path: root/train.py

#!/usr/bin/env python
# Author: Vasil Zlatanov, Nunzio Pucci
# EE4 Pattern Recognition coursework
#
# usage: train.py [-h] -i DATA [-m EIGEN] [-M REIGEN] [-e ENSEMBLE] [-b]
#                 [-R RANDOM] [-n NEIGHBORS] [-f FACES] [-c] [-s SEED]
#                 [-t SPLIT] [-2] [-p] [-l] [-r RECONSTRUCT] [-cm] [-q] [-pr]
#                 [-alt]

import warnings
with warnings.catch_warnings():
    warnings.filterwarnings("ignore", category=DeprecationWarning)
    import imp

import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
import sys
import random
import os
import psutil
from random import randint

from sklearn.neighbors import KNeighborsClassifier
from sklearn.decomposition import PCA
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import confusion_matrix
from sklearn.metrics import accuracy_score

import argparse
import numpy as np

from numpy import genfromtxt
from numpy import linalg as LA

from timeit import default_timer as timer

n_faces = 52
n_cases = 10
n_pixels = 2576

# subtract the normal face from each row of the face matrix
def normalise_faces(deviations_tr, average_face, faces):
    faces = np.subtract(faces, np.tile(average_face, (faces.shape[0],1)))
    return np.divide(faces, deviations_tr)

# Split data into training and testing sets
def test_split(n_faces, raw_faces, split, seed):
    random.seed(seed)

    raw_faces_split = np.split(raw_faces,n_cases)
    n_training_faces = int(round(n_cases*(1 - split)))
    n_test_faces = n_cases - n_training_faces
    faces_train = np.zeros((n_faces, n_training_faces, n_pixels))
    faces_test = np.zeros((n_faces, n_test_faces, n_pixels))
    target_train = np.repeat(np.arange(n_faces), n_training_faces)
    target_test = np.repeat(np.arange(n_faces), n_test_faces)

    for x in range(n_faces):
        samples = random.sample(range(n_cases), n_training_faces)
        faces_train[x] = [raw_faces[i+n_cases*x] for i in samples]
        faces_test[x] = [raw_faces[i+n_cases*x] for i in range(n_cases) if i not in samples]

    faces_train = faces_train.reshape(n_faces*n_training_faces, n_pixels)
    faces_test = faces_test.reshape(n_faces*n_test_faces, n_pixels)
    return faces_train, faces_test, target_train, target_test

def draw_results(args, target_test, target_pred):
    acc_sc = accuracy_score(target_test, target_pred)
    cm = confusion_matrix(target_test, target_pred)
    print('Accuracy: ', acc_sc)
    if (args.conf_mat):
        plt.matshow(cm, cmap='Blues')
        plt.colorbar()
        plt.ylabel('Actual')
        plt.xlabel('Predicted')
        plt.show()
    return

def test_model(M, faces_train, faces_test, target_train, target_test, args):
    raw_faces_train = faces_train
    raw_faces_test = faces_test

    explained_variances = ()


    distances = np.zeros(faces_test.shape[0])

    if args.pca or args.pca_r:
        # faces_pca containcts the principial components or the M most variant eigenvectors
        average_face = np.mean(faces_train, axis=0)
        #PLOTTING MEAN FACE
        #plt.imshow(average_face.reshape([46,56]).T, cmap = 'gist_gray')
        plt.show()
        if args.classifyalt:
            deviations_tr = np.ones(n_pixels)
        else:
            deviations_tr =  np.std(faces_train, axis=0)
        # deviations_tst = np.std(faces_test, axis=0)
        faces_train = normalise_faces(deviations_tr, average_face, faces_train)
        faces_test = normalise_faces(deviations_tr, average_face, faces_test)
        if (args.pca_r):
            e_vals, e_vecs = LA.eigh(np.dot(faces_train, faces_train.T))
            e_vecs = np.dot(faces_train.T, e_vecs)
            e_vecs = e_vecs/LA.norm(e_vecs, axis = 0)
        else:
            e_vals, e_vecs = LA.eigh(np.cov(faces_train.T))
            # e_vecs = normalise_faces(np.mean(e_vecs,axis=0), e_vecs)
        #PLOTTING NON-ZERO EVALS
        #if args.pca:
        #    plt.semilogy(range(2576), np.absolute(416*np.flip(e_vals)))
        #    plt.show()

        e_vals = np.flip(e_vals)
        e_vecs = np.fliplr(e_vecs).T

        if args.random:
            random_features = random.sample(range(M-args.random, M), args.random)
            for i in range(args.random):
                e_vals[M-i] = e_vals[random_features[i]]
                e_vecs[M-i] = e_vecs[random_features[i]]

        e_vals = e_vals[:M]
        e_vecs = e_vecs[:M]

        deviations_tr = np.flip(deviations_tr)
        # deviations_tst = np.flip(deviations_tst)
        faces_train = np.dot(faces_train, e_vecs.T)
        faces_test = np.dot(faces_test, e_vecs.T)

        rec_vecs = np.add(np.tile(average_face,
            (faces_test.shape[0], 1)), np.dot(faces_test, e_vecs) * deviations_tr)
        distances = LA.norm(raw_faces_test - rec_vecs, axis=1);

        if args.reconstruct:
            rec_vec = np.add(average_face, np.dot(faces_train[args.reconstruct], e_vecs) * deviations_tr)
            ar = plt.subplot(2, 1, 1)
            ar.imshow(rec_vec.reshape([46,56]).T, cmap = 'gist_gray')
            ar = plt.subplot(2, 1, 2)
            ar.imshow(raw_faces_train[args.reconstruct].reshape([46,56]).T, cmap = 'gist_gray')
            plt.show()

    if args.lda:
        if args.pca_r or (args.pca and M > n_training_faces - n_faces):
            lda = LinearDiscriminantAnalysis(n_components=M, solver='svd')
        else:
            lda = LinearDiscriminantAnalysis(n_components=M, store_covariance='True')

        faces_train = lda.fit_transform(faces_train, target_train)
        faces_test = lda.transform(faces_test)
        class_means = lda.means_
        e_vals = lda.explained_variance_ratio_
        # scatter_matrix = lda.covariance_; print("Rank of scatter:", LA.matrix_rank(scatter_matrix))

    if args.faces:
        if args.lda:
            for i in range(10):
                ax = plt.subplot(2, 5, i + 1)
                ax.imshow(class_means[i].reshape([46,56]).T)
        else:
            for i in range(args.faces):
                ax = plt.subplot(2, args.faces/2, i + 1)
                ax.imshow(e_vecs[i].reshape([46, 56]).T, cmap = 'gist_gray')
        plt.show()

    if args.principal:
        e_vals = np.multiply(np.divide(e_vals, np.sum(e_vals)), 100)
        plt.bar(np.arange(M), e_vals[:M])
        plt.ylabel('Varaiance ratio (%)');plt.xlabel('Number')
        plt.show()

    if args.grapheigen:
        #graph_eigen()
        # Colors for distinct individuals
        cols = ['#{:06x}'.format(randint(0, 0xffffff)) for i in range(n_faces)]
        pltCol = [cols[int(k)] for k in target_train]
        fig = plt.figure()
        ax = fig.add_subplot(111, projection='3d')
        ax.scatter(faces_train[:, 0], faces_train[:, 1], faces_train[:, 2], marker='o', color=pltCol)
        plt.show()

    classifier = KNeighborsClassifier(n_neighbors=args.neighbors)
    classifier.fit(faces_train, target_train)
    target_pred = classifier.predict(faces_test)
    if args.prob:
        targer_prob = classifier.predict_proba(faces_test)
        targer_prob_vec = np.zeros(104)
        for i in range (104):
            j = int(np.floor(i/2))
            targer_prob_vec [i] = targer_prob[i][j]
        avg_targer_prob = np.zeros(n_faces)
        for i in range (n_faces):
            avg_targer_prob[i] = (targer_prob_vec[2*i] + targer_prob_vec[2*i + 1])/2
        #WE CAN FIX THIS BY RESHAPING TARGER_PROB_VEC AND TAKING THE MEAN ON THE RIGHT AXIS
        plt.bar(range(n_faces), avg_targer_prob)
        plt.show()

        #Better n_neighbors = 2
    return target_pred, distances

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("-i", "--data", help="Input CSV file", required=True)
    parser.add_argument("-m", "--eigen", help="Number of eigenvalues in model", type=int, default = 10 )
    parser.add_argument("-M", "--reigen", help="Number of eigenvalues in model", type=int)
    parser.add_argument("-e", "--ensemble", help="Number of ensemmbles to use", type=int, default = 0)
    parser.add_argument("-b", "--bagging", help="Number of bags to use", action='store_true')
    parser.add_argument("-R", "--random", help="Number of eigen value to randomise", type=int)
    parser.add_argument("-n", "--neighbors", help="How many neighbors to use", type=int, default = 1)
##USING STANDARD 1 FOR NN ACCURACY
    parser.add_argument("-f", "--faces", help="Show faces", type=int, default = 0)
    parser.add_argument("-c", "--principal", help="Show principal components", action='store_true')
    parser.add_argument("-s", "--seed", help="Seed to use", type=int, default=0)
    parser.add_argument("-t", "--split", help="Fractoin of data to use for testing", type=float, default=0.3)
    ### best split for lda = 22
    ### best plit for pca = 20
    parser.add_argument("-2", "--grapheigen", help="Swow 2D graph of targets versus principal components",
            action='store_true')
    parser.add_argument("-p", "--pca", help="Use PCA", action='store_true')
    parser.add_argument("-l", "--lda", help="Use LDA", action='store_true')
    parser.add_argument("-r", "--reconstruct", help="Use PCA reconstruction, specify face NR", type=int, default=0)
    parser.add_argument("-cm", "--conf_mat", help="Show visual confusion matrix", action='store_true')
    parser.add_argument("-q", "--pca_r", help="Use Reduced PCA", action='store_true')
    parser.add_argument("-pr", "--prob", help="Certainty on each guess", action='store_true')
    parser.add_argument("-alt", "--classifyalt", help="Alternative method ON", action='store_true')
    args = parser.parse_args()

    if args.lda and args.classifyalt:
        sys.exit("LDA and Alt PCA can not be performed together")

    raw_faces = genfromtxt(args.data, delimiter=',')
    targets = np.repeat(np.arange(n_faces),n_cases)

    faces_train, faces_test, target_train, target_test = test_split(n_faces, raw_faces, args.split, args.seed)

    if args.ensemble:
        n_training_faces = int(round(n_cases*(1 - args.split)))
        faces_train_ens = np.zeros((args.ensemble, n_faces, n_training_faces, n_pixels))
        for x in range(args.ensemble):
            if args.bagging:
                for k in range(n_faces):
                    samples = random.choices(range(n_training_faces), k=n_training_faces)
                    faces_train_ens[x][k] = [faces_train[i+n_training_faces*k] for i in samples]
            else:
                faces_train_ens[x] = faces_train.reshape((n_faces, n_training_faces, n_pixels))

        faces_train_ens = faces_train_ens.reshape(args.ensemble, n_faces*n_training_faces, n_pixels)

    if args.classifyalt:
        faces_train = faces_train.reshape(n_faces, int(faces_train.shape[0]/n_faces), n_pixels)
        target_train = target_train.reshape(n_faces, int(target_train.shape[0]/n_faces))

        distances = np.zeros((n_faces, faces_test.shape[0]))
        for i in range(n_faces):
            target_pred, distances[i] = test_model(args.eigen, faces_train[i],
                    faces_test, target_train[i], target_test, args)
        target_pred = np.argmin(distances, axis=0)
    elif args.reigen:
        target_pred = np.zeros((args.reigen-args.eigen, target_test.shape[0]))
        accuracy    = np.zeros(args.reigen-args.eigen)
        rec_error   = np.zeros((args.reigen-args.eigen, target_test.shape[0]))

        for M in range(args.eigen, args.reigen):
            start = timer()
            target_pred[M - args.eigen], rec_error[M - args.eigen] = test_model(M, faces_train,
                    faces_test, target_train, target_test, args)
            end = timer()
            print("Run with", M, "eigenvalues completed in ", end-start, "seconds")
            print("Memory Used:", psutil.Process(os.getpid()).memory_info().rss)
            accuracy[M - args.eigen] = accuracy_score(target_test, target_pred[M-args.eigen])
        # Plot
        print('Max efficiency of ', max(accuracy), '% for M =', np.argmax(accuracy))
        plt.plot(range(args.eigen, args.reigen), 100*accuracy)
        plt.xlabel('Number of Eigenvectors used (M)')
        plt.ylabel('Recognition Accuracy (%)')
        plt.grid(True)
        plt.show()
    elif args.ensemble:
        rec_error = np.zeros((args.ensemble, n_faces, faces_test.shape[0]))
        target_pred = np.zeros((args.ensemble, target_test.shape[0]))
        for i in range(args.ensemble):
            target_pred[i], rec_error[i] = test_model(args.eigen, faces_train_ens[i],
                    faces_test, target_train, target_test, args)

        target_pred_comb = np.zeros(target_pred.shape[1])
        target_pred = target_pred.astype(int).T
        if (args.conf_mat):
            cm = confusion_matrix(np.tile(target_test, args.ensemble), target_pred.flatten('F'))
            plt.matshow(cm, cmap='Blues')
            plt.colorbar()
            plt.ylabel('Actual')
            plt.xlabel('Predicted')
            plt.show()

        for i in range(target_pred.shape[0]):
            target_pred_comb[i] = np.bincount(target_pred[i]).argmax()
        target_pred = target_pred_comb
    else:
        M = args.eigen
        start = timer()
        target_pred, distances = test_model(M, faces_train, faces_test, target_train, target_test, args)
        end = timer()

    draw_results(args, target_test, target_pred)

if __name__ == "__main__":
    main()