pytorch 加载部分参数，以及多GPU 加载单GPUmodel，pop去除全连接参数，然后加载

            ######### multi gpu  load one gpu need mudule.
            print ("model  load  param###########################")
            pretrained_dict = torch.load("model_ir_se50.pth")   ###
            self.model_dict = self.model.state_dict()  #get the name:value
            param={}
            for k, v in pretrained_dict.items():
                if k[7:] !="module" :
                    param["module."+k] = pretrained_dict[k]

            pretrained_dict = { k: v for k, v in param.items() if k in self.model_dict}
            self.model_dict.update(pretrained_dict)
            self.model.load_state_dict(self.model_dict)


实际上就是字典的操作，那么字典的操作，哪些层不要，打印出层的名字就可以了，例如
模型参数的某些层的权重不要，那么重构一个字典参数就可以了，for k in torch.load("**.pth').keys():   打印出来按照名字删除 键值对，
所以字典的pop删除操作也是可以的， 多卡训练参数多了一个module

pretrain = torch.load("/home/imagenet.pth")
new_state_dict = {}#OrderedDict()
for k,v in pretrain.items():
    if "classifier" in k:  #最后分类层的参数是classeifer ,不需要这个模型参数
        continue
    new_state_dict[k[7:]] = v  #remove `module.`  #模型k 有module 不要
model.load_state_dict(new_state_dict, strict=False)  #strict =False ,模型参数和模型不一致可以加载

 model = DPN(num_init_features=64, k_R=96, G=32, k_sec=(3,4,20,3), inc_sec=(16,32,24,128), num_classes=1,decoder=args.decoder)
 http = {'url': 'http://data.lip6.fr/cadene/pretrainedmodels/dpn92_extra-b040e4a9b.pth'}
 pretrained_dict=model_zoo.load_url(http['url'])
 model_dict = model.state_dict()
 pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}#filter out unnecessary keys 
 model_dict.update(pretrained_dict)
 model.load_state_dict(model_dict)
 model = torch.nn.DataParallel(model).cuda()

简单写法

直接pop 去除全连接参数，然后加载，字典的去除键值对的方法

            ###############加载自己训练的模型
            print ("加载自己训练的模型")
            pretrained_dict = torch.load("model_best.pth.tar")["state_dict"] #保存的模型的全部结果包括全连接，优化器
            for k,v in pretrained_dict.items():
                print (k)
                print (v.shape)
            pretrained_dict.pop('_fc.weight') # 加载的参数直接删除全连接层的参数，
            pretrained_dict.pop('_fc.bias')
            model.load_state_dict(pretrained_dict, strict=False)

参数 打印结果，有分类输出层

_blocks.15._se_reduce.bias
torch.Size([48])
_blocks.15._se_expand.weight
torch.Size([1152, 48, 1, 1])
_blocks.15._se_expand.bias
torch.Size([1152])
_blocks.15._project_conv.weight
torch.Size([320, 1152, 1, 1])
_blocks.15._bn2.weight
torch.Size([320])
_blocks.15._bn2.bias
torch.Size([320])
_blocks.15._bn2.running_mean
torch.Size([320])
_blocks.15._bn2.running_var
torch.Size([320])
_blocks.15._bn2.num_batches_tracked
torch.Size([])
_conv_head.weight
torch.Size([1280, 320, 1, 1])
_bn1.weight
torch.Size([1280])
_bn1.bias
torch.Size([1280])
_bn1.running_mean
torch.Size([1280])
_bn1.running_var
torch.Size([1280])
_bn1.num_batches_tracked
torch.Size([])
_fc.weight
torch.Size([115, 1280])
_fc.bias
torch.Size([115])
Using image size 224