import torch.nn as nn
参考翻译 What is torch.nn really?
主要是对 PyTorch 框架的各模块进行简要介绍
一定程度上是 PyTorch 的入门笔记
假设已经对神经网络相关基础知识有了一定了解
(或实现过机器学习梯度下降相关代码)
PyTorch 使用 torch.tensor
,需要将数据进行转换
import torch
x_train, y_train, x_valid, y_valid = map(
torch.tensor,
(x_train, y_train, x_valid, y_valid)
)
x_train.shape
x_train.min()
x_train.max()
map(function, iterable, …)
return iterable
import torch.nn.functional as F
包含 torch.nn
库中所有函数
同时包含大量 loss 和 activation function
import torch.nn.functional as F
loss_func = F.cross_entropy
loss = loss_func(model(x), y)
loss.backward()
其中 loss.backward()
更新模型的梯度,包括 weights 和 bias
PyTorch 中,nn 与 nn.functional 有什么区别?
nn.functional.xxx
是函数接口,nn.Xxx
是.nn.functional.xxx
的类封装,并且nn.Xxx
都继承于一个共同祖先nn.Module
nn.Xxx
除了具有nn.functional.xxx
功能之外,内部附带nn.Module
相关的属性和方法,eg.train()
,eval()
,load_state_dict
,state_dict
nn.Xxx
,实例化 -> 函数调用 -> 传入数据
inputs = torch.rand(64, 3, 28, 28)
conv = nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, padding=1)
out = conv(inputs)
nn.functional.xxx
传入数据 和 weight、bias 等其他参数
weight = torch.rand(64, 3, 3, 3)
bias = torch.rand(64)
out = nn.functional.conv2d(inputs, weight, bias, padding=1)
nn.Sequential
结合使用nn.Xxx
继承于 nn.Module
,能够很好的与 nn.Sequential
结合使用
fm_layer = nn.Sequential(
nn.Conv2d(3, 64, kernel_size=3, padding=1),
nn.BatchNorm2d(num_features=64),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2),
nn.Droput(0.2)
)
而 nn.functional.xxx
无法与 nn.Sequential
结合使用
nn.Xxx
不需要自己定义和管理weightclass CNN(nn.Module):
def __init__(self):
super(CNN, self).__init__()
self.cnn1 = nn.Conv2d(in_channels=1, out_channels=16, kernel_size=16, padding=0)
self.relu1 = nn.ReLU()
self.maxpool1 = nn.MaxPool2d(kernel_size=2)
self.cnn2 = nn.Conv2d(in_channels=16, out_channels=32, kernel_size=5, padding=0)
self.relu2 = nn.ReLU()
self.maxpool2 = nn.MaxPool2d(kernel_size=2)
self.linear1 = nn.Linear(4 * 4 * 32, 10)
def forward(self, x):
x = x.view(x.size(0), -1)
out = self.maxpool1(self.relu1(self.cnn1(x)))
out = self.maxpool2(self.relu2(self.cnn2(x)))
out = self.linear1(out.view(x.size(0), -1))
return out
nn.functional.xxx
需要自己定义 weight,每次调用的时候都需要手动传入 weight,不利于代码复用
class CNN(nn.Module):
"""docstring for CNN"""
def __init__(self):
super(CNN, self).__init__()
self.cnn1_weight = nn.Parameter(torch.rand(16, 1, 5, 5))
self.bias1_weight = nn.Parameter((torch.rand(16)))
self.cnn2_weight = nn.Parameter(torch.rand(32, 16, 5, 5))
self.bias2_weight = nn.Parameter(torch.rand(32))
self.linear1_weight = nn.Parameter(torch.rand(4 * 4 * 32, 10))
self.bias3_weight = nn.Parameter(torch.rand(10))
def forward(self, x):
x = x.view(x.size(0), -1)
out = F.conv2d(x, self.cnn1_weight, self.bias1_weight)
out = F.relu(out)
out = F.max_pool2d(out)
out = F.conv2d(out, self.cnn2_weight, self.bias2_weight)
out = F.relu(out)
out = F.max_pool2d(out)
out = F.linear(out, self.linear1_weight, self.bias3_weight)
上述两中定义方式得到的 CNN 功能都是相同的
PyTorch 官方推荐:
- 具有学习参数的(eg. conv2d, linear, batch_norm) 采用
nn.Xxx
- 没有学习参数的(eg. maxpool, loss_func, activation func) 等根据个人选择使用
nn.functional.xxx
或nn.Xxx
- 最后,关于 dropout,强烈推荐使用
nn.Xxx
方式,因为一般情况下只有训练阶段才进行 dropout,在 eval 阶段不会进行 dropout。使用nn.Xxx
方法定义 dropout,在调用model.eval()
之后,model 中所有的 dropout layer 都关闭,但以nn.functional.dropout
方式定义 dropout,在调用model.eval()
之后并不能关闭 dropout。需要使用F.dropout(x, trainig=self.training
。
继承 nn.Module
,构造一个保存 weights,bias 和具有前向传播方法(forward step)的类
nn.Module
有大量属性和方法(eg. .parameters()
和 .zero_grad()
)
torch.optim
有各种优化算法,可以使用优化器的 step
来进行前向传播,而不用人工的更新所有参数
opt.step()
opt.zero_grad()
optim.zero_grad()
将所有的梯度置为 0,需要在下个批次计算梯度之前调用
TensorDataset
是 Dataset 的 tensor 包装
from torch.utils.data import TensorDataset
train_ds = TensorDataset(x_train, y_train)
DataLoader
用于管理 batches,便于迭代
from torch.utils.data import DataLoader
train_dl = DataLoader(train_ds, batch_size=32)
迭代训练
model, opt = get_model()
for epoch in range(epochs):
for xb, yb in train_dl:
pred = model(xb)
loss = loss_func(pred, yb)
loss.backward()
opt.step()
opt.zero_grad()
print(loss_func(model(xb), yb))
在训练过程中计算并打印每个 epoch 的 validation loss
model, opt = get_model()
for epoch in range(epochs):
# 训练前
model.train()
for xb, yb in train_dl:
pred = model(xb)
loss = loss_func(pred, yb)
loss.backward()
opt.step()
opt.zero_grad()
# 训练后,验证前
# 确保 nn.BatchNorm2d 和 nn.Dropout 采取适当的行为(关闭)
model.eval()
with torch.no_grad():
valid_loss = sum(loss_func(model(xb), yb) for xb, yb in valid_dl)
print(epoch, valid_loss / len(valid_dl))
为了简化代码,增强可读性,可以构建 fit()
和 get_data()
函数
def get_data(train_ds, valid_ds, bs):
return (
DataLoader(train_ds, batch_size=bs, shuffle=True),
DataLoader(valid_ds, batch_size=bs *2)
)
def loss_batch(model, loss_func, xb, yb, opt=None):
loss = loss_func(model(xb), yb)
if opt is not None:
loss.backward()
opt.step()
opt.zero_grad()
return loss.item(), len(xb)
import numpy as np
def fit(epochs, model, loss_func, opt, train_dl, valid_dl):
for epoch in range(epochs):
model.train()
# 遍历 batch 中的每个样本
for xb, yb in train_dl:
loss_batch(model, loss_func, xb, yb, opt)
model.eval()
with torch.no_grad():
losses, nums = zip(*[loss_batch(model, loss_func, xb, yb) for xb, yb in valid_dl])
val_loss = np.sum(np.multiply(losses, nums)) / np.sum(nums)
print(epoch, val_loss)
主要代码简化为
train_dl, valid_dl = get_data(train_ds, valid_ds, bs)
model, opt = get_model()
fit(epochs, model, loss_func, opt, train_dl, valid_dl)
参考 Keras 中的 Sequential Model
model = nn.Sequential(
Lambda(preprocess),
nn.Conv2d(1, 16, kernel_size=3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(16, 16, kernel_size=3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(16, 10, kernel_size=3, stride=2, padding=1),
nn.ReLU(),
nn.AvgPool2d(4),
Lambda(lambda x: x.view(x.size(0), -1))
)
其中,可以 PyTorch 没有提供 view layer,需要构造(Sequential中的Lambda)
class Lambda(nn.Module):
def __init__(self, func):
super(Lambda, self).__init__()
self.func = func
def forward(self, x):
return self.func(x)
def preprocess(x):
return x.view(-1, 1, 28, 28)
GPU 和 CPU 训练的模型的加载不一样,参数需要设置
首先,判断 GPU 是否可以使用
print(torch.cuda.is_available())
使用指定 GPU
dev = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
将数据(batch)移到GPU(使用 .to(torch.device("cuda"))
或 .cuda()
)
xb.to(dev) # xb.cuda()
yb.to(dev) # yb.cuda()
最后,需要将模型移到 GPU
model.to(dev) # model.cuda()