统计模型的参数

统计模型的参数量以及大小

以pytorch官方的resnet18为例。首先获取模型的参数，有一下两种方法：

named_parameters() 方法返回一个迭代器，用于同时访问模型中的参数和它们的名称。

parameters() 方法返回一个迭代器，用于访问模型中所有需要梯度更新的参数。

# model.named_parameters()
for name, param in model.named_parameters():
    print(name, param.shape)

# 输出：
# conv1.weight torch.Size([64, 3, 7, 7])
# bn1.weight torch.Size([64])
# bn1.bias torch.Size([64])
# layer1.0.conv1.weight torch.Size([64, 64, 3, 3])

# model.parameters()
for params in model18.parameters():
    print(params.shape)

# torch.Size([64, 3, 7, 7])
# torch.Size([64])
# torch.Size([64])
# torch.Size([64, 64, 3, 3])

---

计算模型的参数量以及大小

import torchvision.models as models

model18 = models.resnet18()

total_size_resnet18 = 0
total_num_resnet18 = 0
for param in model18.parameters():
    param_size = torch.tensor(param.size()).prod() * param.element_size()
    param_num = torch.tensor(param.size()).prod()
    total_num_resnet18 += param_num.item()
    total_size_resnet18 += param_size.item()

print(f"Total parameters size: {total_size_resnet18 / (1024 ** 2):.2f} MB")
print(f"Total parameters number: {total_num_resnet18 } 个")

# Total parameters size: 44.59 MB
# Total parameters number: 11689512 个

---

保存的’'.pth文件大小也说明了这一点，虽然不是完全一样大，但大小差不多，可能与状态字典多的键值有关。

---

resnet18的构造

layer 2, 3, 4层的downsample是对残差传过来的上一层处理结果进行的，kernel_size = 1, stride = 2，维度翻倍，尺寸减半。

ResNet(
  (conv1): 3, 64, 7, 2, 3            # conv1.weight torch.Size([64, 3, 7, 7])
  (bn1):                             # bn1.weight torch.Size([64]) / bn1.bias torch.Size([64])
  (relu):                         
  (maxpool):3, 2, 1
  (layer1): Sequential(
    (0): BasicBlock(
      (conv1): 64, 64, 3, 1, 1       # layer1.0.conv1.weight torch.Size([64, 64, 3, 3])
      (bn1): 
      (relu): 
      (conv2): 64, 64, 3, 1, 1
      (bn2): 
    )
    (1): BasicBlock(
      (conv1): 64, 64, 3, 1, 1
      (bn1): 
      (relu): 
      (conv2): 64, 64, 3, 1, 1
      (bn2): 
    )
  )
  (layer2): Sequential(
    (0): BasicBlock(
      (conv1): 64, 128, 3, 2, 1
      (bn1): 
      (relu): 
      (conv2): 128, 128, 3, 1, 1
      (bn2): 
      (downsample): Sequential(
        (0): 64, 128, 1, 2
        (1): 
      )
    )
    (1): BasicBlock(
      (conv1): 128, 128, 3, 1, 1
      (bn1):
      (relu): 
      (conv2): 128, 128, 3, 1, 1
      (bn2): 
    )
  )
  (layer3): Sequential(
    (0): BasicBlock(
      (conv1): 128, 256, 3, 2, 1
      (bn1): 
      (relu): 
      (conv2): 256, 256, 3, 1, 1
      (bn2): 
      (downsample): Sequential(
        (0): 128, 256, 1, 2
        (1): 
      )
    )
    (1): BasicBlock(
      (conv1): 256, 256, 3, 1, 1
      (bn1): 
      (relu): 
      (conv2): 256, 256, 3, 1, 1
      (bn2): 
    )
  )
  (layer4): Sequential(
    (0): BasicBlock(
      (conv1): 256, 512, 3, 2, 1
      (bn1): 
      (relu): 
      (conv2): 512, 512, 3, 1, 1
      (bn2): 
      (downsample): Sequential(
        (0): 256, 512, 3, 1
        (1): 
      )
    )
    (1): BasicBlock(
      (conv1): 512, 512, 3, 1, 1
      (bn1): 
      (relu): 
      (conv2): 512, 512, 3, 1, 1
      (bn2): 
    )
  )
  (avgpool): 1,1
  (fc): 512, 1000
)