深度学习步骤:
- 1. 数据处理:读取数据 和 预处理操作
- 2. 模型设计:网络结构(假设)
- 3. 训练配置:优化器(寻解算法)
- 4. 训练过程:循环调用训练过程,包括前向计算 + 计算损失(优化目标) + 后向传播
- 5. 保存模型并测试:将训练好的模型保存
一、数据处理
import paddle
import paddle.fluid as fluid
import paddle.fluid.dygraph as dygraph
from paddle.fluid.dygraph import FC
import numpy as np
# 加载数据
def load_data():
# 从文件导入数据
datafile = 'housing.data'
data = np.fromfile(datafile, sep=' ')
# 每条数据包含14项,其中前面13项是影响因素,,第14项是相应的房屋价格中位数
feature_names = ['CRIM', 'ZN', 'INDUS', 'CHAS', 'NOX', 'RM', 'AGE', 'DIS', 'RAD', 'TAX',
'PTRATIO', 'B', 'LSTAT', 'MEDV']
feature_num = len(feature_names)
# 将原始数据进行reshape, 变成[N.14]这样的形状
data = data.reshape([data.shape[0]//feature_num, feature_num])
# 将原始数据集拆分成训练数据集和测试集
# 这里使用80%的数据集做训练,20%的数据集做测试
# 测试集和训练集必须是没有交集的
ratio = 0.8
offset = int(data.shape[0] * 0.8)
train_data = data[:offset]
# 计算train数据集的最大值,最小值,均值
maximums, minimums, avgs = train_data.max(axis=0), \
train_data.min(axis=0), train_data.sum(axis=0)/train_data.shape[0]
# 随数据进行归一化处理
for i in range(feature_num):
data[:, i] = (data[:, i] - avgs[i])/(maximums[i] - minimums[i])
# 训练集和测试集的划分
train_data = data[:offset]
test_data = data[offset:]
return train_data, test_data
二、模型设计
class Regressor(fluid.dygraph.Layer):
def __init__(self, name_scope):
super(Regressor, self).__init__(name_scope)
name_scope = self.full_name()
self.fc = FC(name_scope, size=1, act=None)
def forward(self, inputs):
x = self.fc(inputs)
return x
三、训练配置
# 定义飞桨动态图的工作环境
with fluid.dygraph.guard():
# 声明定义好的线性回归模型
model = Regressor("Regressor")
# 开启模型训练模式
model.train()
# 加载数据
training_data, test_data = load_data()
# 定义优化算法,这里使用随机梯度下降-SGD
# 学习率设置为0.01
opt = fluid.optimizer.SGD(learning_rate=0.01)
四、训练过程
两层循环,四个计算(前向计算,损失计算,反向传播,梯度计算)
with dygraph.guard():
EPOCH_NUM = 10 # 设置外层循环次数
BATCH_SIZE = 10 # 设置batch大小
# 定义外层循环
for epoch_id in range(EPOCH_NUM):
# 在每轮迭代开始之前,将训练数据的顺序随机的打乱
np.random.shuffle(training_data)
# 将训练数据进行拆分,每个batch包含10条数据
mini_batches = [training_data[k:k+BATCH_SIZE] for k in range(0, len(training_data), BATCH_SIZE)]
# 定义内层循环
for iter_id, mini_batch in enumerate(mini_batches):
x = np.array(mini_batch[:, :-1]).astype('float32') # 获得当前批次训练数据
y = np.array(mini_batch[:, -1:]).astype('float32') # 获得当前批次训练标签(真实房价)
# 将numpy数据转为飞桨动态图variable形式
house_features = dygraph.to_variable(x)
prices = dygraph.to_variable(y)
# 前向计算
predicts = model(house_features)
# 计算损失
loss = fluid.layers.square_error_cost(predicts, label=prices)
avg_loss = fluid.layers.mean(fluid.layers.sqrt(loss))
if iter_id%20==0:
print("epoch: {}, iter: {}, loss is: {}".format(epoch_id, iter_id, avg_loss.numpy()))
# 反向传播
avg_loss.backward()
# 最小化loss,更新参数
opt.minimize(avg_loss)
# 清除梯度
model.clear_gradients()
# 保存模型
fluid.save_dygraph(model.state_dict(), 'LR_model')
五、保存模型
# 定义飞桨动态图工作环境
with fluid.dygraph.guard():
# 保存模型参数,文件名为LR_model
fluid.save_dygraph(model.state_dict(), 'LR_model')
print("模型保存成功,模型参数保存在LR_model中")
六、预测
def load_one_example(data_dir):
f = open(data_dir, 'r')
datas = f.readlines()
# 选择倒数第10条数据用于测试
tmp = datas[-10]
tmp = tmp.strip().split()
one_data = [float(v) for v in tmp]
# 对数据进行归一化处理
for i in range(len(one_data)-1):
one_data[i] = (one_data[i] - avg_values[i]) / (max_values[i] - min_values[i])
data = np.reshape(np.array(one_data[:-1]), [1, -1]).astype(np.float32)
label = one_data[-1]
return data, label
with dygraph.guard():
# 参数为保存模型参数的文件地址
model_dict, _ = fluid.load_dygraph('LR_model')
model.load_dict(model_dict)
model.eval()
# 参数为数据集的文件地址
test_data, label = load_one_example('./work/housing.data')
# 将数据转为动态图的variable格式
test_data = dygraph.to_variable(test_data)
results = model(test_data)
# 对结果做反归一化处理
results = results * (max_values[-1] - min_values[-1]) + avg_values[-1]
print("Inference result is {}, the corresponding label is {}".format(results.numpy(), label))
预测结果:
七、完整程序
import paddle
import paddle.fluid as fluid
import paddle.fluid.dygraph as dygraph
from paddle.fluid.dygraph import FC
import numpy as np
# 加载数据
def load_data():
# 从文件导入数据
datafile = 'housing.data'
data = np.fromfile(datafile, sep=' ')
# 每条数据包含14项,其中前面13项是影响因素,,第14项是相应的房屋价格中位数
feature_names = ['CRIM', 'ZN', 'INDUS', 'CHAS', 'NOX', 'RM', 'AGE', 'DIS', 'RAD', 'TAX',
'PTRATIO', 'B', 'LSTAT', 'MEDV']
feature_num = len(feature_names)
# 将原始数据进行reshape, 变成[N.14]这样的形状
data = data.reshape([data.shape[0]//feature_num, feature_num])
# 将原始数据集拆分成训练数据集和测试集
# 这里使用80%的数据集做训练,20%的数据集做测试
# 测试集和训练集必须是没有交集的
ratio = 0.8
offset = int(data.shape[0] * 0.8)
train_data = data[:offset]
# 计算train数据集的最大值,最小值,均值
maximums, minimums, avgs = train_data.max(axis=0), \
train_data.min(axis=0), train_data.sum(axis=0)/train_data.shape[0]
# 随数据进行归一化处理
for i in range(feature_num):
data[:, i] = (data[:, i] - avgs[i])/(maximums[i] - minimums[i])
# 训练集和测试集的划分
train_data = data[:offset]
test_data = data[offset:]
return train_data, test_data
# 网络构建
class Regressor(fluid.dygraph.Layer):
def __init__(self, name_scope):
super(Regressor, self).__init__(name_scope)
name_scope = self.full_name()
self.fc = FC(name_scope, size=1, act=None)
def forward(self, inputs):
x = self.fc(inputs)
return x
def trainer():
use_gpu = False
place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
# 训练配置
# 定义飞浆动态图的工作环境
with fluid.dygraph.guard(place):
# 申明定义好的回归模型
model = Regressor("Regressor")
# 开启模型训练模式
model.train()
# 加载数据
train_data, test_data = load_data()
# 定义优化算法,SGD,学习率为0.01
opt = fluid.optimizer.SGD(learning_rate=0.01)
with dygraph.guard(place):
EPOCH_NUM = 10 # 设置外层循环次数
BATCH_SIZE = 10 # 设置batch尺寸
# 定义外层循环
for epoch_id in range(EPOCH_NUM):
# 在每轮迭代开始之前,将训练数据的顺序随机打乱
np.random.shuffle(train_data)
# 将训练数据集进行拆分,每个batch包含10条数据
mini_batchs = [train_data[k:k+BATCH_SIZE] for k in range(0, len(train_data), BATCH_SIZE)]
# 定义内层循环
for iter_id, mini_batch in enumerate(mini_batchs):
x = np.array(mini_batch[:, :-1]).astype('float32')
t = np.array(mini_batch[:, -1:]).astype('float32')
# 将numpy数据转为飞浆动态图variable形式
house_features = dygraph.to_variable(x)
prices = dygraph.to_variable(t)
# 前向计算
predicts = model(house_features)
# 损失计算
loss = fluid.layers.square_error_cost(predicts, label=prices)
avg_loss = fluid.layers.mean(fluid.layers.sqrt(loss))
if iter_id % 20 == 0:
print('epoch: {}, iter: {}, loss is: {}'.format(epoch_id, iter_id, avg_loss.numpy()))
# 反向传播
avg_loss.backward()
# 最小化loss,更新参数
opt.minimize(avg_loss)
# 清除梯度
model.clear_gradients()
# 保存模型
fluid.save_dygraph(model.state_dict(), 'LR_model')
print('模型保存成功,模型参数保存在LR_model中')
def load_one_example(data_dir):
f = open(data_dir, 'r')
datas = f.readlines()
# 选择倒数第10条数据用于测试
tmp = datas[-10]
tmp = tmp.strip().split()
one_data = [float(v) for v in tmp]
data = np.reshape(np.array(one_data[:-1]), [1, -1]).astype(np.float32)
label = one_data[-1]
return data, label
def predict():
model = Regressor("Regressor")
use_gpu = False
place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
with dygraph.guard(place):
# 参数为保存模型参数的文件地址
model_dict, _ = fluid.load_dygraph('LR_model')
model.load_dict(model_dict)
model.eval()
# 参数为数据集的文件地址
test_data, label = load_one_example('housing.data')
test_data = dygraph.to_variable(test_data)
results = model(test_data)
print('Inference result is {}, the corresponding label is {}'.format(results.numpy(), label))
if __name__ == "__main__":
# trainer()
predict()