pytorc加载模型后，修改最后一层的分类输出层，in |
您所在的位置：网站首页 › feature的含义 › pytorc加载模型后，修改最后一层的分类输出层，in |
|
def forward(self, inputs):
""" Calls extract_features to extract features, applies final linear layer, and returns logits. """
# Convolution layers
x = self.extract_features(inputs)
# Pooling and final linear layer
x = F.adaptive_avg_pool2d(x, 1).squeeze(-1).squeeze(-1)
if self._dropout:
x = F.dropout(x, p=self._dropout, training=self.training)
x = self._fc(x)
return x
修改分类输出层1、，用in_features,得到，该层的输入，重写这一层 from efficientnet_pytorch import EfficientNet from torch import nn model = EfficientNet.from_pretrained('efficientnet-b5') feature = model._fc.in_features model._fc = nn.Linear(in_features=feature,out_features=45,bias=True) print(model)修改分类输出层2、 用 out_features，得到该层的输出，直接修改分类输出个数 from efficientnet_pytorch import EfficientNet model = EfficientNet.from_pretrained('efficientnet-b5') model._fc.out_features = 9 print(model) 不同层，不同学习率设置 class net(nn.Module): def __init__(self): super(net, self).__init__() self.conv1 = nn.Conv2d(3, 64, 1) self.conv2 = nn.Conv2d(64, 64, 1) self.conv3 = nn.Conv2d(64, 64, 1) self.conv4 = nn.Conv2d(64, 64, 1) self.conv5 = nn.Conv2d(64, 64, 1) def forward(self, x): out = conv5(conv4(conv3(conv2(conv1(x))))) return out ———————————————— 版权声明：本文为CSDN博主「BinWang-cvlab」的原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接及本声明。 原文链接：https://blog.csdn.net/wangbin12122224/article/details/79949824 某个层，学习率更大一点，其他层，用统一的学习率 net = net() lr = 0.1 conv5_params = list(map(id, net.conv5.parameters())) #这一层学习率，不一样,lr * 100 base_params = filter(lambda p: id(p) not in conv5_params, ##过滤 conv5 net.parameters()) optimizer = torch.optim.SGD([ {'params': base_params}, {'params': net.conv5.parameters(), 'lr': lr * 100} ], , lr=lr, momentum=0.9)多个层，写法conv5 conv4是定义结构的网络的名字，根据不同的网络结构自己修改名字 conv5_params = list(map(id, net.conv5.parameters())) conv4_params = list(map(id, net.conv4.parameters())) base_params = filter(lambda p: id(p) not in conv5_params + conv4_params, net.parameters()) ##注意写法是这样的 用 + optimizer = torch.optim.SGD([ {'params': base_params}, {'params': net.conv5.parameters(), 'lr': lr * 100}, {'params': net.conv4.parameters(), 'lr': lr * 100}, , lr=lr, momentum=0.9) 学习率设置完成后的学习率衰减，操作 #################### # model structure #------------------- model = Mymodel() if use_gpu: model = model.cuda() #################### # loss #------------------- criterion = nn.CrossEntropyLoss() #################### # optimizer #------------------- ignored_params = list(map(id, model.ViewModel.viewclassifier.parameters())) + list(map(id, model.Block.parameters())) base_params = filter(lambda p: id(p) not in ignored_params, model.parameters()) optimizer_ft = optim.SGD([ {'params': base_params, 'lr': 0.01}, {'params': model.ViewModel.viewclassifier.parameters(), 'lr': 0.001}, {'params': model.Block.parameters(), 'lr': 0.03}], weight_decay=1e-3, momentum=0.9, nesterov=True) #################### #** Set lr_decay ** #------------------- ##能分别设定不同层的衰减率 def adjust_lr(epoch): step_size = 60 lr = args.lr * (0.1 ** (epoch // 30)) #epoch // 30取整操作， 0 1 2 3 .。。。。 for g in optimizer.param_groups: g['lr'] = lr * g.get('lr') ###################################### ### optimizer.param_groups 类型与内容 [ { 'params': base_params, 'lr': 0.01, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0.001, 'nesterov': True, 'initial_lr': 0.01 }, { 'params': model.ViewModel.viewclassifier.parameters(), 'lr': 0.001, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0.001, 'nesterov': True, 'initial_lr': 0.001 }, { 'params': model.Block.parameters(), 'lr': 0.03, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0.001, 'nesterov': True, 'initial_lr': 0.03 } ] ### optimizer.param_groups 类型与内容 ###################################### for epoch in range(start_epoch, args.epochs): adjust_lr(epoch) # 每epoch更新一次。 model.train(True) # Set model to training mode .... #### 或者 exp_lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=60, gamma=0.1) for epoch in range(start_epoch, args.epochs): scheduler.step() # put it before model.train(True) model.train(True) # Set model to training mode学习率设置，其他方式 def lr_poly(base_lr, iter, max_iter, power): return base_lr*((1-float(iter)/max_iter)**(power)) def adjust_learning_rate(optimizer, learning_rate, i_iter, max_iter, power): """Sets the learning rate to the initial LR divided by 5 at 60th, 120th and 160th epochs""" lr = lr_poly(learning_rate, i_iter, max_iter, power) optimizer.param_groups[0]['lr'] = lr return lr ———————————————— 版权声明：本文为CSDN博主「咆哮的阿杰」的原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接及本声明。 原文链接：https://blog.csdn.net/qq_34914551/article/details/87699317 冻结某个层的写法，优化器需要filter 冻结的层 for name, p in net.named_parameters(): #两种循环网络的名字 for p in net.parameters(): ##冻结，参数 model = EfficientNet.from_pretrained(args.arch,num_classes=29) for param in model._bn0.parameters(): #网络结构的名字，想冻结到哪里就写那一层的名字 param.requires_grad = False #直接设置 false optimizer = optim.SGD( #优化器 需要过滤掉所有的 梯度为0 的东接层 filter(lambda p: p.requires_grad, model.parameters()), # 记住一定要加上filter()，不然会报错 lr=0.01, weight_decay=1e-5, momentum=0.9, nesterov=True) https://blog.csdn.net/qq_34914551/article/details/87699317 Adam 优化器，写法，以及model.module打印 |
今日新闻 |
推荐新闻 |
| CopyRight 2018-2019 办公设备维修网 版权所有 豫ICP备15022753号-3 |