Finetune预训练模型

1、PyTorch学习笔记（1）-finetune网络的一些注意事项https://blog.csdn.net/u014448054/article/details/80623514k+

2、全局微调和局部微调

https://blog.csdn.net/u012759136/article/details/65634477

3、局部微调选择层数

https://blog.csdn.net/u012436149/article/details/78038098

4、CS231关于transfer learning的教程

http://cs231n.github.io/transfer-learning/

5、Pytorch Tutorial for Fine Tuning/Transfer Learning a Resnet for Image Classification

https://github.com/Spandan-Madan/Pytorch_fine_tuning_Tutorial

6、pytorch官方教程

https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html

https://pytorch.org/docs/stable/torchvision/models.html

github网络配置文件

https://github.com/pytorch/vision/tree/master/torchvision/models

例程：

import torch

from torchvisionimport models

model = models.resnet18(pretrained=True)

# #######全局finetune，使用不同的learning rate#########

# ignored_params = list(map(id,model.fc.parameters()))

# base_params = filter(lambda p: id(p) not in ignored_params, model.parameters())

#

# optimizer = torch.optim.SGD([{'params': base_params}, {'params': model.fc.parameters(), 'lr': 1e-5}], lr=1e-4, momentum=0.9)

#

#######局部finetune，只学习fc#######################

# in_features = model.fc.in_features

# #######change the output feature of fc######

# model.fc = torch.nn.Linear(in_features, 15)

# for param in model.parameters():

#    param.requires_grad = False

# model.fc = torch.nn.Linear(in_features, 15)

# optimizer = torch.optim.SGD(model.fc.parameters(), lr=1e-2, momentum=0.9)

#

######局部finetune, 学习fc和top parameters#########

fc_params = list(map(id,model.fc.parameters()))

top_params = list(map(id, model.layer4.parameters()))

ignored_params = fc_params.append(top_params)

base_params_model = filter(lambda p: id(p)not in ignored_params, model.parameters())

fc_params_model = filter(lambda p: id(p)in fc_params, model.parameters())

top_params_model = filter(lambda p: id(p)in top_params, model.parameters())

optimizer = torch.optim.SGD([{'params': fc_params_model}, {'params': top_params_model, 'lr':1e-3}], lr=1e-2, momentum=0.9)

报错：fc size mismatch

inception_v3 RuntimeError: size mismatch

原因是input image的size应该是224（vgg、resnet)或者299（inception）不是448