解决PyTorch多标签分类中批次大小不一致问题：模型架构与张量形变管理

# 初始尝试：修改预训练模型的分类头 (不适用多头输出)
# model2.classifier_artist = torch.nn.Sequential(...)
# model2.classifier_style = torch.nn.Sequential(...)
# model2.classifier_genre = torch.nn.Sequential(...)

import torch
import torch.nn as nn
import torch.nn.functional as F

class WikiartModel(nn.Module):
    def __init__(self, num_artists, num_genres, num_styles):
        super(WikiartModel, self).__init__()

        # 共享卷积层
        self.conv1 = nn.Conv2d(3, 64, kernel_size=3, padding =1)
        self.conv2 = nn.Conv2d(64, 128, kernel_size=3, padding=1)
        self.conv3 = nn.Conv2d(128, 256, kernel_size=3, padding=1)
        self.pool = nn.MaxPool2d(2, 2) # 最大池化层

        # 艺术家分类分支
        self.fc_artist1 = nn.Linear(256 * 16 * 16, 512) # 错误：输入特征维度计算有误
        self.fc_artist2 = nn.Linear(512, num_artists)

        # 流派分类分支
        self.fc_genre1 = nn.Linear(256 * 16 * 16, 512) # 错误：输入特征维度计算有误
        self.fc_genre2 = nn.Linear(512, num_genres)

        # 风格分类分支
        self.fc_style1 = nn.Linear(256 * 16 * 16, 512) # 错误：输入特征维度计算有误
        self.fc_style2 = nn.Linear(512, num_styles)

    def forward(self, x):
        # 共享卷积层处理
        x = self.pool(F.relu(self.conv1(x)))   
        x = self.pool(F.relu(self.conv2(x)))       
        x = self.pool(F.relu(self.conv3(x)))

        # 张量形变：将多维特征图展平为一维向量
        x = x.view(-1, 256 * 16  * 16) # 错误：展平后的维度计算有误，且-1可能导致意外行为

        # 艺术家分类分支
        artists_out = F.relu(self.fc_artist1(x))
        artists_out = self.fc_artist2(artists_out)

        # 流派分类分支
        genre_out = F.relu(self.fc_genre1(x))
        genre_out = self.fc_genre2(genre_out) 

        # 风格分类分支 
        style_out = F.relu(self.fc_style1(x))
        style_out = self.fc_style2(style_out)

        return artists_out, genre_out, style_out

# 设置类别数量
num_artists = 129
num_genres = 11
num_styles = 27

import torch
import torch.nn as nn
import torch.nn.functional as F

class WikiartModel(nn.Module):
    def __init__(self, num_artists, num_genres, num_styles):
        super(WikiartModel, self).__init__()

        # 共享卷积层
        self.conv1 = nn.Conv2d(3, 64, kernel_size=3, padding =1)
        self.conv2 = nn.Conv2d(64, 128, kernel_size=3, padding=1)
        self.conv3 = nn.Conv2d(128, 256, kernel_size=3, padding=1)
        self.pool = nn.MaxPool2d(2, 2)

        # 计算卷积层最终输出的特征图尺寸（例如，对于224x224输入，经过三次conv+pool后为28x28）
        # 建议在模型初始化时或通过一个小的dummy_input计算得出
        # 确保这里的尺寸与实际计算结果一致
        self.feature_map_size = 28 # 经过三次池化后，224 -> 112 -> 56 -> 28
        self.flattened_features = 256 * self.feature_map_size * self.feature_map_size # 256 * 28 * 28 = 200704

        # 艺术家分类分支
        self.fc_artist1 = nn.Linear(self.flattened_features, 512) # 修正此处输入特征维度
        self.fc_artist2 = nn.Linear(512, num_artists)

        # 流派分类分支
        self.fc_genre1 = nn.Linear(self.flattened_features, 512) # 修正此处输入特征维度
        self.fc_genre2 = nn.Linear(512, num_genres)

        # 风格分类分支
        self.fc_style1 = nn.Linear(self.flattened_features, 512) # 修正此处输入特征维度
        self.fc_style2 = nn.Linear(512, num_styles)

    def forward(self, x):
        # 共享卷积层处理
        x = self.pool(F.relu(self.conv1(x)))   
        x = self.pool(F.relu(self.conv2(x)))       
        x = self.pool(F.relu(self.conv3(x)))

        # 张量形变：展平张量，保留批次大小
        # x.size(0) 获取当前批次大小，-1让PyTorch自动计算剩余维度
        x = x.view(x.size(0), -1) 

        # 艺术家分类分支
        artists_out = F.relu(self.fc_artist1(x))
        artists_out = self.fc_artist2(artists_out)

        # 流派分类分支
        genre_out = F.relu(self.fc_genre1(x))
        genre_out = self.fc_genre2(genre_out) 

        # 风格分类分支 
        style_out = F.relu(self.fc_style1(x))
        style_out = self.fc_style2(style_out)

        return artists_out, genre_out, style_out

# 设置类别数量
num_artists = 129
num_genres = 11
num_styles = 27

# 实例化模型并进行测试 (示例)
model = WikiartModel(num_artists, num_genres, num_styles)
dummy_input = torch.randn(32, 3, 224, 224) # 批次大小为32的模拟输入
artist_output, genre_output, style_output = model(dummy_input)

print(f"Artist Output Shape: {artist_output.shape}") # 预期: [32, 129]
print(f"Genre Output Shape: {genre_output.shape}")   # 预期: [32, 11]
print(f"Style Output Shape: {style_output.shape}")   # 预期: [32, 27]

# 此时，torchinfo的输出也将显示正确的批次大小
# from torchinfo import summary
# summary(model, input_size=(32, 3, 224, 224))