def model_dnn_3():
#train:0.8596
#test:0.9308
input_data = Input((28 * 28, ))
temp_data = Dense(128)(input_data)
temp_data = Activation('relu')(temp_data)
temp_data = Dense(64)(temp_data)
temp_data = Activation('relu')(temp_data)
temp_data = Dense(48)(temp_data)
temp_data = Activation('relu')(temp_data)
temp_data = Dense(10)(temp_data)
output_data = Activation('softmax')(temp_data)
model = Model(inputs=[input_data], outputs=[output_data])
modelcheck = ModelCheckpoint('model/model.hdf5',
monitor='loss',
verbose=1,
save_best_only=True)
model.compile(loss='categorical_crossentropy',
optimizer='adadelta',
metrics=['accuracy'])
model.fit([mnist.train.images], [mnist.train.labels],
batch_size=256,
epochs=1,
callbacks=[modelcheck],
validation_data=(mnist.test.images, mnist.test.labels))
print('acc:{}'.format(accuracy_mnist(model, mnist)))
def test_bias(model_path):
model = load_model(model_path)
dic = {}
dic[5] = []
dic[10] = []
dic[20] = []
dic[30] = []
dic[50] = []
for extent in [5, 10, 20, 30, 50]:
ratio = 0.0
step = 0.1
for i in range(5):
ratio_temp = 0.0
change = False
acc = 0.98
while change | (acc > 0.9):
ratio = ratio + step
if ratio == ratio_temp:
break
_, _, _, _, model_change = model_mutation_single_neuron(
model, cls='bias', random_ratio=ratio, extent=extent)
acc = accuracy_mnist(model_change, mnist)
print(acc)
if acc <= 0.90:
if step != 0.001:
ratio_temp = ratio
step = step / 10.0
ratio = ratio - step * 10.0
change = True
print(step, ratio)
dic[extent].append((ratio, acc))
ratio = 0.0
step = 0.1
return dic
def test_del(model_path):
model = load_model(model_path)
ratio_lst = [0.05, 0.1, 0.15, 0.2, 0.3]
dic = {}
dic[0.05] = []
dic[0.1] = []
dic[0.15] = []
dic[0.2] = []
dic[0.3] = []
for ratio in ratio_lst:
for i in range(5):
model_change = random_del_neuron(model, ratio=ratio)
acc = accuracy_mnist(model_change, mnist)
print(acc, ratio)
dic[ratio].append(acc)
return dic
def bp_bias(model_path):
model = load_model(model_path)
extent_lst = [0.01, 0.1, 0.5, 1.5, 2, 3, 5, 10]
statistic = {i: [] for i in extent_lst}
ratio_lst = [0.01, 0.03, 0.05, 0.1, 0.2]
for extent in extent_lst:
for ratio in ratio_lst:
lst = []
for i in range(10):
print(i)
_, _, _, _, model_change = model_mutation_single_neuron(
model, cls='bias', random_ratio=ratio, extent=extent)
acc = accuracy_mnist(model_change, mnist)
lst.append(acc)
statistic[extent].append(lst)
return statistic
def cnn_kernel(model_path):
model = load_model(model_path)
extent_lst = [0.01, 0.1, 0.5, 1.5, 2, 3, 5, 10]
statistic = {i: [] for i in extent_lst}
ratio_lst = [0.01, 0.03, 0.05, 0.1, 0.2]
for extent in extent_lst:
for ratio in ratio_lst:
lst = []
for i in range(10):
print(i)
_, _, _, _, model_change = model_mutation_single_neuron_cnn(
model,
cls='kernel',
layers='conv',
random_ratio=ratio,
extent=extent)
acc = accuracy_mnist(model_change, mnist, cnn=True)
lst.append(acc)
print(i, acc)
statistic[extent].append(lst)
return statistic
input_data = Input((28 * 28, ))
temp_data = Dense(128)(input_data)
temp_data = Activation('relu')(temp_data)
temp_data = Dense(64)(temp_data)
temp_data = Activation('relu')(temp_data)
temp_data = Dense(48)(temp_data)
temp_data = Activation('relu')(temp_data)
temp_data = Dense(10)(temp_data)
output_data = Activation('softmax')(temp_data)
model = Model(inputs=[input_data], outputs=[output_data])
modelcheck = ModelCheckpoint('model/model.hdf5',
monitor='loss',
verbose=1,
save_best_only=True)
model.compile(loss='categorical_crossentropy',
optimizer='adadelta',
metrics=['accuracy'])
model.fit([mnist.train.images], [mnist.train.labels],
batch_size=256,
epochs=1,
callbacks=[modelcheck],
validation_data=(mnist.test.images, mnist.test.labels))
print('acc:{}'.format(accuracy_mnist(model, mnist)))
if __name__ == '__main__':
#model_dnn_3()
model_path = './model/model.hdf5'
model = load_model(model_path)
print('acc:{}'.format(accuracy_mnist(model, mnist)))