ConvGeNCode/library/generators/autoencoder.py

import numpy as np

from library.interfaces import GanBaseClass
from library.dataset import DataSet

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 sklearn.neighbors import NearestNeighbors
from sklearn.utils import shuffle
from imblearn.datasets import fetch_datasets

from keras.layers import Dense, Input, Multiply, Flatten, Conv1D, Reshape
from keras.models import Model
from keras import backend as K
from tqdm import tqdm

import tensorflow as tf
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.layers import Lambda

from library.NNSearch import NNSearch

import warnings
warnings.filterwarnings("ignore")


lossFunction = "mean_squared_logarithmic_error"
#lossFunction = "mse"


def newDense(size, activation="relu"):  # softsign
    initializer = tf.keras.initializers.RandomUniform(minval=0.00001, maxval=float(size))
    initializer = "glorot_uniform"

    return Dense(int(size)
        , activation=activation
        , kernel_initializer=initializer
        , bias_initializer=initializer
        )


class Autoencoder(GanBaseClass):
    """
    This is a toy example of a GAN.
    It repeats the first point of the training-data-set.
    """
    def __init__(self, n_feat, middleSize=4, eps=0.0001, debug=True):
        self.isTrained = False
        self.n_feat = n_feat
        self.middleSize = middleSize
        self.eps = eps
        self.debug = debug
        self.dataSet = None
        self.decoder = None
        self.encoder = None
        self.autoencoder = None
        self.cg = None
        self.scaler = 1.0
        self.lossFn = lossFunction #"mse"
        self.lossFn = "mean_squared_logarithmic_error"

    def reset(self):
        """
        Resets the trained GAN to an random state.
        """
        self.isTrained = False
        self.scaler = 1.0
        ## instanciate discriminator network and visualize architecture
        self.encoder = self._createEncoder()

        ## instanciate generator network and visualize architecture
        self.decoder = self._createDecoder()

        ## instanciate network and visualize architecture
        self.autoencoder = self._createAutoencoder(self.encoder, self.decoder)

    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.")

        d = dataSet.data1
        self.data1 = d
        self.scaler = 1.5 * tf.reduce_max(tf.abs(d)).numpy()
        scaleDown = 1.0 / self.scaler

        lastLoss = 0.0
        print(f"scaler: {self.scaler}")

        dScaled = scaleDown * d
        
        for epoch in range(1000):
            h = self.autoencoder.fit(d, dScaled, epochs=1, shuffle=True)
            #print(str(d[0]) + " →")
            #print(self.scaler * self.autoencoder.predict(np.array([d[0]])))
            loss = h.history["loss"][-1]
            if loss < self.eps:
                print(f"done in {epoch} rounds")
                break

            if epoch == 0:
                lastLoss = loss
            else:
                print(f"Loss: {lastLoss} → {loss}")
                if abs(lastLoss - loss) < (0.1 * self.eps) and epoch > 10:
                    print(f"converged in {epoch} rounds")
                    break
                else:
                    lastLoss = loss


        code = self.encoder.predict(d)
        center = np.zeros(self.middleSize)
        for c in code:
            center = center + c
        center = (1.0 / float(d.shape[0])) * center

        d = 0.0
        for c in code:
            d = max(d, tf.reduce_max(tf.abs(c - center)).numpy())

        self.noise = (center, d)

        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.")

        noise = self.noise[0] + np.random.normal(0.0, self.noise[1], [numOfSamples, self.middleSize])
        syntheticPoints = self.decoder.predict(noise)
        
        # syntheticPoints = []
        # while len(syntheticPoints) < numOfSamples:
        #     nRest = max(0, numOfSamples - len(syntheticPoints))
        #     nBatch = min(nRest, len(self.data1))
        #     syntheticPoints.extend(self.autoencoder.predict(self.data1[:nBatch]))

        return self.scaler * np.array(syntheticPoints)

    # ###############################################################
    # Hidden internal functions
    # ###############################################################

    # Creating the GAN
    def _createEncoder(self):
        """
        the generator network to generate synthetic samples from the convex space
        of arbitrary minority neighbourhoods
        """

        ## takes minority batch as input
        dataIn = Input(shape=(self.n_feat,))
        x = dataIn
        x = newDense(self.n_feat)(x)

        ## 
        n = self.n_feat // 2
        x = newDense(max(n, self.middleSize))(x)

        x = newDense(self.middleSize)(x)

        model = Model(inputs=dataIn, outputs=x)
        opt = Adam(learning_rate=0.01)
        model.compile(loss=lossFunction, optimizer=opt)

        print("encoder")
        model.summary()
        return model

    def _createDecoder(self):
        """
        the generator network to generate synthetic samples from the convex space
        of arbitrary minority neighbourhoods
        """

        ## takes minority batch as input
        dataIn = Input(shape=(self.middleSize,))
        x = dataIn

        ## 
        n = self.n_feat // 2
        #x = newDense(max(n, self.middleSize))(x)

        #x = newDense(self.n_feat)(x)
        x = newDense(self.n_feat)(x)

        model = Model(inputs=dataIn, outputs=x)
        opt = Adam(learning_rate=0.01)
        model.compile(loss=lossFunction, optimizer=opt)

        print("decoder")
        model.summary()
        return model


    def _createAutoencoder(self, encoder, decoder):
        """
        for joining the generator and the discriminator
        conv_coeff_generator-> generator network instance
        maj_min_discriminator -> discriminator network instance
        """

        #encoder.trainable = False
        ## input receives a neighbourhood minority batch
        ## and a proximal majority batch concatenated
        dataIn = Input(shape=(self.n_feat,))
        #x = newDense(self.middleSize)(dataIn)
        #x = newDense(self.n_feat)(x)
        #x = newDense(self.n_feat)(x)
        
        x = encoder(dataIn )
        x = decoder(x)

        ## note that, the discriminator will not be traied but will make decisions based
        ## on its previous training while using this function
        model = Model(inputs=dataIn, outputs=x)
        opt = Adam(learning_rate=0.01)
        model.compile(loss=self.lossFn, optimizer=opt)

        print("autoencoder")
        model.summary()
        return model