my paper
• to change from knn to mlp implementation all you need to do is uncomment line 52 and comment line 53
• with mlp classifier epoch + loss will be printed to console during execution and you can uncomment line 58-62 to display a plot of this
• the end of the file will display plot for predictions and projections separately, regardless of implementation

3. mlp.py contains TensorFlow implementation of MLP:

• the file will load data and train for 100 epochs
• accuracy is displayed to console every 10 epochs
• line 64 - 71 can be uncommented to visualize last 10 epoch predictions
• plot of accuracy over 100 epochs is displayed at end

4. cnn_grey.py has implementation of CNN with greyscale input:

• the file will load and split data then begin training for 20 epochs
• accuracy at epoch will be printed to console for each epoch
• top 4 patch plots along with orignal images (one for each class) will display at end

5. cnn_rgb.py has implementation of CNN with RGB input:

• the file will load and split data then begin training for 20 epochs
• accuracy at epoch will be printed to console and patch plots will display at end
• you may find it useful to reduce number of epochs as each version of the cnn takes about 10 minutes for 20

The /experiments folder contains four text files each with examples of accuracy results for the network specified:

• cnn_greyscale_accuracy_different_ksize: results from testing different ksize on max pool layer on CNN with greyscale input
• cnn_rgb_accuracy_different_ksize: results from teseting different ksize on max pool layer but with CNN that takes RGB input
• mlp_accuracy_different_weights: results from different weight configurations of the 2 hidden layers of TensorFlow implementation of MLP
• sklearn_mlp_loss_different_hidden: results from mlp fully-dependent on Scikit-Learn library from testing different number of hidden layer sizes