ConvGeNCode/library/convGAN.py

import os
import math
import random
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import random
from scipy import ndarray
from sklearn.neighbors import NearestNeighbors
from sklearn.decomposition import PCA
from sklearn.metrics import confusion_matrix
from sklearn.metrics import f1_score
from sklearn.metrics import cohen_kappa_score
from sklearn.metrics import precision_score
from sklearn.metrics import recall_score
from collections import Counter
from imblearn.datasets import fetch_datasets
from sklearn.preprocessing import StandardScaler

import keras
from keras.layers import Dense, Dropout, Input
from keras.models import Model,Sequential
from tqdm import tqdm
from keras.layers.advanced_activations import LeakyReLU
from tensorflow.keras.optimizers import Adam
from keras import losses
from keras import backend as K
import tensorflow as tf

import warnings
warnings.filterwarnings("ignore")

from sklearn.neighbors import KNeighborsClassifier
from sklearn.ensemble import RandomForestClassifier
from sklearn.ensemble import GradientBoostingClassifier

from numpy.random import seed
seed_num=1
seed(seed_num)
tf.random.set_seed(seed_num) 

from library.interfaces import GanBaseClass


## Import dataset
data = fetch_datasets()['yeast_me2']

## Creating label and feature matrices
labels_x=data.target ## labels of the data
labels_x.shape

features_x=data.data ## features of the data
features_x.shape

# Until now we have obtained the data. We divided it into training and test sets. we separated obtained seperate variables for the majority and miority classes and their labels for both sets.


class ConvGAN(GanBaseClass):
    """
    This is a toy example of a GAN.
    It repeats the first point of the training-data-set.
    """
    def __init__(self):
        self.isTrained = False

    def reset(self):
        """
        Resets the trained GAN to an random state.
        """
        self.isTrained = False

    def train(self, dataSet):
        """
        Trains the GAN.

        It stores the data points in the training data set and mark as trained.

        *dataSet* is a instance of /library.dataset.DataSet/. It contains the training dataset.
        We are only interested in the first *maxListSize* points in class 1.
        """
        if dataSet.data1.shape[0] <= 0:
            raise AttributeError("Train: Expected data class 1 to contain at least one point.")

        # TODO: do actually training

        self.isTrained = True

    def generateDataPoint(self):
        """
        Returns one synthetic data point by repeating the stored list.
        """
        return (self.generateData(1))[0]


    def generateData(self, numOfSamples=1):
        """
        Generates a list of synthetic data-points.

        *numOfSamples* is a integer > 0. It gives the number of new generated samples.
        """
        if not self.isTrained:
            raise ValueError("Try to generate data with untrained Re.")


        syntheticPoints = [] # TODO

        return np.array(syntheticPoints)



## convGAN
def unison_shuffled_copies(a, b,seed_perm):
    'Shuffling the feature matrix along with the labels with same order'
    np.random.seed(seed_perm)##change seed 1,2,3,4,5
    assert len(a) == len(b)
    p = np.random.permutation(len(a))
    return a[p], b[p]


def BMB(data_min,data_maj, neb, gen):
    
    ## Generate a borderline majority batch
    ## data_min -> minority class data
    ## data_maj -> majority class data
    ## neb -> oversampling neighbourhood
    ## gen -> convex combinations generated from each neighbourhood
    
    from sklearn.neighbors import NearestNeighbors
    from sklearn.utils import shuffle
    neigh = NearestNeighbors(neb)
    n_feat=data_min.shape[1]
    neigh.fit(data_maj)
    bmbi=[]
    for i in range(len(data_min)):
        indices=neigh.kneighbors([data_min[i]],neb,return_distance=False)
        bmbi.append(indices)
    bmbi=np.unique(np.array(bmbi).flatten())
    bmbi=shuffle(bmbi)
    bmb=data_maj[np.random.randint(len(data_maj),size=gen)]
    bmb=tf.convert_to_tensor(bmb)
    return bmb


def NMB_guided(data_min, neb, index):
    
    ## generate a minority neighbourhood batch for a particular minority sample
    ## we need this for minority data generation
    ## we will generate synthetic samples for each training data neighbourhood
    ## index -> index of the minority sample in a training data whose neighbourhood we want to obtain
    ## data_min -> minority class data
    ## neb -> oversampling neighbourhood
    
    from sklearn.neighbors import NearestNeighbors
    from sklearn.utils import shuffle
    neigh = NearestNeighbors(neb)
    neigh.fit(data_min)
    ind=index
    nmbi=neigh.kneighbors([data_min[ind]],neb,return_distance=False)
    nmbi=shuffle(nmbi)
    nmb=data_min[nmbi]
    nmb=tf.convert_to_tensor(nmb[0])
    return (nmb)

def conv_sample_gen():
    
    ## the generator network to generate synthetic samples from the convex space of arbitrary minority neighbourhoods
    
    min_neb_batch = keras.layers.Input(shape=(n_feat,)) ## takes minority batch as input
    x=tf.reshape(min_neb_batch, (1,neb,n_feat), name=None) ## reshaping the 2D tensor to 3D for using 1-D convolution, otherwise 1-D convolution won't work.
    x= keras.layers.Conv1D(n_feat, 3, activation='relu')(x) ## using 1-D convolution, feature dimension remains the same
    x= keras.layers.Flatten()(x) ## flatten after convolution
    x= keras.layers.Dense(neb*gen, activation='relu')(x) ## add dense layer to transform the vector to a convenient dimension
    x= keras.layers.Reshape((neb,gen))(x)## again, witching to 2-D tensor once we have the convenient shape
    s=K.sum(x,axis=1) ## row wise sum
    s_non_zero=tf.keras.layers.Lambda(lambda x: x+.000001)(s) ## adding a small constant to always ensure the row sums are non zero. if this is not done then during initialization the sum can be zero
    sinv=tf.math.reciprocal(s_non_zero) ## reprocals of the approximated row sum
    x=keras.layers.Multiply()([sinv,x]) ## At this step we ensure that row sum is 1 for every row in x. That means, each row is set of convex co-efficient
    aff=tf.transpose(x[0]) ## Now we transpose the matrix. So each column is now a set of convex coefficients
    synth=tf.matmul(aff,min_neb_batch) ## We now do matrix multiplication of the affine combinations with the original minority batch taken as input. This generates a convex transformation of the input minority batch
    model = Model(inputs=min_neb_batch, outputs=synth) ## finally we compile the generator with an arbitrary minortiy neighbourhood batch as input and a covex space transformation of the same number of samples as output
    opt = Adam(learning_rate=0.001)
    model.compile(loss='mean_squared_logarithmic_error', optimizer=opt)
    return model

def maj_min_disc():
    
    ## the discriminator is trained intwo phase:  
    ## first phase: while training GAN the discriminator learns to differentiate synthetic minority samples generated from convex minority data space against the borderline majority samples
    ## second phase: after the GAN generator learns to create synthetic samples, it can be used to generate synthetic samples to balance the dataset
    ## and then rettrain the discriminator with the balanced dataset
    
    samples=keras.layers.Input(shape=(n_feat,)) ## takes as input synthetic sample generated as input stacked upon a batch of borderline majority samples 
    y= keras.layers.Dense(250, activation='relu')(samples) ## passed through two dense layers 
    y= keras.layers.Dense(125, activation='relu')(y)
    output= keras.layers.Dense(2, activation='sigmoid')(y) ## two output nodes. outputs have to be one-hot coded (see labels variable before)
    model = Model(inputs=samples, outputs=output) ## compile model
    opt = Adam(learning_rate=0.0001)
    model.compile(loss='binary_crossentropy', optimizer=opt)
    return model


def convGAN(generator,discriminator):
    
    ## for joining the generator and the discriminator
    ## conv_coeff_generator-> generator network instance
    ## maj_min_discriminator -> discriminator network instance
    
    maj_min_disc.trainable=False ## by default the discriminator trainability is switched off. 
    ## Thus training the GAN means training the generator network as per previously trained discriminator network.
    batch_data = keras.layers.Input(shape=(n_feat,)) ## input receives a neighbourhood minority batch and a proximal majority batch concatenated
    min_batch = tf.keras.layers.Lambda(lambda x: x[:neb])(batch_data) ## extract minority batch
    maj_batch = tf.keras.layers.Lambda(lambda x: x[neb:])(batch_data) ## extract majority batch 
    conv_samples=generator(min_batch) ## pass minority batch into generator to obtain convex space transformation (synthetic samples) of the minority neighbourhood input batch
    new_samples=tf.concat([conv_samples,maj_batch],axis=0) ## concatenate the synthetic samples with the majority samples  
    output=discriminator(new_samples) ## pass the concatenated vector into the discriminator to know its decisions
    ## note that, the discriminator will not be traied but will make decisions based on its previous training while using this function
    model = Model(inputs=batch_data, outputs=output)
    opt = Adam(learning_rate=0.0001)
    model.compile(loss='mse', optimizer=opt)
    return model


## this is the main training process where the GAn learns to generate appropriate samples from the convex space
## this is the first training phase for the discriminator and the only training phase for the generator.


def rough_learning(neb_epochs,data_min,data_maj,neb,gen,generator, discriminator,GAN):

    
    step=1
    loss_history=[] ## this is for stroring the loss for every run
    min_idx=0
    neb_epoch_count=1
    
    labels=[]
    for i in range(2*gen):
        if i<gen:
            labels.append(np.array([1,0]))
        else:
            labels.append(np.array([0,1]))
    labels=np.array(labels)
    labels=tf.convert_to_tensor(labels)
    
    
    while step<(neb_epochs*len(data_min)):

        
        min_batch=NMB_guided(data_min, neb, min_idx) ## generate minority neighbourhood batch for every minority class sampls by index
        min_idx=min_idx+1 
        maj_batch=BMB(data_min,data_maj,neb,gen) ## generate random proximal majority batch 

        conv_samples=generator.predict(min_batch) ## generate synthetic samples from convex space of minority neighbourhood batch using generator
        concat_sample=tf.concat([conv_samples,maj_batch],axis=0) ## concatenate them with the majority batch

        discriminator.trainable=True ## switch on discriminator training
        discriminator.fit(x=concat_sample,y=labels,verbose=0) ## train the discriminator with the concatenated samples and the one-hot encoded labels 
        discriminator.trainable=False ## switch off the discriminator training again

        gan_loss_history=GAN.fit(concat_sample,y=labels,verbose=0) ## use the GAN to make the generator learn on the decisions made by the previous discriminator training

        loss_history.append(gan_loss_history.history['loss']) ## store the loss for the step

        if step%10 == 0:
            print(str(step)+' neighbourhood batches trained; running neighbourhood epoch ' + str(neb_epoch_count))

        if min_idx==len(data_min)-1:
            print(str('Neighbourhood epoch '+ str(neb_epoch_count) +' complete'))
            neb_epoch_count=neb_epoch_count+1
            min_idx=0


        step=step+1
    run_range=range(1,len(loss_history)+1)
    plt.rcParams["figure.figsize"] = (16,10)
    plt.xticks(fontsize=20)
    plt.yticks(fontsize=20)
    plt.xlabel('runs',fontsize=25)
    plt.ylabel('loss', fontsize=25)
    plt.title('Rough learning loss for discriminator', fontsize=25)
    plt.plot(run_range, loss_history)
    plt.show()
    return generator, discriminator, GAN, loss_history
    


def rough_learning_predictions(discriminator,test_data_numpy,test_labels_numpy):
    
    ## after the first phase of training the discriminator can be used for classification 
    ## it already learns to differentiate the convex minority points with majority points during the first training phase
    y_pred_2d=discriminator.predict(tf.convert_to_tensor(test_data_numpy))
    ## discretisation of the labels
    y_pred=np.digitize(y_pred_2d[:,0], [.5])
    ## prediction shows a model with good recall and less precision
    c=confusion_matrix(test_labels_numpy, y_pred)
    f=f1_score(test_labels_numpy, y_pred)
    pr=precision_score(test_labels_numpy, y_pred)
    rc=recall_score(test_labels_numpy, y_pred)
    k=cohen_kappa_score(test_labels_numpy, y_pred)
    print('Rough learning confusion matrix:', c)
    print('Rough learning f1 score', f)
    print('Rough learning precision score', pr)
    print('Rough learning recall score', rc)
    print('Rough learning kappa score', k)
    return c,f,pr,rc,k


def generate_data_for_min_point(data_min,neb,index,synth_num,generator):
    
    ## generate synth_num synthetic points for a particular minoity sample 
    ## synth_num -> required number of data points that can be generated from a neighbourhood
    ## data_min -> minority class data
    ## neb -> oversampling neighbourhood
    ## index -> index of the minority sample in a training data whose neighbourhood we want to obtain
    
    runs=int(synth_num/neb)+1
    synth_set=[]
    for run in range(runs):
        batch=NMB_guided(data_min, neb, index)
        synth_batch=generator.predict(batch)
        for i in range(len(synth_batch)):
            synth_set.append(synth_batch[i])
    synth_set=synth_set[:synth_num]
    synth_set=np.array(synth_set)
    return(synth_set)


def generate_synthetic_data(data_min,data_maj,neb,generator):
    
    ## roughly claculate the upper bound of the synthetic samples to be generated from each neighbourhood
    synth_num=((len(data_maj)-len(data_min))//len(data_min))+1

    ## generate synth_num synthetic samples from each minority neighbourhood
    synth_set=[]
    for i in range(len(data_min)):
        synth_i=generate_data_for_min_point(data_min,neb,i,synth_num,generator)
        for k in range(len(synth_i)):
            synth_set.append(synth_i[k])
    synth_set=synth_set[:(len(data_maj)-len(data_min))] ## extract the exact number of synthetic samples needed to exactly balance the two classes
    synth_set=np.array(synth_set)
    ovs_min_class=np.concatenate((data_min,synth_set),axis=0)
    ovs_training_dataset=np.concatenate((ovs_min_class,data_maj),axis=0)
    ovs_pca_labels=np.concatenate((np.zeros(len(data_min)),np.zeros(len(synth_set))+1,np.zeros(len(data_maj))+2))
    ovs_training_labels=np.concatenate((np.zeros(len(ovs_min_class))+1,np.zeros(len(data_maj))+0))
    ovs_training_labels_oh=[]
    for i in range(len(ovs_training_dataset)):
        if i<len(ovs_min_class):
            ovs_training_labels_oh.append(np.array([1,0]))
        else:
            ovs_training_labels_oh.append(np.array([0,1]))
    ovs_training_labels_oh=np.array(ovs_training_labels_oh)
    ovs_training_labels_oh=tf.convert_to_tensor(ovs_training_labels_oh)
    
    
    ## PCA visualization of the synthetic sata
    ## observe how the minority samples from convex space have optimal variance and avoids overlap with the majority
    pca = PCA(n_components=2)
    pca.fit(ovs_training_dataset)
    data_pca= pca.transform(ovs_training_dataset)
    
    ## plot PCA
    plt.rcParams["figure.figsize"] = (12,12)

    colors=['r', 'b', 'g']
    plt.xticks(fontsize=20)
    plt.yticks(fontsize=20)
    plt.xlabel('PCA1',fontsize=25)
    plt.ylabel('PCA2', fontsize=25)
    plt.title('PCA plot of oversampled data',fontsize=25)
    classes = ['minority', 'synthetic minority', 'majority']

    scatter=plt.scatter(data_pca[:,0], data_pca[:,1], c=ovs_pca_labels, cmap='Set1')
    plt.legend(handles=scatter.legend_elements()[0], labels=classes, fontsize=20)
    plt.show()
    
    return ovs_training_dataset, ovs_pca_labels, ovs_training_labels_oh


def final_learning(discriminator, ovs_training_dataset, ovs_training_labels_oh, test_data_numpy, test_labels_numpy, num_epochs):
    
    print('\n')
    print('Final round training of the discrminator as a majority-minority classifier')
    print('\n')
    ## second phase training of the discriminator with balanced data
    
    history_second_learning=discriminator.fit(x=ovs_training_dataset,y=ovs_training_labels_oh, batch_size=20, epochs=num_epochs)
    
    ## loss of the second phase learning smoothly decreses 
    ## this is because now the data is fixed and diverse convex combinations are no longer fed into the discriminator at every training step
    run_range=range(1,num_epochs+1)
    plt.rcParams["figure.figsize"] = (16,10)
    plt.xticks(fontsize=20)
    plt.yticks(fontsize=20)
    plt.xlabel('runs',fontsize=25)
    plt.ylabel('loss', fontsize=25)
    plt.title('Final learning loss for discriminator', fontsize=25)
    plt.plot(run_range, history_second_learning.history['loss'])
    plt.show()
    
    ## finally after second phase training the discriminator classifier has a more balanced performance
    ## meaning better F1-Score
    ## the recall decreases but the precision improves
    print('\n')

    y_pred_2d=discriminator.predict(tf.convert_to_tensor(test_data_numpy))
    y_pred=np.digitize(y_pred_2d[:,0], [.5])
    c=confusion_matrix(test_labels_numpy, y_pred)
    f=f1_score(test_labels_numpy, y_pred)
    pr=precision_score(test_labels_numpy, y_pred)
    rc=recall_score(test_labels_numpy, y_pred)
    k=cohen_kappa_score(test_labels_numpy, y_pred)
    print('Final learning confusion matrix:', c)
    print('Final learning f1 score', f)
    print('Final learning precision score', pr)
    print('Final learning recall score', rc)
    print('Final learning kappa score', k)
    return c,f,pr,rc,k


def convGAN_train_end_to_end(training_data,training_labels,test_data,test_labels, neb, gen, neb_epochs,epochs_retrain_disc):
    
    ##minority class
    data_min=training_data[np.where(training_labels == 1)[0]]
    ##majority class
    data_maj=training_data[np.where(training_labels == 0)[0]]

   
    ## instanciate generator network and visualize architecture
    conv_sample_generator=conv_sample_gen()
    print(conv_sample_generator.summary())
    print('\n')

    ## instanciate discriminator network and visualize architecture
    maj_min_discriminator=maj_min_disc()
    print(maj_min_discriminator.summary())
    print('\n')

    ## instanciate network and visualize architecture
    cg=convGAN(conv_sample_generator, maj_min_discriminator)
    print(cg.summary())
    print('\n')
    
    print('Training the GAN, first round training of the discrminator as a majority-minority classifier')
    print('\n')

    ## train gan generator ## rough_train_discriminator
    conv_sample_generator, maj_min_discriminator_r ,cg , loss_history=rough_learning(neb_epochs, data_min,data_maj, neb, gen, conv_sample_generator, maj_min_discriminator, cg)
    print('\n')
    
    ## rough learning results
    c_r,f_r,pr_r,rc_r,k_r=rough_learning_predictions(maj_min_discriminator_r, test_data,test_labels)
    print('\n')
    
    ## generate synthetic data
    ovs_training_dataset, ovs_pca_labels, ovs_training_labels_oh=generate_synthetic_data(data_min, data_maj, neb, conv_sample_generator)
    print('\n')
    
    ## final training results
    c,f,pr,rc,k=final_learning(maj_min_discriminator, ovs_training_dataset, ovs_training_labels_oh, test_data, test_labels, epochs_retrain_disc)
    
    return ((c_r,f_r,pr_r,rc_r,k_r),(c,f,pr,rc,k))


## specify parameters

neb=gen=5 ##neb=gen required
neb_epochs=10
epochs_retrain_disc=50
n_feat=len(features_x[1]) ## number of features


## Training
np.random.seed(42)
strata=5
results=[]
for seed_perm in range(strata):
    
    features_x,labels_x=unison_shuffled_copies(features_x,labels_x,seed_perm)

    ### Extracting all features and labels
    print('Extracting all features and labels for seed:'+ str(seed_perm)+'\n')
    
    ## Dividing data into training and testing datasets for 10-fold CV
    print('Dividing data into training and testing datasets for 10-fold CV for seed:'+ str(seed_perm)+'\n')
    label_1=list(np.where(labels_x == 1)[0])
    features_1=features_x[label_1]
    
    label_0=list(np.where(labels_x != 1)[0])
    features_0=features_x[label_0]
    
    a=len(features_1)//5
    b=len(features_0)//5

    fold_1_min=features_1[0:a]
    fold_1_maj=features_0[0:b]
    fold_1_tst=np.concatenate((fold_1_min,fold_1_maj))
    lab_1_tst=np.concatenate((np.zeros(len(fold_1_min))+1, np.zeros(len(fold_1_maj))))

    fold_2_min=features_1[a:2*a]
    fold_2_maj=features_0[b:2*b]
    fold_2_tst=np.concatenate((fold_2_min,fold_2_maj))
    lab_2_tst=np.concatenate((np.zeros(len(fold_1_min))+1, np.zeros(len(fold_1_maj))))

    fold_3_min=features_1[2*a:3*a]
    fold_3_maj=features_0[2*b:3*b]
    fold_3_tst=np.concatenate((fold_3_min,fold_3_maj))
    lab_3_tst=np.concatenate((np.zeros(len(fold_1_min))+1, np.zeros(len(fold_1_maj))))

    fold_4_min=features_1[3*a:4*a]
    fold_4_maj=features_0[3*b:4*b]
    fold_4_tst=np.concatenate((fold_4_min,fold_4_maj))
    lab_4_tst=np.concatenate((np.zeros(len(fold_1_min))+1, np.zeros(len(fold_1_maj))))


    fold_5_min=features_1[4*a:]
    fold_5_maj=features_0[4*b:]
    fold_5_tst=np.concatenate((fold_5_min,fold_5_maj))
    lab_5_tst=np.concatenate((np.zeros(len(fold_5_min))+1, np.zeros(len(fold_5_maj))))

    fold_1_trn=np.concatenate((fold_2_min,fold_3_min,fold_4_min,fold_5_min, fold_2_maj,fold_3_maj,fold_4_maj,fold_5_maj))

    lab_1_trn=np.concatenate((np.zeros(3*a+len(fold_5_min))+1,np.zeros(3*b+len(fold_5_maj))))

    fold_2_trn=np.concatenate((fold_1_min,fold_3_min,fold_4_min,fold_5_min,fold_1_maj,fold_3_maj,fold_4_maj,fold_5_maj))

    lab_2_trn=np.concatenate((np.zeros(3*a+len(fold_5_min))+1,np.zeros(3*b+len(fold_5_maj))))

    fold_3_trn=np.concatenate((fold_2_min,fold_1_min,fold_4_min,fold_5_min,fold_2_maj,fold_1_maj,fold_4_maj,fold_5_maj))

    lab_3_trn=np.concatenate((np.zeros(3*a+len(fold_5_min))+1,np.zeros(3*b+len(fold_5_maj))))

    fold_4_trn=np.concatenate((fold_2_min,fold_3_min,fold_1_min,fold_5_min,fold_2_maj,fold_3_maj,fold_1_maj,fold_5_maj))

    lab_4_trn=np.concatenate((np.zeros(3*a+len(fold_5_min))+1,np.zeros(3*b+len(fold_5_maj))))

    fold_5_trn=np.concatenate((fold_2_min,fold_3_min,fold_4_min,fold_1_min,fold_2_maj,fold_3_maj,fold_4_maj,fold_1_maj))

    lab_5_trn=np.concatenate((np.zeros(4*a)+1,np.zeros(4*b)))


    training_folds_feats=[fold_1_trn,fold_2_trn,fold_3_trn,fold_4_trn,fold_5_trn]

    testing_folds_feats=[fold_1_tst,fold_2_tst,fold_3_tst,fold_4_tst,fold_5_tst]

    training_folds_labels=[lab_1_trn,lab_2_trn,lab_3_trn,lab_4_trn,lab_5_trn]

    testing_folds_labels=[lab_1_tst,lab_2_tst,lab_3_tst,lab_4_tst,lab_5_tst]
    
    
    
    for i in range(5):
        
        print('\n')
        print('Executing fold: '+str(i+1))
        print('\n')
        
        r1,r2=convGAN_train_end_to_end(training_folds_feats[i],training_folds_labels[i],testing_folds_feats[i],testing_folds_labels[i], neb, gen, neb_epochs, epochs_retrain_disc)
        results.append(np.array([list(r1[1:]),list(r2[1:])]))
results=np.array(results)



## Benchmark
mean_rough=np.mean(results[:,0], axis=0)
data_r={'F1-Score_r':[mean_rough[0]], 'Precision_r' : [mean_rough[1]], 'Recall_r' : [mean_rough[2]], 'Kappa_r': [mean_rough[3]]}
df_r=pd.DataFrame(data=data_r)


print('Rough training results:')
print('\n')
print(df_r)


mean_final=np.mean(results[:,1], axis=0)
data_f={'F1-Score_f':[mean_final[0]], 'Precision_f' : [mean_final[1]], 'Recall_f' : [mean_final[2]], 'Kappa_f': [mean_final[3]]}
df_f=pd.DataFrame(data=data_f)


print('Final training results:')
print('\n')
print(df_f)