人工智能 - 如何获取 ResNet-50 模型的前 48 层的输出？

Question

我想看看 resnet50 模型，第 48 层的输出，我写了下面的代码，但是运行报错了

import torch
import torchvision
import torchvision.transforms as transforms
from PIL import Image
from torch import Tensor
import torch.nn as nn

# 加载ResNet-50模型
model = torchvision.models.resnet50(pretrained=True)

# 获取前48层的子模型
model = nn.Sequential(*list(model.children())[:48])

# 修改fc层
# model.fc = nn.Linear(2048, 512)

# 设置模型为评估模式
model.eval()

# 图像预处理
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
    )
])

# 加载并预处理图像
image = Image.open('std.jpg')
image = transform(image).unsqueeze(0)  # 添加批次维度

# 使用模型进行推理
with torch.no_grad():
    features: Tensor = model(image)
    print(features.shape)

报错如下：

/Users/ponponon/.local/share/virtualenvs/image2vector-n-kX1tX6/lib/python3.10/site-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and may be removed in the future, please use 'weights' instead.
  warnings.warn(
/Users/ponponon/.local/share/virtualenvs/image2vector-n-kX1tX6/lib/python3.10/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and may be removed in the future. The current behavior is equivalent to passing `weights=ResNet50_Weights.IMAGENET1K_V1`. You can also use `weights=ResNet50_Weights.DEFAULT` to get the most up-to-date weights.
  warnings.warn(msg)
Traceback (most recent call last):
  File "/Users/ponponon/Desktop/code/me/resnet_example/resnet48_handle_image_into_vector.py", line 35, in <module>
    features: Tensor = model(image)
  File "/Users/ponponon/.local/share/virtualenvs/image2vector-n-kX1tX6/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/Users/ponponon/.local/share/virtualenvs/image2vector-n-kX1tX6/lib/python3.10/site-packages/torch/nn/modules/container.py", line 217, in forward
    input = module(input)
  File "/Users/ponponon/.local/share/virtualenvs/image2vector-n-kX1tX6/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/Users/ponponon/.local/share/virtualenvs/image2vector-n-kX1tX6/lib/python3.10/site-packages/torch/nn/modules/linear.py", line 114, in forward
    return F.linear(input, self.weight, self.bias)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (2048x1 and 2048x1000)

我明明把后面的 fc 给丢掉了呀，为什么还报错呢？

我该如何修改？

Answer 1

改成下面这样就可以了

import torchvision.models as models
import torch.nn.functional as F
import torch
import torchvision
import torchvision.transforms as transforms
from PIL import Image
from torch import Tensor
import torch.nn as nn


class ImageRetrievalNet(nn.Module):

    def __init__(self, dim: int = 512):
        super().__init__()
        resnet50_model = models.resnet50()
        features = list(resnet50_model.children())[:-2]

        self.features = nn.Sequential(*features)

    def forward(self, x: Tensor):
        # featured_t shape: torch.Size([1, 2048, 7, 7])
        featured_t: Tensor = self.features(x)

        print(featured_t.shape)

        return featured_t


# 加载ResNet-50模型
model = ImageRetrievalNet()

# 修改fc层
# model.fc = nn.Linear(2048, 512)

# 设置模型为评估模式
model.eval()

# 图像预处理
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
    )
])

# 加载并预处理图像
image = Image.open('std.jpg')
image = transform(image).unsqueeze(0)  # 添加批次维度

# 使用模型进行推理
with torch.no_grad():
    features: Tensor = model(image)
    print(features.shape)