# 關於pytorch置換backbone

###### tags: `beginner`

## 前言
常在進行機器學習時，我們會需要用到網路上開源的pretrained weights，但因為每個架構的名稱不同，常會導致pretrained weights無法順利讀入使用的狀況，故在這篇對pytorch的pretrained model如何置換backbone、新增或刪除layer進行一些筆記和紀錄

*筆者實際都是使用ResNet50進行*

## load pretrained model

pretrained_model = torch.hub.load(
            'pytorch/vision:v0.5.0', 
            'resnet50', pretrained=True
            ) 
            #在pretrained使用True就可以下載hub中的pretrained weights
            #
            #如果只是想要模型架構，可以設為False，並透過下方程式碼
            #對權重進行讀取
checkpoint = torch.load(
        "checkpoint.pth.tar", map_location="cuda"
    )

## add layer
* 方法一：將原先的model需要的架構直接包入，可以忽略掉

pretrained = model #your pretrained model
class MyResNet(torch.nn.Module):
    def __init__(self, my_pretrained_model):
        super(MyResNet, self).__init__()
        self.pretrained = my_pretrained_model
        self.my_new_layers = layer #New layer you want to add
    
    def forward(self, x):
        x = self.pretrained(x)
        # Sometimes,we take sub-model, and the code below might be helpful
        # x = self.pretrained(x)['3']
        x = self.my_new_layers(x)
        return x

* 方法二：對pretrained model的名稱進行置換，換成符合自己模型的名稱

from collections import OrderedDict
pretrain_weight = {'state_dict':OrderedDict()}
for key in checkpoint['state_dict']:
    pretrain_weight['state_dict'][key[7:]]=checkpoint['state_dict'][key]
new_model.load_state_dict(pretrain_weight['state_dict'],strict=False)

## delete layer and output feature from middle layer

# *Still working...*