def __init__(self,type='nn'):
self.num_classes = 10
self.history = None
self.epoch = 10
self.verbose = True
self.info = Info()
if type == 'nn':
self.model = Sequential()
self.model.add(Dense(128,input_shape=(784,),activation='relu'))
self.model.add(Dense(128,activation='relu'))
self.model.add(Dense(self.num_classes,activation='softmax'))
elif type == 'cnn':
self.model = Sequential()
self.model.add(Conv2D(32, (3, 3), padding='same', input_shape=(32, 32, 3),activation='relu'))
#model.add(Activation('relu'))
self.model.add(Conv2D(32,(3, 3),activation='relu'))
#model.add(Activation('relu'))
self.model.add(MaxPooling2D(pool_size=(2, 2)))
#model.add(Dropout(0.25))
self.model.add(Conv2D(64, (3, 3), padding='same',activation='relu'))
#model.add(Activation('relu'))
self.model.add(Conv2D(64, (3,3),activation='relu'))
#model.add(Activation('relu'))
self.model.add(MaxPooling2D(pool_size=(2, 2)))
#model.add(Dropout(0.25))
self.model.add(Flatten())
self.model.add(Dense(512,activation='relu'))
#model.add(Activation('relu'))
#model.add(Dropout(0.5))
self.model.add(Dense(self.num_classes,activation='softmax'))
def __init__(self):
self.dim = 64
self.num_classes = 10
self.history = None
self.epoch = 1
input_shape = (32, 32, 3)
self.info = Info()
self.model = Sequential()
self.model.add(
Conv2D(32, (3, 3),
padding='same',
input_shape=input_shape,
activation='relu'))
self.model.add(Conv2D(32, (3, 3), activation='relu'))
self.model.add(MaxPooling2D(pool_size=(2, 2)))
self.model.add(Dropout(0.25))
self.model.add(Conv2D(64, (3, 3), padding='same', activation='relu'))
self.model.add(Conv2D(64, (3, 3), activation='relu'))
self.model.add(MaxPooling2D(pool_size=(2, 2)))
self.model.add(Dropout(0.25))
self.model.add(Flatten())
self.model.add(Dense(512, activation='relu'))
self.model.add(Dropout(0.5))
self.model.add(Dense(self.num_classes, activation='softmax'))
def __init__(self):
self.dim = 64
self.num_classes = 10
self.history = None
self.epoch = 10
self.info = Info()
self.model = Sequential()
self.model.add(Dense(50, input_shape=(784, ), activation='relu'))
self.model.add(Dense(self.num_classes, activation='softmax'))
self.tbCallBack = TensorBoard(log_dir='./logs/mnist_drift/kal/',
histogram_freq=0,
write_graph=True,
write_grads=True,
write_images=True,
embeddings_freq=0,
embeddings_layer_names=None,
embeddings_metadata=None)
def __init__(self): self.dim = 64 self.num_classes = 10 self.history = None self.epoch = 10 self.info = Info() self.count = 0 self.model = Sequential() self.model.add(Dense(50,input_shape=(784,),activation='relu')) self.model.add(Dense(self.num_classes,activation='softmax')) self.y = self.model.output self.var_list = self.model.trainable_weights
def run():
pygame.init()
pygame.font.init()
set = Setting()
screen = pygame.display.set_mode((set.width, set.height))
info = Info(screen, set)
pygame.display.set_caption("TicTac")
blocks = Group()
winpage = WinPage(screen, set, info)
pausepage = PausePage(screen, set, info)
pages = {}
pages['Win'] = winpage
pages['Pause'] = pausepage
pages['Play'] = PlayPage(screen, set, info, blocks)
create_board(screen, set, blocks)
while True:
check_event(screen, set, blocks, info, pages)
update_screen(screen, set, blocks, info, pages)
class Model(object):
'''
- Initialize the model
- important parameters:
- history: it will record all the accuracy on training dataset and validation dataset
'''
def __init__(self,type='nn'):
self.num_classes = 10
self.history = None
self.epoch = 10
self.verbose = True
self.info = Info()
if type == 'nn':
self.model = Sequential()
self.model.add(Dense(128,input_shape=(784,),activation='relu'))
self.model.add(Dense(128,activation='relu'))
self.model.add(Dense(self.num_classes,activation='softmax'))
elif type == 'cnn':
self.model = Sequential()
self.model.add(Conv2D(32, (3, 3), padding='same', input_shape=(32, 32, 3),activation='relu'))
#model.add(Activation('relu'))
self.model.add(Conv2D(32,(3, 3),activation='relu'))
#model.add(Activation('relu'))
self.model.add(MaxPooling2D(pool_size=(2, 2)))
#model.add(Dropout(0.25))
self.model.add(Conv2D(64, (3, 3), padding='same',activation='relu'))
#model.add(Activation('relu'))
self.model.add(Conv2D(64, (3,3),activation='relu'))
#model.add(Activation('relu'))
self.model.add(MaxPooling2D(pool_size=(2, 2)))
#model.add(Dropout(0.25))
self.model.add(Flatten())
self.model.add(Dense(512,activation='relu'))
#model.add(Activation('relu'))
#model.add(Dropout(0.5))
self.model.add(Dense(self.num_classes,activation='softmax'))
#model.add(Activation('softmax'))
#What data is used for validation
def val_data(self,X_test,y_test):
self.X_test = X_test
self.y_test = y_test
#train the model by normal gradient descent algorithm
def fit(self,X_train,y_train):
self.model.compile(optimizer='sgd',loss='categorical_crossentropy',metrics=['accuracy'])
if self.history is None:
self.history = self.model.fit(X_train,y_train,epochs=self.epoch,batch_size=128,verbose=self.verbose,validation_data=(self.X_test,self.y_test))
else:
history = self.model.fit(X_train,y_train,epochs=self.epoch,batch_size=128,verbose=self.verbose,validation_data=(self.X_test,self.y_test))
self.history.history['acc'].extend(history.history['acc'])
self.history.history['val_acc'].extend(history.history['val_acc'])
def set_fisher_data(self,X,y):
self.info.set_fisher_data(X,y)
'''
- This function is used for 'kal' algorithm.
- The model will calculate the gradient on D1[batch0], and never access to D1 again
'''
def transfer(self,X_train,y_train,num=None):
self.info.set_train_data(X_train,y_train)
kalman_filter = Kalman_filter_modifier(info=self.info,num=num)
history = self.model.fit(X_train,y_train,epochs=self.epoch,batch_size=128,verbose=self.verbose,callbacks=[kalman_filter],validation_data=(self.X_test,self.y_test))
self.history.history['acc'].extend(history.history['acc'])
self.history.history['val_acc'].extend(history.history['val_acc'])
def save(self,name):
import json
with open('./logs/{}.txt'.format(name),'w') as f:
json.dump(self.history.history,f)
self.model.save('./models/{}.h5'.format(name))
def evaluate(self,X_test,y_test):
score=self.model.evaluate(X_test,y_test,batch_size=128)
print("Convolutional neural network test loss:",score[0])
print('Convolutional neural network test accuracy:',score[1])
return score[1]
def get_history(self):
return self.history
'''
Plot the history of accuracy
'''
def plot(self,name,model,shift=2):
plt.subplot(211)
plt.title('accuracy on current training data')
for i in range(shift):
plt.vlines(self.epoch*(i+1),0,1,color='r',linestyles='dashed')
plt.plot(self.history.history['acc'],label='{}'.format(model))
plt.ylabel('acc')
plt.xlabel('training time')
plt.legend(loc='upper right')
plt.subplot(212)
plt.title('validation accuracy on original data')
plt.plot(self.history.history['val_acc'],label='{}'.format(model))
plt.ylabel('acc')
plt.xlabel('training time')
for i in range(shift):
plt.vlines(self.epoch*(i+1),0,1,color='r',linestyles='dashed')
plt.legend(loc='upper right')
plt.subplots_adjust(wspace=1,hspace=1)
plt.savefig('./images/{}.png'.format(name))
display.display(plt.gcf())
display.clear_output(wait=True)
def enable_tensorboard(self):
self.tbCallBack = TensorBoard(log_dir='./logs/mnist_drift/kal/',
histogram_freq=0,
write_graph=True,
write_grads=True,
write_images=True,
embeddings_freq=0,
embeddings_layer_names=None,
embeddings_metadata=None)
class cnn_model(object):
'''
- Initialize the model
- important parameters:
- history: it will record all the accuracy on training dataset and validation dataset
'''
def __init__(self):
self.dim = 64
self.num_classes = 10
self.history = None
self.epoch = 10
self.info = Info()
self.model = Sequential()
self.model.add(Dense(50,input_shape=(784,),activation='relu'))
self.model.add(Dense(self.num_classes,activation='softmax'))
self.tbCallBack = TensorBoard(log_dir='./logs/mnist_drift/kal/',
histogram_freq=0,
write_graph=True,
write_grads=True,
write_images=True,
embeddings_freq=0,
embeddings_layer_names=None,
embeddings_metadata=None)
#What data is used for validation
def val_data(self,X_test,y_test):
self.X_test = X_test
self.y_test = y_test
#train the model by normal gradient descent algorithm
def fit(self,X_train,y_train):
self.model.compile(optimizer='sgd',loss='categorical_crossentropy',metrics=['accuracy'])
if self.history is None:
self.history = self.model.fit(X_train,y_train,epochs=self.epoch,batch_size=128,verbose=True,validation_data=(self.X_test,self.y_test))
else:
history = self.model.fit(X_train,y_train,epochs=self.epoch,batch_size=128,verbose=True,validation_data=(self.X_test,self.y_test))
self.history.history['acc'].extend(history.history['acc'])
self.history.history['val_acc'].extend(history.history['val_acc'])
self.info.add_data(X_train,y_train)
'''
- This function is used for 'kal' algorithm.
- The model will calculate the gradient on D1[batch0], and never access to D1 again
'''
def transfer(self,X_train,y_train,num=2):
self.info.add_data(X_train,y_train)
opcallback = op_pre_callback(info=self.info,use_pre=False)
#self.model.compile(optimizer='sgd',loss='categorical_crossentropy',metrics=['accuracy'])
history = self.model.fit(X_train,y_train,epochs=self.epoch,batch_size=128,verbose=True,callbacks=[opcallback],validation_data=(self.X_test,self.y_test))
self.history.history['acc'].extend(history.history['acc'])
self.history.history['val_acc'].extend(history.history['val_acc'])
'''
- This function is used for 'kal_pre' algorithm
- The model will access to D1 to calculate the gradient during all training process
'''
def use_pre(self,X_train,y_train):
opcallback = op_pre_callback(use_pre=True)
self.model.compile(optimizer='sgd',loss='categorical_crossentropy',metrics=['accuracy'])
history = self.model.fit(X_train,y_train,epochs=self.epoch,batch_size=128,verbose=True,callbacks=[opcallback],validation_data=(self.X_test,self.y_test))
self.history.history['acc'].extend(history.history['acc'])
self.history.history['val_acc'].extend(history.history['val_acc'])
'''
- This funciton is used for 'kal_cur' algorithm
- The model won't access to D1. All the gradients are calculated on current training data
'''
def use_cur(self,X_train,y_train):
self.info.set_value('X_train',X_train)
self.info.set_value('y_train',y_train)
opcallback = op_pre_callback(use_pre=False)
self.model.compile(optimizer='sgd',loss='categorical_crossentropy',metrics=['accuracy'])
history = self.model.fit(X_train,y_train,epochs=self.epoch,batch_size=128,verbose=True,callbacks=[opcallback],validation_data=(self.X_test,self.y_test))
self.history.history['acc'].extend(history.history['acc'])
self.history.history['val_acc'].extend(history.history['val_acc'])
'''
- A simple transer learning function
- It will pop the last two layers, and add another classifier to the end of the model
'''
def nor_trans(self,X_train,y_train):
self.model.pop()
self.model.pop()
for layer in self.model.layers:
layer.trainable = False
self.model.add(Dense(32,activation='relu'))
self.model.add(Dense(self.num_classes,activation='softmax'))
self.model.compile(optimizer='sgd',loss='categorical_crossentropy',metrics=['accuracy'])
history = self.model.fit(X_train,y_train,epochs=self.epoch,batch_size=128,verbose=True,validation_data=(self.X_test,self.y_test))
self.history.history['acc'].extend(history.history['acc'])
self.history.history['val_acc'].extend(history.history['val_acc'])
def save(self,name):
import json
with open('./logs/{}.txt'.format(name),'w') as f:
json.dump(self.history.history,f)
self.model.save('./models/{}.h5'.format(name))
def evaluate(self,X_test,y_test):
score=self.model.evaluate(X_test,y_test,batch_size=128)
print("Convolutional neural network test loss:",score[0])
print('Convolutional neural network test accuracy:',score[1])
return score[1]
def get_history(self):
return self.history
'''
Plot the history of accuracy
'''
def plot(self,name,model,shift=2):
plt.subplot(211)
plt.title('accuracy on current training data')
for i in range(shift):
plt.vlines(self.epoch*(i+1),0,1,color='r',linestyles='dashed')
plt.plot(self.history.history['acc'],label='{}'.format(model))
plt.ylabel('acc')
plt.xlabel('training time')
plt.legend(loc='upper right')
plt.subplot(212)
plt.title('validation accuracy on original data')
plt.plot(self.history.history['val_acc'],label='{}'.format(model))
plt.ylabel('acc')
plt.xlabel('training time')
for i in range(shift):
plt.vlines(self.epoch*(i+1),0,1,color='r',linestyles='dashed')
plt.legend(loc='upper right')
plt.subplots_adjust(wspace=1,hspace=1)
plt.savefig('./images/{}.png'.format(name))
class cnn_model(object):
'''
- Initialize the model
- important parameters:
- history: it will record all the accuracy on training dataset and validation dataset
'''
def __init__(self):
self.dim = 64
self.num_classes = 10
self.history = None
self.epoch = 1
input_shape = (32, 32, 3)
self.info = Info()
self.model = Sequential()
self.model.add(
Conv2D(32, (3, 3),
padding='same',
input_shape=input_shape,
activation='relu'))
self.model.add(Conv2D(32, (3, 3), activation='relu'))
self.model.add(MaxPooling2D(pool_size=(2, 2)))
self.model.add(Dropout(0.25))
self.model.add(Conv2D(64, (3, 3), padding='same', activation='relu'))
self.model.add(Conv2D(64, (3, 3), activation='relu'))
self.model.add(MaxPooling2D(pool_size=(2, 2)))
self.model.add(Dropout(0.25))
self.model.add(Flatten())
self.model.add(Dense(512, activation='relu'))
self.model.add(Dropout(0.5))
self.model.add(Dense(self.num_classes, activation='softmax'))
#What data is used for validation
def val_data(self, X_test, y_test):
self.X_test = X_test
self.y_test = y_test
#train the model by normal gradient descent algorithm
def fit(self, X_train, y_train):
self.model.compile(optimizer='sgd',
loss='categorical_crossentropy',
metrics=['accuracy'])
if self.history is None:
self.history = self.model.fit(X_train,
y_train,
epochs=self.epoch,
batch_size=128,
verbose=True,
validation_data=(self.X_test,
self.y_test))
else:
history = self.model.fit(X_train,
y_train,
epochs=self.epoch,
batch_size=128,
verbose=True,
validation_data=(self.X_test,
self.y_test))
self.history.history['acc'].extend(history.history['acc'])
self.history.history['val_acc'].extend(history.history['val_acc'])
self.info.add_data(X_train, y_train)
'''
- This function is used for 'kal' algorithm.
- The model will calculate the gradient on D1[batch0], and never access to D1 again
'''
def transfer(self, X_train, y_train, num=2):
self.info.add_data(X_train, y_train)
opcallback = op_pre_callback(info=self.info, use_pre=False)
#self.model.compile(optimizer='sgd',loss='categorical_crossentropy',metrics=['accuracy'])
history = self.model.fit(X_train,
y_train,
epochs=self.epoch,
batch_size=128,
verbose=True,
callbacks=[opcallback],
validation_data=(self.X_test, self.y_test))
self.history.history['acc'].extend(history.history['acc'])
self.history.history['val_acc'].extend(history.history['val_acc'])
def save(self, name):
import json
with open('./logs/{}.txt'.format(name), 'w') as f:
json.dump(self.history.history, f)
self.model.save('./models/{}.h5'.format(name))
def evaluate(self, X_test, y_test):
score = self.model.evaluate(X_test, y_test, batch_size=128)
print("Convolutional neural network test loss:", score[0])
print('Convolutional neural network test accuracy:', score[1])
return score[1]
def get_history(self):
return self.history
'''
Plot the history of accuracy
'''
def plot(self, name, model, shift=2):
plt.subplot(211)
plt.title('accuracy on current training data')
for i in range(shift):
plt.vlines(self.epoch * (i + 1),
0,
1,
color='r',
linestyles='dashed')
plt.plot(self.history.history['acc'], label='{}'.format(model))
plt.ylabel('acc')
plt.xlabel('training time')
plt.legend(loc='upper right')
plt.subplot(212)
plt.title('validation accuracy on original data')
plt.plot(self.history.history['val_acc'], label='{}'.format(model))
plt.ylabel('acc')
plt.xlabel('training time')
for i in range(shift):
plt.vlines(self.epoch * (i + 1),
0,
1,
color='r',
linestyles='dashed')
plt.legend(loc='upper right')
plt.subplots_adjust(wspace=1, hspace=1)
plt.savefig('./images/{}.png'.format(name))