# -*- coding=utf-8 -*-
import numpy as np
import keras
from keras.models import Sequential
from keras.layers import Dense,Flatten,Dropout
from keras.optimizers import Adadelta
from keras.datasets import cifar10
from keras import applications

import matplotlib.pyplot as plt
%matplotlib inline

vgg_model=applications.VGG19(include_top=False,weights='imagenet')
vgg_model.summary()

(train_x,train_y),(test_x,test_y)=cifar10.load_data()
print(train_x.shape,train_y.shape,test_x.shape,test_y.shape)

n_classes=10
train_y=keras.utils.to_categorical(train_y,n_classes)
test_y=keras.utils.to_categorical(test_y,n_classes)

bottleneck_feature_train=vgg_model.predict(train_x,verbose=1)
bottleneck_feature_test=vgg_model.predict(test_x,verbose=1)

print(bottleneck_feature_train.shape,bottleneck_feature_test.shape)

my_model=Sequential()
my_model.add(Flatten())###my_model.add(Flatten(input_shape=?))
my_model.add(Dense(512,activation='relu'))
my_model.add(Dropout(0.5))
my_model.add(Dense(256,activation='relu'))
my_model.add(Dropout(0.5))
my_model.add(Dense(n_classes,activation='softmax'))
my_model.compile(optimizer=Adadelta(),loss="categorical_crossentropy",\
metrics=['accuracy'])
my_model.fit(bottleneck_feature_train,train_y,batch_size=128,epochs=50,verbose=1)

evaluation=my_model.evaluate(bottleneck_feature_test,test_y,batch_size=128,verbose=0)
print("loss:",evaluation[0],"accuracy:",evaluation[1])

def predict_label(img_idx,show_proba=True):
plt.imshow(train_x[img_idx],aspect='auto')
plt.title("Image to be labeled")
plt.show()
img_4D=(bottleneck_feature_train[img_idx])[np.newaxis,:,:,:]
prediction=my_model.predict_classes(img_4D,batch_size=1,verbose=0)
print("Actual class:{0}\nPredict class:{1}".format(np.argmax(train_y[img_idx],0),prediction))

if show_proba:
pred=my_model.predict_proba(img_4D,batch_size=1,verbose=0)
print(pred)

for i in range(3):
predict_label(i,show_proba=True)