def run(iterations): nn = Network(10, 1, [ 10, 10, 10 ]) _input = [ 1 ] * 10 [ nn.train(_input, [1]) for i in range(0, int(iterations))]
def nn_init(sizes, batch_sizes, filters, learning_rate=0.1):
# size = (SUB_IMG_LAYERS, SUB_IMG_HEIGHT // 4, SUB_IMG_WIDTH // 4)
layers, height, width = sizes[NET_12]
first_net = Network(learning_rate=learning_rate,
input_shape=(batch_sizes[NET_12], layers, height, width),
random_state=123
)
first_filter = filters[0]
first_net.add_convolution_layer(filter_numbers=first_filter[0], filter_size=first_filter[1])
first_net.add_pooling_layer(pool_size=(2, 2))
first_net.add_dropout_layer(p=.5)
first_net.add_fully_connected_layer(hidden_layer_size=500)
first_net.add_dropout_layer(p=.5)
first_net.add_softmax_layer(unit_numbers=2)
first_net.initialize()
# SUB_IMG_LAYERS, SUB_IMG_HEIGHT // 2, SUB_IMG_WIDTH // 2
layers, height, width = sizes[NET_24]
second_net = Network(learning_rate=learning_rate,
input_shape=(batch_sizes[NET_24], layers, height, width),
random_state=123
)
second_filter = filters[1]
second_net.add_convolution_layer(filter_numbers=second_filter[0], filter_size=second_filter[1])
second_net.add_pooling_layer(pool_size=(2, 2))
second_net.add_dropout_layer(p=.5)
second_net.add_fully_connected_layer(hidden_layer_size=500)
second_net.add_dropout_layer(p=.5)
second_net.add_softmax_layer(unit_numbers=2)
second_net.initialize()
# SUB_IMG_LAYERS, SUB_IMG_HEIGHT, SUB_IMG_WIDTH
layers, height, width = sizes[NET_48]
third_net = Network(learning_rate=learning_rate,
input_shape=(batch_sizes[NET_48], layers, height, width),
random_state=123
)
third_filter = filters[2]
third_net.add_convolution_layer(filter_numbers=third_filter[0], filter_size=third_filter[1])
third_net.add_pooling_layer(pool_size=(2, 2))
third_net.add_dropout_layer(p=.5)
third_net.add_fully_connected_layer(hidden_layer_size=500)
third_net.add_dropout_layer(p=.5)
third_net.add_softmax_layer(unit_numbers=2)
third_net.initialize()
return [first_net, second_net, third_net]
import numpy as np
from layers import SigmoidLayer, LogitLayer
from nn import Network
train_images = np.load('train_images.npy')
train_labels = np.load('train_labels.npy')
test_images = np.load('test_images.npy')
test_labels = np.load('test_labels.npy')
net = Network([SigmoidLayer(28 * 28, 200), LogitLayer(200, 10)])
net.train(train_images, train_labels, test_images, test_labels, 1000, 0.1, 100)
def benchmark(data_path, act_func, save_path):
"""
测试在线性模型、单隐含层和双隐含层下的功耗模型
:param data_path: 数据路径
:param act_func: 激活函数
:param save_path: 存储路径
:return:
"""
if not os.path.exists(data_path + "/" + program + "/data"):
os.makedirs(data_path + "/" + program + "/data")
if not os.path.exists(save_path):
os.makedirs(save_path)
# mcpat.init(data_path,program)
event, power = read_data(data_path, program)
# 测试线性模型
file_name = save_path + "/linear/prediction_acc"
if os.path.exists(file_name):
os.remove(file_name)
model = Network.linear_model_train(event,
power,
path=save_path,
program_name=program)
Network.graph(model(event),
power,
program_name=program,
save_path=save_path + "/linear/" + program)
# 测试单隐含层模型
file_name = save_path + "/onelayer/prediction_acc"
if os.path.exists(file_name):
os.remove(file_name)
for node in range(4, 14):
model = Network.onelayer_model(event,
power,
layer_node=node,
act_func=act_func,
path=save_path,
program_name=program)
Network.graph(model(event),
power,
program_name=program,
save_path=save_path + "/onelayer/" + program + "_" +
str(node))
with open(save_path + "/onelayer/best_acc", "w",
encoding="utf-8") as f:
f.write("BEST_ACC=%.4f node1=%d\n" %
(Network.best_acc, Network.node1))
Network.best_acc, Network.node1 = 0, -1
# 测试双隐含层模型
file_name = save_path + "/twolayer/prediction_acc"
if os.path.exists(file_name):
os.remove(file_name)
for node1 in range(4, 14):
for node2 in range(4, 14):
model = Network.twolayer_model(event,
power,
layer1_node=node1,
layer2_node=node2,
act_func=act_func,
path=save_path,
program_name=program)
Network.graph(model(event),
power,
program_name=program,
save_path=save_path + "/twolayer/" + program + "_" +
str(node1) + "_" + str(node2))
with open(save_path + "/twolayer/best_acc", "w", encoding="utf-8") as f:
f.write("BEST_ACC=%.4f node1=%d node2=%d\n" %
(Network.best_acc, Network.node1, Network.node2))
Network.best_acc, Network.node1, Network.node2 = 0, -1, -1