pytorch和tensorflow的爱恨情仇之定义可训练的参数
pytorch和tensorflow的爱恨情仇之基本数据类型
pytorch版本:1.6.0
tensorflow版本:1.15.0
之前我们就已经了解了pytorch和tensorflow中的变量,本节我们深入了解可训练的参数-变量
接下来我们将使用sklearn自带的iris数据集来慢慢品味。
1、pytorch
(1)第一种方式,不使用nn.Module或nn.Sequntial()来建立模型的情况下自定义参数;
加载数据集并转换为tensot:
import torch import torch.nn.functional as F import numpy as np from sklearn.datasets import load_iris iris = load_iris() data=iris.data target = iris.target
data = torch.from_numpy(data).float() #(150,4) target = torch.from_numpy(target).long() #(150,3) batch_size=data.shape[0] #设置batchsize的大小就是所有数据 dataset = torch.utils.data.TensorDataset(data, target) # 设置数据集 train_iter = torch.utils.data.DataLoader(dataset, batch_size, shuffle=True) # 设置获取数据方式
自己定义好要训练的参数:
classes = 3 input = 4 hidden = 10 w_0 = torch.tensor(np.random.normal(0, 0.01, (input, hidden)), dtype=torch.float) b_0 = torch.zeros(hidden, dtype=torch.float) w_1 = torch.tensor(np.random.normal(0, 0.01, (hidden, classes)), dtype=torch.float) b_1 = torch.zeros(classes, dtype=torch.float)
我们可以在定义参数的时候指定requires_grad=True使其为可训练的参数,也可以使用如下方式:
params = [w_0, b_0, w_1, b_1] for param in params: param.requires_grad_(requires_grad=True)
定义学习率、优化器、损失函数、网络
lr = 5 optimizer = None criterion = torch.nn.CrossEntropyLoss() epoch = 1000 def sgd(params, lr, batch_size): for param in params: param.data -= lr * param.grad / batch_size # 注意这里更改param时用的param.data def net(x): h = torch.matmul(x,w_0)+b_0 h = F.relu(h) output = torch.matmul(h,w_1)+b_1 #output = F.softmax(output,dim=1) return output
为了更加清楚参数训练的过程,这里我们不使用pytorch自带的,而是我们自己定义的随机梯度下降。
定义训练主函数:
def train(net,params,lr,train_iter): for i in range(1,epoch+1): for x,y in train_iter: output = net(x) loss = criterion(output,y) # 梯度清零 if optimizer is not None: optimizer.zero_grad() elif params is not None and params[0].grad is not None: for param in params: param.grad.data.zero_() loss.backward() if optimizer is None: sgd(params, lr, batch_size) else: optimizer.step() # “softmax回归的简洁实现”一节将用到 acc = (output.argmax(dim=1) == y).sum().item() / data.shape[0] print("epoch:{:03d} loss:{:.4f} acc:{:.4f}".format(i,loss.item(),acc)) train(net=net,params=params,lr=lr,train_iter=train_iter)
从这里我们也可以看到optimizer.zero_grad()和optimizer.step()的作用了,以上便是我们自定义训练参数的完整过程了,看下结果:
epoch:994 loss:0.0928 acc:0.9800 epoch:995 loss:0.0927 acc:0.9800 epoch:996 loss:0.0926 acc:0.9800 epoch:997 loss:0.0926 acc:0.9800 epoch:998 loss:0.0925 acc:0.9800 epoch:999 loss:0.0925 acc:0.9800 epoch:1000 loss:0.0924 acc:0.9800
(2)使用nn.Sequential()来构建模型,进行参数初始化:
导入相应的包并加载数据集:
import torch import torch.nn as nn import torch.nn.init as init import torch.nn.functional as F import numpy as np from sklearn.datasets import load_iris iris = load_iris() data=iris.data target = iris.target
转换为pytorch数据格式:
data = torch.from_numpy(data).float() target = torch.from_numpy(target).long() batch_size=data.shape[0] dataset = torch.utils.data.TensorDataset(data, target) # 设置数据集 train_iter = torch.utils.data.DataLoader(dataset, batch_size, shuffle=True) # 设置获取数据方式
定义相关超参数:
classes = 3 input = 4 hidden = 10 lr = 4 optimizer = None
定义网络:
net = nn.Sequential(
nn.Linear(input,hidden),
nn.ReLU(),
nn.Linear(hidden,classes),
)
参数初始化:
for name,param in net.named_parameters(): #使用model.named_parameters()可以获得相应层的名字的参数以及具体值 if "weight" in name: init.normal_(param, mean=0, std=0.01) if "bias" in name: init.zeros_(param)
自定义随机梯度下降优化器:
def sgd(params, lr, batch_size): for param in params: param.data -= lr * param.grad / batch_size # 注意这里更改param时用的param.data
训练主循环:
epoch = 1000 criterion = torch.nn.CrossEntropyLoss() def train(net,lr,train_iter): for i in range(1,epoch+1): for x,y in train_iter: output = net(x) loss = criterion(output,y) # 梯度清零 if optimizer is not None: optimizer.zero_grad() elif net.parameters() is not None: for param in net.parameters(): if param.grad is not None: param.grad.data.zero_() loss.backward() if optimizer is None: sgd(net.parameters(), lr, batch_size) else: optimizer.step() # “softmax回归的简洁实现”一节将用到 acc = (output.argmax(dim=1) == y).sum().item() / data.shape[0] print("epoch:{:03d} loss:{:.4f} acc:{:.4f}".format(i,loss.item(),acc)) return train(net=net,lr=lr,train_iter=train_iter)
结果:
(3) 使用pytorch自带的优化器
我们只需要将opyimizer设置为以下即可:
optimizer = torch.optim.SGD(net.parameters(), lr=0.05)
需要注意的是学习率这里需要设置的比较小一点,和上面设置的有所不同,结果如下:
(4) 使用nn.Module来构建网络,自定义参数并进行初始化
我们只需要修改以下地方即可:
class Net(nn.Module): def __init__(self,input,hidden,classes): super(Net, self).__init__() self.input = input self.hidden = hidden self.classes = classes self.w0 = nn.Parameter(torch.Tensor(self.input,self.hidden)) self.b0 = nn.Parameter(torch.Tensor(self.hidden)) self.w1 = nn.Parameter(torch.Tensor(self.hidden,self.classes)) self.b1 = nn.Parameter(torch.Tensor(self.classes)) self.reset_parameters() def reset_parameters(self): nn.init.normal_(self.w0) nn.init.constant_(self.b0,0) nn.init.normal_(self.w1) nn.init.constant_(self.b1,0) def forward(self,x): out = torch.matmul(x,self.w0)+self.b0 out = F.relu(out) out = torch.matmul(out,self.w1)+self.b1 return out net = Net(input,hidden,classes) optimizer = torch.optim.SGD(net.parameters(), lr=0.05)
结果:
(4) 使用nn.Module()构建网路,并使用各层中的参数并进行初始化
class Net(nn.Module): def __init__(self,input,hidden,classes): super(Net, self).__init__() self.input = input self.hidden = hidden self.classes = classes self.fc1 = nn.Linear(self.input,self.hidden) self.fc2 = nn.Linear(self.hidden,self.classes) for m in self.modules(): if isinstance(m, nn.Linear): nn.init.normal_(m.weight,0,0.01) nn.init.constant_(m.bias, 0) def forward(self,x): out = self.fc1(x) out = F.relu(out) out = self.fc2(out) return out net = Net(input,hidden,classes) optimizer = torch.optim.SGD(net.parameters(), lr=0.05)
结果:
PyTorch 中参数的默认初始化在各个层的 reset_parameters()
方法
我们看下官方的Linear层的实现:
官方Linear层: class Linear(Module): def __init__(self, in_features, out_features, bias=True): super(Linear, self).__init__() self.in_features = in_features self.out_features = out_features self.weight = Parameter(torch.Tensor(out_features, in_features)) if bias: self.bias = Parameter(torch.Tensor(out_features)) else: self.register_parameter('bias', None) self.reset_parameters() def reset_parameters(self): stdv = 1. / math.sqrt(self.weight.size(1)) self.weight.data.uniform_(-stdv, stdv) if self.bias is not None: self.bias.data.uniform_(-stdv, stdv) def forward(self, input): return F.linear(input, self.weight, self.bias) def extra_repr(self): return 'in_features={}, out_features={}, bias={}'.format( self.in_features, self.out_features, self.bias is not None )
(5) 最后我们来看下从网络中获取参数名字和参数值的一些例子
我们以这个网络为例:
class Net(nn.Module): def __init__(self,input,hidden,classes): super(Net, self).__init__() self.input = input self.hidden = hidden self.classes = classes self.fc1 = nn.Linear(self.input,self.hidden) self.fc2 = nn.Linear(self.hidden,self.classes) for m in self.modules(): if isinstance(m, nn.Linear): nn.init.normal_(m.weight,0,0.01) nn.init.constant_(m.bias, 0) def forward(self,x): out = self.fc1(x) out = F.relu(out) out = self.fc2(out) return out net = Net(input,hidden,classes)
首先是model.state_dict():是一个参数字典,键是参数的名称,值是参数的值:
for name,value in net.state_dict().items(): print(name,value)
接着是:model.parameters():返回的是一个generator,我们之前也经常使用,通过param.data,param.data.grad来获取参数的值以及梯度
for param in net.parameters(): print(param.data,param.grad)
接着是model.named_parameters():返回的是一个具名参数,也就是包含了参数的名称
for name,param in net.named_parameters(): print(name,param)
最后讲下的是self.modules():一般是在网络初始化中使用,返回的是网络中的具体层,我们可以通过其对不同层进行参数初始化,比如nn.Conv2d、nn.Linear等;
参考:
https://www.cnblogs.com/KaifengGuan/p/12332072.html
https://www.geekschool.org/2020/08/02/13455.html
https://blog.csdn.net/weixin_44058333/article/details/92691656
(2)tensorflow
导入相应的包并加载数据:
import tensorflow as tf import numpy as np from sklearn.datasets import load_iris from sklearn.preprocessing import OneHotEncoder iris = load_iris() data=iris.data target = iris.target
将标签转换为onehot编码:
oneHotEncoder = OneHotEncoder(sparse=False) onehot_target = oneHotEncoder.fit_transform(target.reshape(-1,1)) print(onehot_target)
定义超参数以及可训练的参数:
input=4 hidden=10 classes=3 w0=tf.Variable(tf.random.normal([input,hidden],stddev=0.01,seed=1)) b0=tf.Variable(tf.zeros([hidden])) w1=tf.Variable(tf.random.normal([hidden,classes],stddev=0.01,seed=1)) b1=tf.Variable(tf.zeros([classes]))
定义计算图中的占位符:
x = tf.placeholder(tf.float32,shape=(None,input),name="x-input") #输入数据 y_ = tf.placeholder(tf.float32,shape=(None,classes),name="y-input") #真实标签
定义网络、损失函数和优化器:
def net(x): hid = tf.add(tf.matmul(x,w0),b0) hid = tf.nn.relu(hid) out = tf.add(tf.matmul(hid,w1),b1) out = tf.nn.softmax(out) return out y = net(x) cross_entropy = -tf.reduce_mean(y_*tf.log(tf.clip_by_value(y,1e-10,1.0)) \ + (1-y_)*tf.log(tf.clip_by_value(1-y,1e-10,1.0))) optimizer=tf.compat.v1.train.GradientDescentOptimizer(learning_rate=0.05).minimize(cross_entropy)
训练循环:
epoch = 1000 with tf.compat.v1.Session() as sess: #建立会话 init_op = tf.global_variables_initializer() #初始化参数 sess.run(init_op) for epoch in range(1,epoch+1): sess.run(optimizer,feed_dict={x:data,y_:onehot_target}) #传入数据给优化器 y_pred = sess.run(y,feed_dict={x:data}) #计算输出 total_cross_entropy = sess.run(cross_entropy,feed_dict={y:y_pred,y_:onehot_target}) #计算交叉熵 pred = tf.argmax(y_pred,axis = 1) # 取出行中最大值的索引,也就是取出其中概率最大的索引 correct = tf.cast(tf.equal(pred,target),dtype=tf.int32) # 判断与测试集的标签是否相等并且转换bool为int型 correct = tf.reduce_sum(correct) # 沿着指定维度的和,不指定axis则默认为所有元素的和 acc = correct.eval() / data.shape[0] print("epoch:{} loss:{:.4f} acc:{:.4f}".format(epoch, total_cross_entropy,acc))
结果:
但感觉训练1000个epoch比pytorch慢好多。。