1. 微调¶
2. 总结¶
1. 微调¶
In [1]:
%matplotlib inline
import os
import torch
import torchvision
from torch import nn
from d2l import torch as d2l
In [2]:
d2l.DATA_HUB['hotdog'] = (d2l.DATA_URL + 'hotdog.zip','fba480ffa8aa7e0febbb511d181409f899b9baa5')
data_dir = d2l.download_extract('hotdog')
train_imgs = torchvision.datasets.ImageFolder(os.path.join(data_dir,'train'))
test_imgs = torchvision.datasets.ImageFolder(os.path.join(data_dir,'test'))
In [3]:
# 图片的大小和纵横比各有不同
hotdogs = [train_imgs[i][0] for i in range(8)]
print(train_imgs[0]) # 图片和标签，合为一个元组
print(train_imgs[0][0]) # 元组第一个元素为图片
not_hotdogs = [train_imgs[-i - 1][0] for i in range(8)]
d2l.show_images(hotdogs + not_hotdogs, 2, 8, scale=1.4)
(<PIL.Image.Image image mode=RGB size=122x144 at 0x1F2CBDF9AC8>, 0) <PIL.Image.Image image mode=RGB size=122x144 at 0x1F2CBDF9C18>
Out[3]:
array([<AxesSubplot:>, <AxesSubplot:>, <AxesSubplot:>, <AxesSubplot:>,
<AxesSubplot:>, <AxesSubplot:>, <AxesSubplot:>, <AxesSubplot:>,
<AxesSubplot:>, <AxesSubplot:>, <AxesSubplot:>, <AxesSubplot:>,
<AxesSubplot:>, <AxesSubplot:>, <AxesSubplot:>, <AxesSubplot:>],
dtype=object)
In [4]:
# 数据增广
normalize = torchvision.transforms.Normalize([0.485,0.456,0.406],
[0.229,0.224,0.225]) # 按该均值、方差做归一化
train_augs = torchvision.transforms.Compose([
torchvision.transforms.RandomResizedCrop(224),
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.ToTensor(),
normalize ])
test_augs = torchvision.transforms.Compose([
torchvision.transforms.Resize(256),
torchvision.transforms.CenterCrop(224),
torchvision.transforms.ToTensor(),
normalize ])
In [5]:
# 定义和初始化模型
pretrained_net = torchvision.models.resnet18(pretrained=True) # 把模型和在ImageNet上定义好的参数拿过来
pretrained_net.fc # full connection全连接层，最后一层，查看最后一层的输入和输出结构
Out[5]:
Linear(in_features=512, out_features=1000, bias=True)
In [6]:
finetune_net = torchvision.models.resnet18(pretrained=True)
finetune_net.fc = nn.Linear(finetune_net.fc.in_features,2) # 最后一层修改为输出类别数为2
nn.init.xavier_uniform_(finetune_net.fc.weight) # 只对最后一层的weight做随即初始化
Out[6]:
Parameter containing:
tensor([[ 0.0004, -0.0395, -0.0163, ..., 0.0185, -0.0238, 0.0693],
[ 0.0307, 0.0278, 0.0082, ..., -0.0852, 0.0642, -0.0302]],
requires_grad=True)
In [7]:
# 微调座位
def train_fine_tuning(net, learning_rate, batch_size=128, num_epochs=5, param_group=True):
train_iter = torch.utils.data.DataLoader(
torchvision.datasets.ImageFolder(os.path.join(data_dir,'train'),transform=train_augs),
batch_size = batch_size,shuffle=True)
test_iter = torch.utils.data.DataLoader(
torchvision.datasets.ImageFolder(os.path.join(data_dir,'test'),transform=test_augs),
batch_size=batch_size)
devices = d2l.try_all_gpus()
loss = nn.CrossEntropyLoss(reduction="none")
if param_group:
# 除了最后一层的learning rate外，用的是默认的learning rate
# 最后一层的learning rate用的是十倍的learning rate
params_lx = [
param for name, param in net.named_parameters()
if name not in ["fc.weight","fc.bias"] ]
trainer = torch.optim.SGD([
{'params': params_lx},
{'params': net.fc.parameters(), 'lr': learning_rate * 10}],
lr=learning_rate, weight_decay=0.001)
else:
trainer = torch.optim.SGD(net.parameters(),lr=learning_rate,weight_decay=0.001)
d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs,devices)
In [8]:
# 使用较小的学习率
train_fine_tuning(finetune_net,5e-5)
loss 0.163, train acc 0.932, test acc 0.935 265.9 examples/sec on [device(type='cuda', index=0)]
In [9]:
# 为了进行比较，所有模型参数初始化为随机值
scratch_net = torchvision.models.resnet18() # 这里没有pretrained=True，没有拿预训练的参数
scratch_net.fc = nn.Linear(scratch_net.fc.in_features,2)
train_fine_tuning(scratch_net,5e-4,param_group=False) # param_group=False使得所有层的参数都为默认的学习率
loss 0.349, train acc 0.847, test acc 0.806 422.1 examples/sec on [device(type='cuda', index=0)]
