一、深度学习模型训练流程“八股文”
1.1 机器学习的开发应用步骤:
数据搜集——>数据预处理——>特征工程——>划分训练集和测试集——>选择模型——>建立模型(模型+超参数设置)——>设置损失函数——>进行训练以及预测
1.2 深度学习开发应用步骤与机器学习的不同
上面的机器学习应用开发步骤也同样适用于深度学习应用开发,当然也会存在一些不同,具体如下:
(1)深度学习一般所需要的数据样本很大,无法一次性的加载全部数据到模型中,需要设置参数定义每次固定送入的样本数量,通过分批次对模型进行训练;
(2)深度学习模型框架在搭建时都需要我们使用函数去实现,由上往下“逐层”进行搭建,当然也可以分开搭建,最后进行拼接组装;
(3)损失函数反向传播回网络最前面的层,同时使用优化器自动调整相关的网络参数;
(4)机器学习一般是默认在CPU上训练,当然少数除外,然而深度学习,由于计算量的过大,一般在GPU上进行训练。
二、基本配置
2.1 导包
这里导包和机器学习一样,选择需要的包进行导入,常见的有:pandas、numpy、torch、torch.nn、torch.utils.data.Dataset、torch.utils.data.DataLoader、torch.optimizer等等,如下代码所示:
import os
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
import torch.optim as optimizer
2.2 设置一些超参数
在导包完成之后,我们一般开始设置一些在下面需要用到的一些参数,这里的参数比如包含:batch size(每次模型进行训练送入的样本数)、学习率、训练次数、GPU配置、随机数等,(当然有时候也可以使用一个文件来设置参数,将参数保存到yaml文件中,方便对参数进行调整)
batch_size = 16
# 批次的大小
lr = 1e-4
# 优化器的学习率
max_epochs = 100
GPU设置的两种使用方式
# 方案一:使用os.environ,这种情况如果使用GPU不需要设置
os.environ['CUDA_VISIBLE_DEVICES'] = '0,1'
# 方案二:使用“device”,后续对要使用GPU的变量用.to(device)即可
device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
三、数据读入
3.1 Dataset类重要的三个函数
定义自己的Dataset类来实现灵活的数据读取,定义的类需要继承PyTorch自身的Dataset类,其中重要的三个函数的用法很重要,介绍如下表所示。
函数 | 介绍 |
__init__ | 用于向类中传入外部参数,同时定义样本集 |
__getitem__ | 用于逐个读取样本集合中的元素,可以进行一定的变换,并将返回训练/验证所需的数据 |
__len__ | 用于返回数据集的样本数 |
3.2 案例
class MyDataset(Dataset):
def __init__(self, data_dir, info_csv, image_list, transform=None):
"""
Args:
data_dir: path to image directory.
info_csv: path to the csv file containing image indexes
with corresponding labels.
image_list: path to the txt file contains image names to training/validation set
transform: optional transform to be applied on a sample.
"""
label_info = pd.read_csv(info_csv)
image_file = open(image_list).readlines()
self.data_dir = data_dir
self.image_file = image_file
self.label_info = label_info
self.transform = transform
def __getitem__(self, index):
"""
Args:
index: the index of item
Returns:
image and its labels
"""
image_name = self.image_file[index].strip('\n')
raw_label = self.label_info.loc[self.label_info['Image_index'] == image_name]
label = raw_label.iloc[:,0]
image_name = os.path.join(self.data_dir, image_name)
image = Image.open(image_name).convert('RGB')
if self.transform is not None:
image = self.transform(image)
return image, label
def __len__(self):
return len(self.image_file)
3.3 数据加载DataLoader
from torch.utils.data import DataLoader
train_loader = torch.utils.data.DataLoader(train_data, batch_size=batch_size, num_workers=4, shuffle=True, drop_last=True)
val_loader = torch.utils.data.DataLoader(val_data, batch_size=batch_size, num_workers=4, shuffle=False)
其中重要参数介绍
参数 | 介绍 |
batch_size | 样本是按“批”读入的,batch_size就是每次读入的样本数 |
num_workers | 有多少个进程用于读取数据 |
shuffle | 是否将读入的数据打乱 |
drop_last | 对于样本最后一部分没有达到批次数的样本,使其不再参与训练 |
四、模型构建
4.1 PyTorch中神经网络的构造方法
Module 类是 nn 模块里提供的一个模型构造类,是所有神经⽹网络模块的基类。在使用它构造自己的模型框架时,通常是对其类中的init函数和forward函数进行重载。下面以MLP作为例子:
import torch
from torch import nn
class MLP(nn.Module):
# 声明带有模型参数的层,这里声明了两个全连接层
def __init__(self, **kwargs):
# 调用MLP父类Block的构造函数来进行必要的初始化。这样在构造实例时还可以指定其他函数
super(MLP, self).__init__(**kwargs)
self.hidden = nn.Linear(784, 256)
self.act = nn.ReLU()
self.output = nn.Linear(256,10)
# 定义模型的前向计算,即如何根据输入x计算返回所需要的模型输出
def forward(self, x):
o = self.act(self.hidden(x))
return self.output(o)
init函数中定义我们需要的各个网络层以及进行参数设定
forward函数中定义正向传播
4.2 常见网络层的构造
4.2.1 不含参数的层
即对于init函数中我们不需要进行重载,只需写一个函数调用在那即可,下面以定义一个将输入减掉均值后输出的网络层·为例:
import torch
from torch import nn
class MyLayer(nn.Module):
def __init__(self, **kwargs):
super(MyLayer, self).__init__(**kwargs)
def forward(self, x):
return x - x.mean()
4.2.2 含参数的层
将参数传入Parmeter类,Parameter 类其实是 Tensor 的子类。
(1)一个是Parameter的Tensor会自动添加到模型的参数列表中;
(2)ParameterList,将需要设定的参数组合成列表放入其中;
(3)ParameterDict ,将需要设定的参数设置成字典,不同模型层的参数按照对应网络层传入。
ParameterList
class MyListDense(nn.Module):
def __init__(self):
super(MyListDense, self).__init__()
self.params = nn.ParameterList([nn.Parameter(torch.randn(4, 4)) for i in range(3)])
self.params.append(nn.Parameter(torch.randn(4, 1)))
def forward(self, x):
for i in range(len(self.params)):
x = torch.mm(x, self.params[i])
return x
net = MyListDense()
print(net)
ParameterDict
class MyDictDense(nn.Module):
def __init__(self):
super(MyDictDense, self).__init__()
self.params = nn.ParameterDict({
'linear1': nn.Parameter(torch.randn(4, 4)),
'linear2': nn.Parameter(torch.randn(4, 1))
})
self.params.update({'linear3': nn.Parameter(torch.randn(4, 2))}) # 新增
def forward(self, x, choice='linear1'):
return torch.mm(x, self.params[choice])
net = MyDictDense()
print(net)
在进行实际项目或者比赛的时候,会发现init函数中几乎定义的都是单个不同网络层的输入
4.3 常见的网络层
这里详细展示二维卷积层和池化层的底层代码,一般我们是直接调用nn里面已经写好的网络层。
4.3.1 二维卷积层
import torch
from torch import nn
# 卷积运算(二维互相关)
def corr2d(X, K):
h, w = K.shape
X, K = X.float(), K.float()
Y = torch.zeros((X.shape[0] - h + 1, X.shape[1] - w + 1))
for i in range(Y.shape[0]):
for j in range(Y.shape[1]):
Y[i, j] = (X[i: i + h, j: j + w] * K).sum()
return Y
# 二维卷积层
class Conv2D(nn.Module):
def __init__(self, kernel_size):
super(Conv2D, self).__init__()
self.weight = nn.Parameter(torch.randn(kernel_size))
self.bias = nn.Parameter(torch.randn(1))
def forward(self, x):
return corr2d(x, self.weight) + self.bias
4.3.2 池化层
import torch
from torch import nn
def pool2d(X, pool_size, mode='max'):
p_h, p_w = pool_size
Y = torch.zeros((X.shape[0] - p_h + 1, X.shape[1] - p_w + 1))
for i in range(Y.shape[0]):
for j in range(Y.shape[1]):
if mode == 'max':
Y[i, j] = X[i: i + p_h, j: j + p_w].max()
elif mode == 'avg':
Y[i, j] = X[i: i + p_h, j: j + p_w].mean()
return Y
4.4 模型示例——LeNet和AlexNet
4.4.1 LeNet
(1)模型架构
(2)代码实现
import torch
import torch.nn as nn
import torch.nn.functional as F
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
# 输入图像channel:1;输出channel:6;5x5卷积核
self.conv1 = nn.Conv2d(1, 6, 5)
self.conv2 = nn.Conv2d(6, 16, 5)
# an affine operation: y = Wx + b
self.fc1 = nn.Linear(16 * 5 * 5, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10)
def forward(self, x):
# 2x2 Max pooling
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
# 如果是方阵,则可以只使用一个数字进行定义
x = F.max_pool2d(F.relu(self.conv2(x)), 2)
x = x.view(-1, self.num_flat_features(x))
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def num_flat_features(self, x):
size = x.size()[1:] # 除去批处理维度的其他所有维度
num_features = 1
for s in size:
num_features *= s
return num_features
4.4.2 AlexNet
(1)模型架构
(2) 代码实现
class AlexNet(nn.Module):
def __init__(self):
super(AlexNet, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(1, 96, 11, 4), # in_channels, out_channels, kernel_size, stride, padding
nn.ReLU(),
nn.MaxPool2d(3, 2), # kernel_size, stride
# 减小卷积窗口,使用填充为2来使得输入与输出的高和宽一致,且增大输出通道数
nn.Conv2d(96, 256, 5, 1, 2),
nn.ReLU(),
nn.MaxPool2d(3, 2),
# 连续3个卷积层,且使用更小的卷积窗口。除了最后的卷积层外,进一步增大了输出通道数。
# 前两个卷积层后不使用池化层来减小输入的高和宽
nn.Conv2d(256, 384, 3, 1, 1),
nn.ReLU(),
nn.Conv2d(384, 384, 3, 1, 1),
nn.ReLU(),
nn.Conv2d(384, 256, 3, 1, 1),
nn.ReLU(),
nn.MaxPool2d(3, 2)
)
# 这里全连接层的输出个数比LeNet中的大数倍。使用丢弃层来缓解过拟合
self.fc = nn.Sequential(
nn.Linear(256*5*5, 4096),
nn.ReLU(),
nn.Dropout(0.5),
nn.Linear(4096, 4096),
nn.ReLU(),
nn.Dropout(0.5),
# 输出层。由于这里使用Fashion-MNIST,所以用类别数为10,而非论文中的1000
nn.Linear(4096, 10),
)
def forward(self, img):
feature = self.conv(img)
output = self.fc(feature.view(img.shape[0], -1))
return output
五、模型初始化
5.1 torch.nn.init内容
常见的nn中的初始化,主要展示每次初始化的时候应该传入的参数形式
torch.nn.init.uniform_ (tensor, a=0.0, b=1.0) |
torch.nn.init.normal_ (tensor, mean=0.0, std=1.0) |
torch.nn.init.constant_ (tensor, val) |
torch.nn.init.ones_ (tensor) |
torch.nn.init.zeros_ (tensor) |
torch.nn.init.eye_ (tensor) |
torch.nn.init.dirac_ (tensor, groups=1) |
torch.nn.init.xavier_uniform_ (tensor, gain=1.0) |
torch.nn.init.xavier_normal_ (tensor, gain=1.0) |
torch.nn.init.kaiming_uniform_ (tensor, a=0, mode='fan__in', nonlinearity='leaky_relu') |
torch.nn.init.kaiming_normal_ (tensor, a=0, mode='fan_in', nonlinearity='leaky_relu') |
torch.nn.init.orthogonal_ (tensor, gain=1) |
torch.nn.init.sparse_ (tensor, sparsity, std=0.01) |
torch.nn.init.calculate_gain (nonlinearity, param=None) |
5.2 初始化函数的封装
一般定义为一个initialize_weights()
的函数并在模型初始后进行使用
def initialize_weights(self):
for m in self.modules():
# 判断是否属于Conv2d
if isinstance(m, nn.Conv2d):
torch.nn.init.xavier_normal_(m.weight.data)
# 判断是否有偏置
if m.bias is not None:
torch.nn.init.constant_(m.bias.data,0.3)
elif isinstance(m, nn.Linear):
torch.nn.init.normal_(m.weight.data, 0.1)
if m.bias is not None:
torch.nn.init.zeros_(m.bias.data)
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zeros_()
六、损失函数
将模型训练并进行预测后得到的结果与真实结果之间的“差距”,这个“差距”是不同的,这取决于我们定义的计算方法,也就是我们常说的损失函数。
6.1 二分类交叉熵损失函数
torch.nn.BCELoss(weight=None, size_average=None, reduce=None, reduction='mean')
6.2 交叉熵损失函数
torch.nn.CrossEntropyLoss(weight=None, size_average=None, ignore_index=-100, reduce=None, reduction='mean')
6.3 L1损失函数
torch.nn.L1Loss(size_average=None, reduce=None, reduction='mean')
6.4 MSE损失函数
torch.nn.MSELoss(size_average=None, reduce=None, reduction='mean')
6.5 平滑L1 (Smooth L1)损失函数
torch.nn.SmoothL1Loss(size_average=None, reduce=None, reduction='mean', beta=1.0)
6.6 目标泊松分布的负对数似然损失
torch.nn.PoissonNLLLoss(log_input=True, full=False, size_average=None, eps=1e-08, reduce=None, reduction='mean')
6.7 KL散度
torch.nn.KLDivLoss(size_average=None, reduce=None, reduction='mean', log_target=False)
6.8 MarginRankingLoss
torch.nn.MarginRankingLoss(margin=0.0, size_average=None, reduce=None, reduction='mean')
6.9 多标签边界损失函数
torch.nn.MultiLabelMarginLoss(size_average=None, reduce=None, reduction='mean')
6.10 二分类损失函数
torch.nn.SoftMarginLoss(size_average=None, reduce=None, reduction='mean')torch.nn.(size_average=None, reduce=None, reduction='mean')
6.11 多分类的折页损失
torch.nn.MultiMarginLoss(p=1, margin=1.0, weight=None, size_average=None, reduce=None, reduction='mean')
6.12 三元组损失
torch.nn.TripletMarginLoss(margin=1.0, p=2.0, eps=1e-06, swap=False, size_average=None, reduce=None, reduction='mean')
6.13 HingEmbeddingLoss
torch.nn.HingeEmbeddingLoss(margin=1.0, size_average=None, reduce=None, reduction='mean')
6.14 余弦相似度
torch.nn.CosineEmbeddingLoss(margin=0.0, size_average=None, reduce=None, reduction='mean')
6.15 CTC损失函数
torch.nn.CTCLoss(blank=0, reduction='mean', zero_infinity=False)
七、训练和评估
训练模型存在两种状态:
(1)训练态:应该参数支持反向传播(主要是对模型参数不断更新,训练出最优的模型)
(2)验证/测试态:不应该修改模型参数(加载上面已经训练好的模型进行预测结果)
7.1 训练
def train(epoch):
# 训练状态
model.train()
# 初始化训练loss为0
train_loss = 0
# 加载训练数据集和标签
for data, label in train_loader:
# 调用cuda将计算放到GPU上
data, label = data.cuda(), label.cuda()
# 因梯度会累加上一次的,因此在每次进行训练的时候,为了放在累加,都要将优化器的梯度置零
optimizer.zero_grad()
# 开始加载模型并传入数据进行模型训练,并得到训练集的预测标签
output = model(data)
# 加载定义好的损失函数进行损失计算
loss = criterion(label, output)
# 通过反向传播传回网络
loss.backward()
# 使用优化器对参数进行更新
optimizer.step()
# 累加loss,方便最后的loss计算
train_loss += loss.item()*data.size(0)
train_loss = train_loss/len(train_loader.dataset)
print('Epoch: {} \tTraining Loss: {:.6f}'.format(epoch, train_loss))
7.2 评估
def val(epoch):
# 验证/测试状态
model.eval()
# 定义验证的loss(一般比赛我们是无法知道测试集的loss的)
val_loss = 0
# 不存入梯度计算
with torch.no_grad():
# 加载验证的数据集,并传入验证集标签(还是一样如果是测试集就只有一个训练数据,没有标签
for data, label in val_loader:
# 调用GPU计算
data, label = data.cuda(), label.cuda()
# 加载模型并传入数据,得到预测的结果
output = model(data)
# 将输出定义为我们想要的输出,增加一个输出层(这里输出为二分类)
preds = torch.argmax(output, 1)
# 计算损失
loss = criterion(output, label)
# 验证loss
val_loss += loss.item()*data.size(0)
# 统计和标签相同的
running_accu += torch.sum(preds == label.data)
val_loss = val_loss/len(val_loader.dataset)
print('Epoch: {} \tTraining Loss: {:.6f}'.format(epoch, val_loss))
八、Pytorch优化器
8.1 Pytorch提供的优化器
torch.optim.ASGD |
torch.optim.Adadelta |
torch.optim.Adagrad |
torch.optim.Adam |
torch.optim.AdamW |
torch.optim.Adamax |
torch.optim.LBFGS |
torch.optim.RMSprop |
torch.optim.Rprop |
torch.optim.SGD |
torch.optim.SparseAdam |
优化算法继承于Optimizer,其定义如下(把原文中的代码合在了一起放在Optimizer类中)
class Optimizer(object):
def __init__(self, params, defaults):
self.defaults = defaults
self.state = defaultdict(dict)
self.param_groups = []
def zero_grad(self, set_to_none: bool = False):
for group in self.param_groups:
for p in group['params']:
if p.grad is not None: #梯度不为空
if set_to_none:
p.grad = None
else:
if p.grad.grad_fn is not None:
p.grad.detach_()
else:
p.grad.requires_grad_(False)
p.grad.zero_()# 梯度设置为0
def step(self, closure):
raise NotImplementedError
def add_param_group(self, param_group):
assert isinstance(param_group, dict), "param group must be a dict"
# 检查类型是否为tensor
params = param_group['params']
if isinstance(params, torch.Tensor):
param_group['params'] = [params]
elif isinstance(params, set):
raise TypeError('optimizer parameters need to be organized in ordered collections, but '
'the ordering of tensors in sets will change between runs. Please use a list instead.')
else:
param_group['params'] = list(params)
for param in param_group['params']:
if not isinstance(param, torch.Tensor):
raise TypeError("optimizer can only optimize Tensors, "
"but one of the params is " + torch.typename(param))
if not param.is_leaf:
raise ValueError("can't optimize a non-leaf Tensor")
for name, default in self.defaults.items():
if default is required and name not in param_group:
raise ValueError("parameter group didn't specify a value of required optimization parameter " +
name)
else:
param_group.setdefault(name, default)
params = param_group['params']
if len(params) != len(set(params)):
warnings.warn("optimizer contains a parameter group with duplicate parameters; "
"in future, this will cause an error; "
"see github.com/pytorch/pytorch/issues/40967 for more information", stacklevel=3)
# 上面好像都在进行一些类的检测,报Warning和Error
param_set = set()
for group in self.param_groups:
param_set.update(set(group['params']))
if not param_set.isdisjoint(set(param_group['params'])):
raise ValueError("some parameters appear in more than one parameter group")
# 添加参数
self.param_groups.append(param_group)
def load_state_dict(self, state_dict):
r"""Loads the optimizer state.
Arguments:
state_dict (dict): optimizer state. Should be an object returned
from a call to :meth:`state_dict`.
"""
# deepcopy, to be consistent with module API
state_dict = deepcopy(state_dict)
# Validate the state_dict
groups = self.param_groups
saved_groups = state_dict['param_groups']
if len(groups) != len(saved_groups):
raise ValueError("loaded state dict has a different number of "
"parameter groups")
param_lens = (len(g['params']) for g in groups)
saved_lens = (len(g['params']) for g in saved_groups)
if any(p_len != s_len for p_len, s_len in zip(param_lens, saved_lens)):
raise ValueError("loaded state dict contains a parameter group "
"that doesn't match the size of optimizer's group")
# Update the state
id_map = {old_id: p for old_id, p in
zip(chain.from_iterable((g['params'] for g in saved_groups)),
chain.from_iterable((g['params'] for g in groups)))}
def cast(param, value):
r"""Make a deep copy of value, casting all tensors to device of param."""
# Copy state assigned to params (and cast tensors to appropriate types).
# State that is not assigned to params is copied as is (needed for
# backward compatibility).
state = defaultdict(dict)
for k, v in state_dict['state'].items():
if k in id_map:
param = id_map[k]
state[param] = cast(param, v)
else:
state[k] = v
# Update parameter groups, setting their 'params' value
def update_group(group, new_group):
...
param_groups = [
update_group(g, ng) for g, ng in zip(groups, saved_groups)]
self.__setstate__({'state': state, 'param_groups': param_groups})
def state_dict(self):
r"""Returns the state of the optimizer as a :class:`dict`.
It contains two entries:
* state - a dict holding current optimization state. Its content
differs between optimizer classes.
* param_groups - a dict containing all parameter groups
"""
# Save order indices instead of Tensors
param_mappings = {}
start_index = 0
def pack_group(group):
param_groups = [pack_group(g) for g in self.param_groups]
# Remap state to use order indices as keys
packed_state = {(param_mappings[id(k)] if isinstance(k, torch.Tensor) else k): v
for k, v in self.state.items()}
return {
'state': packed_state,
'param_groups': param_groups,
}
defaults | 存储的是优化器的超参数 |
state | 参数的缓存 |
param_groups | 管理的参数组,是一个list,其中每个元素是一个字典,顺序是params,lr,momentum,dampening,weight_decay,nesterov |
zero_grad() | 清空所管理参数的梯度,PyTorch的特性是张量的梯度不自动清零,因此每次反向传播后都需要清空梯度。 |
step() | 执行一步梯度更新,参数更新 |
add_param_group() | 添加参数组 |
load_state_dict() | 加载状态参数字典,可以用来进行模型的断点续训练,继续上次的参数进行训练 |
state_dict() | 获取优化器当前状态信息字典 |
8.2 实际操作
import os
import torch
# 设置权重,服从正态分布 --> 2 x 2
weight = torch.randn((2, 2), requires_grad=True)
# 设置梯度为全1矩阵 --> 2 x 2
weight.grad = torch.ones((2, 2))
# 输出现有的weight和data
print("The data of weight before step:\n{}".format(weight.data))
print("The grad of weight before step:\n{}".format(weight.grad))
# 实例化优化器
optimizer = torch.optim.SGD([weight], lr=0.1, momentum=0.9)
# 进行一步操作
optimizer.step()
# 查看进行一步后的值,梯度
print("The data of weight after step:\n{}".format(weight.data))
print("The grad of weight after step:\n{}".format(weight.grad))
# 权重清零
optimizer.zero_grad()
# 检验权重是否为0
print("The grad of weight after optimizer.zero_grad():\n{}".format(weight.grad))
# 输出参数
print("optimizer.params_group is \n{}".format(optimizer.param_groups))
# 查看参数位置,optimizer和weight的位置一样,我觉得这里可以参考Python是基于值管理
print("weight in optimizer:{}\nweight in weight:{}\n".format(id(optimizer.param_groups[0]['params'][0]), id(weight)))
# 添加参数:weight2
weight2 = torch.randn((3, 3), requires_grad=True)
optimizer.add_param_group({"params": weight2, 'lr': 0.0001, 'nesterov': True})
# 查看现有的参数
print("optimizer.param_groups is\n{}".format(optimizer.param_groups))
# 查看当前状态信息
opt_state_dict = optimizer.state_dict()
print("state_dict before step:\n", opt_state_dict)
# 进行5次step操作
for _ in range(50):
optimizer.step()
# 输出现有状态信息
print("state_dict after step:\n", optimizer.state_dict())
# 保存参数信息
torch.save(optimizer.state_dict(),os.path.join(r"D:\pythonProject\Attention_Unet", "optimizer_state_dict.pkl"))
print("----------done-----------")
# 加载参数信息
state_dict = torch.load(r"D:\pythonProject\Attention_Unet\optimizer_state_dict.pkl") # 需要修改为你自己的路径
optimizer.load_state_dict(state_dict)
print("load state_dict successfully\n{}".format(state_dict))
# 输出最后属性信息
print("\n{}".format(optimizer.defaults))
print("\n{}".format(optimizer.state))
print("\n{}".format(optimizer.param_groups))
8.3 输出结果
# 进行更新前的数据,梯度
The data of weight before step:
tensor([[-0.3077, -0.1808],
[-0.7462, -1.5556]])
The grad of weight before step:
tensor([[1., 1.],
[1., 1.]])
# 进行更新后的数据,梯度
The data of weight after step:
tensor([[-0.4077, -0.2808],
[-0.8462, -1.6556]])
The grad of weight after step:
tensor([[1., 1.],
[1., 1.]])
# 进行梯度清零的梯度
The grad of weight after optimizer.zero_grad():
tensor([[0., 0.],
[0., 0.]])
# 输出信息
optimizer.params_group is
[{'params': [tensor([[-0.4077, -0.2808],
[-0.8462, -1.6556]], requires_grad=True)], 'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'nesterov': False}]
# 证明了优化器的和weight的储存是在一个地方,Python基于值管理
weight in optimizer:1841923407424
weight in weight:1841923407424
# 输出参数
optimizer.param_groups is
[{'params': [tensor([[-0.4077, -0.2808],
[-0.8462, -1.6556]], requires_grad=True)], 'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'nesterov': False}, {'params': [tensor([[ 0.4539, -2.1901, -0.6662],
[ 0.6630, -1.5178, -0.8708],
[-2.0222, 1.4573, 0.8657]], requires_grad=True)], 'lr': 0.0001, 'nesterov': True, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0}]
# 进行更新前的参数查看,用state_dict
state_dict before step:
{'state': {0: {'momentum_buffer': tensor([[1., 1.],
[1., 1.]])}}, 'param_groups': [{'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'nesterov': False, 'params': [0]}, {'lr': 0.0001, 'nesterov': True, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'params': [1]}]}
# 进行更新后的参数查看,用state_dict
state_dict after step:
{'state': {0: {'momentum_buffer': tensor([[0.0052, 0.0052],
[0.0052, 0.0052]])}}, 'param_groups': [{'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'nesterov': False, 'params': [0]}, {'lr': 0.0001, 'nesterov': True, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'params': [1]}]}
# 存储信息完毕
----------done-----------
# 加载参数信息成功
load state_dict successfully
# 加载参数信息
{'state': {0: {'momentum_buffer': tensor([[0.0052, 0.0052],
[0.0052, 0.0052]])}}, 'param_groups': [{'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'nesterov': False, 'params': [0]}, {'lr': 0.0001, 'nesterov': True, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'params': [1]}]}
# defaults的属性输出
{'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'nesterov': False}
# state属性输出
defaultdict(<class 'dict'>, {tensor([[-1.3031, -1.1761],
[-1.7415, -2.5510]], requires_grad=True): {'momentum_buffer': tensor([[0.0052, 0.0052],
[0.0052, 0.0052]])}})
# param_groups属性输出
[{'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'nesterov': False, 'params': [tensor([[-1.3031, -1.1761],
[-1.7415, -2.5510]], requires_grad=True)]}, {'lr': 0.0001, 'nesterov': True, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'params': [tensor([[ 0.4539, -2.1901, -0.6662],
[ 0.6630, -1.5178, -0.8708],
[-2.0222, 1.4573, 0.8657]], requires_grad=True)]}]
8.4 注意
(1)每个优化器都是一个类,我们一定要进行实例化才能使用。
class Net(nn.Moddule):
···
net = Net()
optim = torch.optim.SGD(net.parameters(),lr=lr)
optim.step()
(2)optimizer在一个神经网络的epoch中需要实现下面两个步骤:梯度置零、梯度更新。
optimizer = torch.optim.SGD(net.parameters(), lr=1e-5)
for epoch in range(EPOCH):
...
optimizer.zero_grad() #梯度置零
loss = ... #计算loss
loss.backward() #BP反向传播
optimizer.step() #梯度更新
(3)给网络不同的层赋予不同的优化器参数。
from torch import optim
from torchvision.models import resnet18
net = resnet18()
optimizer = optim.SGD([
{'params':net.fc.parameters()},#fc的lr使用默认的1e-5
{'params':net.layer4[0].conv1.parameters(),'lr':1e-2}],lr=1e-5)
# 可以使用param_groups查看属性
九、基础实战——FashionMNIST时装分类
9.1 导包
import os
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
9.2 配置训练环境和超参数
# 配置GPU,这里有两种方式
## 方案一:使用os.environ
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
# 方案二:使用“device”,后续对要使用GPU的变量用.to(device)即可
device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
## 配置其他超参数,如batch_size, num_workers, learning rate, 以及总的epochs
batch_size = 256
num_workers = 4 # 对于Windows用户,这里应设置为0,否则会出现多线程错误
lr = 1e-4
epochs = 20
9.3 数据读入和加载
首先设置数据变换
# 首先设置数据变换
from torchvision import transforms
image_size = 28
data_transform = transforms.Compose([
transforms.ToPILImage(),
# 这一步取决于后续的数据读取方式,如果使用内置数据集读取方式则不需要
transforms.Resize(image_size),
transforms.ToTensor()
])
两种数据读入方式:
9.3.1 使用代码从网上下载
## 读取方式一:使用torchvision自带数据集,下载可能需要一段时间
from torchvision import datasets
train_data = datasets.FashionMNIST(root='./', train=True, download=True, transform=data_transform)
test_data = datasets.FashionMNIST(root='./', train=False, download=True, transform=data_transform)
9.3.2 自行从网站上进行下载,并创建相关的数据文件目录,方便读取。
## 读取方式二:读入csv格式的数据,自行构建Dataset类
# csv数据下载链接:https://www.kaggle.com/zalando-research/fashionmnist
class FMDataset(Dataset):
def __init__(self, df, transform=None):
self.df = df
self.transform = transform
self.images = df.iloc[:,1:].values.astype(np.uint8)
self.labels = df.iloc[:, 0].values
def __len__(self):
return len(self.images)
def __getitem__(self, idx):
image = self.images[idx].reshape(28,28,1)
label = int(self.labels[idx])
if self.transform is not None:
image = self.transform(image)
else:
image = torch.tensor(image/255., dtype=torch.float)
label = torch.tensor(label, dtype=torch.long)
return image, label
train_df = pd.read_csv("./FashionMNIST/fashion-mnist_train.csv")
test_df = pd.read_csv("./FashionMNIST/fashion-mnist_test.csv")
train_data = FMDataset(train_df, data_transform)
test_data = FMDataset(test_df, data_transform)
数据读入后观察数据并对数据进行可视化
import matplotlib.pyplot as plt
image, label = next(iter(train_loader))
print(image.shape, label.shape)
plt.imshow(image[0][0], cmap="gray")
9.4 模型设计
这里搭建的是一个简单的CNN模型架构
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(1, 32, 5),
nn.ReLU(),
nn.MaxPool2d(2, stride=2),
nn.Dropout(0.3),
nn.Conv2d(32, 64, 5),
nn.ReLU(),
nn.MaxPool2d(2, stride=2),
nn.Dropout(0.3)
)
self.fc = nn.Sequential(
nn.Linear(64*4*4, 512),
nn.ReLU(),
nn.Linear(512, 10)
)
def forward(self, x):
x = self.conv(x)
x = x.view(-1, 64*4*4)
x = self.fc(x)
# x = nn.functional.normalize(x)
return x
model = Net()
model = model.cuda()
# model = nn.DataParallel(model).cuda() # 多卡训练时的写法,之后的课程中会进一步讲解
9.5 设定损失函数
交叉熵损失函数为本次模型训练的损失函数
criterion = nn.CrossEntropyLoss()
注意:PyTorch会自动把整数型的label转为one-hot型,用于计算CE loss
9.6 设定优化器
使用Adam优化器
optimizer = optim.Adam(model.parameters(), lr=0.001)
9.7 训练和测试(验证)
9.7.1 训练函数
def train(epoch):
model.train()
train_loss = 0
for data, label in train_loader:
data, label = data.cuda(), label.cuda()
optimizer.zero_grad()
output = model(data)
loss = criterion(output, label)
loss.backward()
optimizer.step()
train_loss += loss.item()*data.size(0)
train_loss = train_loss/len(train_loader.dataset)
print('Epoch: {} \tTraining Loss: {:.6f}'.format(epoch, train_loss))
9.7.2 测试(验证)函数
def val(epoch):
model.eval()
val_loss = 0
gt_labels = []
pred_labels = []
with torch.no_grad():
for data, label in test_loader:
data, label = data.cuda(), label.cuda()
output = model(data)
preds = torch.argmax(output, 1)
gt_labels.append(label.cpu().data.numpy())
pred_labels.append(preds.cpu().data.numpy())
loss = criterion(output, label)
val_loss += loss.item()*data.size(0)
val_loss = val_loss/len(test_loader.dataset)
gt_labels, pred_labels = np.concatenate(gt_labels), np.concatenate(pred_labels)
acc = np.sum(gt_labels==pred_labels)/len(pred_labels)
print('Epoch: {} \tValidation Loss: {:.6f}, Accuracy: {:6f}'.format(epoch, val_loss, acc))
注意:本次案例并没有设置测试,上面是验证,一般在使用pytorch进行比赛及论文训练的时候,都会设置验证和测试,这也就是在比赛中我们常听到的线下得分和线上得分。
9.7.3 开始训练
调用上面写到训练和测试的函数,并根据上面设定的epochs进行循环设置。
for epoch in range(1, epochs+1):
train(epoch)
val(epoch)
9.8 模型保存
save_path = "./FahionModel.pkl"
torch.save(model, save_path)
深入浅出PyTorch:https://github.com/datawhalechina/thorough-pytorch