# IMPORTING ALL THE NECESSARY LIBRARIES
from keras .models import Sequential
from keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout
from keras.utils import np_utils
from keras.callbacks import EarlyStopping

import matplotlib.pyplot as plt

# IMPORTING THE CIFAR10 DATASET BUNDLED WITH KERAS
from keras.datasets import cifar100

# SPLITTING THE DATASET INTO TRAINING AND TESTING DATASETS
(X_train, y_train), (X_test, y_test) = cifar100.load_data(label_mode='fine')

'''
SINCE WE ARE USING TENSORFLOW AS BACKEND, WE NEED TO RESHAPE THE FEATURE MATRIX AS BEING:
1) NUMBER OF SAMPLES, 2)NUMBER OF ROWS, 3)NUMBER OF COLUMNS, 4) NUMBER OF CHANNELS
'''
X_train = X_train.reshape(X_train.shape[0], 32, 32, 3)
X_test = X_test.reshape(X_test.shape[0], 32, 32, 3)

# NORMALIZING INPUTS FROM 0-255 TO 0.0-1.0
X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255
X_test /= 255

# CONVERTING LABELS TO KERAS USABLE CATEGORICAL FORMAT
y_train = np_utils.to_categorical(y_train, 100)
y_test = np_utils.to_categorical(y_test, 100)

# INITIATING THE SEQUENTIAL MODEL
classifier = Sequential()

'''
ADDING A CONVOLUTIONAL LAYER WITH:
1)NUMBER OF FILTERS TO BE USED FOR CONVOLUTION - 64
2)FILTER SIZE - (3,3)
3)INPUT SHAPE OF THE IMAGES(NUMBER OF ROWS, NUMBER OF COLUMNS, NUMBER OF CHANNELS)
4)AN ACTIVATION FUNCTION
'''
classifier.add(Conv2D(64, (3, 3), activation = 'relu', input_shape = (32, 32, 3)))
# ADDING A POOLING LAYER MATRIX OF SIZE 2X2
classifier.add(MaxPooling2D(pool_size = (2, 2)))
# DROPPING OUT RANDOM NODES TO AVOID OVERFITTING
classifier.add(Dropout(0.25))

'''
ADDING ONE MORE CONVOLUTIONAL LAYER, BUT THIS TIME THE INPUT OF THIS LAYER
WILL BE THE OUTPUT OF THE PREVIOUS LAYER
'''
classifier.add(Conv2D(64, (3, 3), activation = 'relu'))
# ADDING A POOLING LAYER MATRIX OF SIZE 2X2
classifier.add(MaxPooling2D(pool_size = (2, 2)))
# DROPPING OUT RANDOM NODES
classifier.add(Dropout(0.25))

# ADDING ONE MORE CONVOLUTIONAL LAYER WITH 64 FILTERS
classifier.add(Conv2D(128, (3, 3), activation = 'relu'))
# ADDING A POOLING LAYER MATRIX OF SIZE 2X2
classifier.add(MaxPooling2D(pool_size = (2, 2)))
# DROPPING OUT RANDOM NODES
classifier.add(Dropout(0.25))

# FLATTENING THE OUTPUT OF THE PREVIOUS LAYER TO 1D
classifier.add(Flatten())

# ADDING A FULLY CONNECTED DENSE NEURAL NETWORK WITH INPUT BEING THE FLATTEND ARRAY
classifier.add(Dense(units = 128, activation = 'relu'))
# DROPPING OUT RANDOM NODES
classifier.add(Dropout(0.25))

# ADDING A FULLY CONNECTED DENSE NEURAL NETWORK WITH INPUT AS PREVIOUS DENSE LAYER
classifier.add(Dense(units = 128, activation = 'relu'))
# DROPPING OUT RANDOM NODES
classifier.add(Dropout(0.25))

# ADDING A FULLY CONNECTED DENSE NEURAL NETWORK WITH INPUT AS PREVIOUS DENSE LAYER
classifier.add(Dense(units = 128, activation = 'relu'))
# DROPPING OUT RANDOM NODES
classifier.add(Dropout(0.25))

'''
ADDING A FULLY CONNECTED DENSE NEURAL NETWORK, BUT THIS TIME THE NETWORK WILL ONLY
CONSIST OF NODES EQUIVALENT TO NUMBER OF CATEGORIES(LABELS). NOTICE THE ACTIVATION
FUNCTION IS ALSO BEEN CHANGED TO A SIGMOID INSTEAD OF A "Rectified Linear Unit"(ReLU)
'''
classifier.add(Dense(units = 100, activation = 'softmax'))

# DECLARING NUMBER OF EPOCH AND THE BATCH SIZE
epochs = 10000
batch_size = 256
VALIDATION_PATIENCE = 20

# COMPILING THE FINAL CREATED MODEL WITH AN OPTIMIZER, LOSS FUNCTION AND EVALUATION METRICS
classifier.compile(optimizer = 'rmsprop', loss = 'categorical_crossentropy', metrics = ['accuracy'])

'''
IMPLEMENTING A CONVERGENCE MONITOR SO THE THE MODEL AUTOMATICALLY STOPS THE ITERATION
AS SOON AS THE CONVERGENCE IS ACHIEVED
'''
stopper = EarlyStopping(monitor='val_loss', patience=VALIDATION_PATIENCE)

# TRAINS THE MODEL ON TRAINING DATA BATCH-BY-BATCH
classifier.fit(X_train, y_train, batch_size=100, callbacks=[stopper], validation_data = (X_test, y_test), epochs=epochs, shuffle = True)