from tensorflow.keras.datasets import mnist
import numpy as np
from tensorflow.keras.layers import Dropout
from sklearn.decomposition import PCA
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
from tensorflow.keras.utils import to_categorical
n_c_list = [154, 331, 486, 713]
pca_list = [0.95, 0.99, 0.999, 1.0]

(x_train, y_train), (x_test, y_test) = mnist.load_data()

x = np.append(x_train, x_test, axis=0)
y = np.append(y_train, y_test, axis=0)
y = to_categorical(y)
x = x.reshape(x.shape[0], -1)

for i in range(len(n_c_list)):
    pca = PCA(n_components=n_c_list[i])
    x_p = pca.fit_transform(x.astype('float32'))
    x_train, x_test, y_train, y_test = train_test_split(x_p, y, train_size=0.8, shuffle=True, random_state=123)

    model =Sequential()
    model.add(Dense(64, input_shape=(28*28,)))
    model.add(Dropout(0.5))
    model.add(Dense(32,activation='relu'))
    model.add(Dense(16,activation='relu'))
    model.add(Dense(10,activation='softmax'))
    
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics='acc')
    model.fit(x_train, y_train, epochs=20, batch_size=16, validation_split=0.2, verbose=0)
    
    acc = model.evaluate(x_test, y_test)[1]
    print(f'PCA {pca_list[i]} test acc : {acc}')
    
    y_pred = np.argmax(model.predict(x_test,verbose=0), axis=1)
    print(f'PCA {pca_list[i]} pred acc :', accuracy_score(np.argmax(y_test, axis=1), y_pred))
    

# CNN acc : 0.9881    
# DNN acc : 0.9685
# PCA 0.95 test acc : 0.8454285860061646
# PCA 0.95 pred acc : 0.8454285714285714
# PCA 0.99 test acc : 0.8495714068412781
# PCA 0.99 pred acc : 0.8495714285714285
# PCA 0.999 test acc : 0.8355714082717896
# PCA 0.999 pred acc : 0.8355714285714285
# PCA 1.0 test acc : 0.818142831325531
# PCA 1.0 pred acc : 0.8181428571428572