|
|
楼主 |
发表于 2024-9-3 09:51:56
|
显示全部楼层
对于一下模型的保存和加载的功能，我要怎么补充：
class NN():
def __init__(self, layers, data):
L = len(layers) # 这个layers是个包含了输入层到输出层的各层神经元个数的列表
# 设置神经元权重参数和偏置参数
self.W = {}
self.b = {}
for l in range(1, L):
# print(layers[l])
self.W["W" + str(l)] = np.mat(np.random.randn(layers[l], layers[l - 1]) * 0.1)
self.b["b" + str(l)] = np.mat(np.random.randn(layers[l], 1) * 0.1)
self.A = {} # 这个字典里存储了输入，和后面各个层的激活值；输出层的激活值在这个字典的最后一个
self.Z = {} # 这个字典里存储了从第二层到输出层的没被激活的值
self.cache = {} # 这个字典里存储了https://www.cnblogs.com/pinard/p/6422831.html里面各个层δl的值，第一个是输出层的δ
# 将数据写入类中
self.data = data
def forward_activation_02(self, L, flag): # 这个L是和上面的L一样的，包含了总共层数
# 初始化输入
self.A["A0"] = self.inputs
for l in range(1, L):
if flag == 0 or l == 1 or l == L - 1:
self.Z["Z" + str(l)] = self.W["W" + str(l)] * self.A["A" + str(l - 1)] + self.b["b" + str(l)]
self.A["A" + str(l)] = sigmoid(self.Z["Z" + str(l)])
else:
# 启用dropout正则化
self.d = np.random.rand(self.A["A" + str(l - 1)].shape[0], self.A["A" + str(l - 1)].shape[1])
self.d = self.d < keep_prob
self.A["A" + str(l - 1)] = np.multiply(self.A["A" + str(l - 1)], self.d)
self.A["A" + str(l - 1)] /= keep_prob
self.Z["Z" + str(l)] = self.W["W" + str(l)] * self.A["A" + str(l - 1)] + self.b["b" + str(l)]
self.A["A" + str(l)] = sigmoid(self.Z["Z" + str(l)])
# 更新cache
for l in reversed(range(1, L)):
if l == L - 1:
self.cache["C" + str(l)] = np.multiply(self.A["A" + str(l)] - self.output,
dsigmoid(self.Z["Z" + str(l)]))
else:
self.cache["C" + str(l)] = np.multiply(self.W["W" + str(l + 1)].T * self.cache["C" + str(l + 1)],
dsigmoid(self.Z["Z" + str(l)]))
err = np.abs(self.A["A" + str(L - 1)] - self.output)
return err
def backPropagate_02(self, learning_rate, L): # 反向传播
alpha = learning_rate
m = self.inputs.shape[1] # 这是样本数
for i in range(L):
l = L - i - 1
if l > 0:
self.b['b' + str(l)] = self.b['b' + str(l)] - alpha * 1.0 / m * (
self.cache["C" + str(l)] * np.ones((m, 1)))
self.W['W' + str(l)] = self.W['W' + str(l)] - alpha * 1.0 / m * (
self.cache["C" + str(l)] * self.A["A" + str(l - 1)].T)
def init_prameter(self, batch_size):
# 每次批量训练batch_size幅图像
train_dataloader = DataLoader(self.data, batch_size=batch_size, drop_last=False, shuffle=True)
for train_data in train_dataloader:
imgs = train_data.numpy()
batch_xs, batch_ys = imgs[:, :feature_num], imgs[:, feature_num]
self.inputs = np.mat(batch_xs).transpose()
self.output = np.mat(sigmoid(batch_ys))
def train(self, iterations, learning_rate, L, batch_size, train_shape):
# 批处理训练
for i in range(iterations):
if (train_shape % batch_size) != 0:
print("batch没选对，要能够被train除进")
for j in range(int(train_shape / batch_size)):
self.init_prameter(batch_size)
err = self.forward_activation_02(L, 1)
self.backPropagate_02(learning_rate, L)
#############################每训练一轮就在测试集上进行测试
asdf, aswe = test_data[:, :feature_num], test_data[:, feature_num]
self.inputs = np.mat(asdf).transpose()
self.output = np.mat(sigmoid(aswe))
test_num = self.output.shape[1]
test_loss = self.forward_activation_02(L, 0)
print("在训练集上，第", i, "轮迭代误差为", np.sum(err) / batch_size, "在测试集上，第", i, "轮迭代误差为",
np.sum(test_loss) / test_num)
if np.sum(test_loss) / test_num < 0.02 or i == (iterations - 1):
true_ = self.output.transpose()
pred_ = self.A["A" + str(L - 1)].transpose()
print("测试样本的实际结果：", true_)
print("测试样本的预测结果：", pred_)
plt.plot(pred_, label="pred") # 绘制预测关系图
plt.legend()
plt.plot(true_, label="true") # 绘制预测关系图
plt.legend()
plt.show()
break
return true_, pred_ |
|
[复制链接]
( 粤ICP备18085999号-1 | 粤公网安备 44051102000585号)
狗仔卡
显身卡
发表于 2024-9-3 09:51:59