source_code/sam3/medsam3_brats/lora.py

#!/usr/bin/env python3
"""
LoRA (Low-Rank Adaptation) 模块用于SAM3微调

参考: https://arxiv.org/abs/2106.09685
"""

import math
from typing import Dict, List, Optional, Tuple, Union

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


class LoRALinear(nn.Module):
    """
    LoRA线性层
    
    将原始线性层分解为: W = W_0 + BA
    其中 B ∈ R^{d×r}, A ∈ R^{r×k}, r << min(d, k)
    """
    
    def __init__(
        self,
        original_layer: nn.Linear,
        rank: int = 4,
        alpha: float = 1.0,
        dropout: float = 0.0,
    ):
        super().__init__()
        
        self.original_layer = original_layer
        self.rank = rank
        self.alpha = alpha
        self.scaling = alpha / rank
        
        in_features = original_layer.in_features
        out_features = original_layer.out_features
        
        # LoRA参数
        self.lora_A = nn.Parameter(torch.zeros(rank, in_features))
        self.lora_B = nn.Parameter(torch.zeros(out_features, rank))
        
        # 初始化
        nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
        nn.init.zeros_(self.lora_B)
        
        # Dropout
        self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
        
        # 冻结原始层
        for param in self.original_layer.parameters():
            param.requires_grad = False
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # 原始前向传播
        result = self.original_layer(x)
        
        # 确保LoRA参数在正确的设备上
        if self.lora_A.device != x.device:
            self.lora_A = nn.Parameter(self.lora_A.to(x.device))
            self.lora_B = nn.Parameter(self.lora_B.to(x.device))
        
        # LoRA分支
        x_dropout = self.dropout(x)
        with torch.amp.autocast('cuda', enabled=False):
            x_fp32 = x_dropout.float()
            lora_A = self.lora_A.float()
            lora_B = self.lora_B.float()
            lora_out = x_fp32 @ lora_A.T @ lora_B.T
            lora_out = lora_out.to(result.dtype)
        
        return result + lora_out * self.scaling
    
    def merge_weights(self):
        """将LoRA权重合并到原始层"""
        if self.rank > 0:
            self.original_layer.weight.data += (
                self.lora_B @ self.lora_A * self.scaling
            )
    
    def unmerge_weights(self):
        """从原始层移除LoRA权重"""
        if self.rank > 0:
            self.original_layer.weight.data -= (
                self.lora_B @ self.lora_A * self.scaling
            )


class LoRAConv2d(nn.Module):
    """
    LoRA卷积层
    使用与原始卷积相同的stride和padding
    """
    
    def __init__(
        self,
        original_layer: nn.Conv2d,
        rank: int = 4,
        alpha: float = 1.0,
        dropout: float = 0.0,
    ):
        super().__init__()
        
        self.original_layer = original_layer
        self.rank = rank
        self.alpha = alpha
        self.scaling = alpha / rank
        
        in_channels = original_layer.in_channels
        out_channels = original_layer.out_channels
        kernel_size = original_layer.kernel_size
        stride = original_layer.stride
        padding = original_layer.padding
        
        # LoRA参数
        # lora_A: 降维 (in_ch -> rank)，使用1x1卷积
        # lora_B: 升维 (rank -> out_ch)，使用原始kernel_size保持空间维度一致
        self.lora_A = nn.Conv2d(
            in_channels, rank, kernel_size=1, stride=1, padding=0, bias=False
        )
        self.lora_B = nn.Conv2d(
            rank, out_channels, kernel_size=kernel_size, stride=stride, padding=padding, bias=False
        )
        
        # 初始化
        nn.init.kaiming_uniform_(self.lora_A.weight, a=math.sqrt(5))
        nn.init.zeros_(self.lora_B.weight)
        
        self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
        
        # 冻结原始层
        for param in self.original_layer.parameters():
            param.requires_grad = False
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        result = self.original_layer(x)
        # 确保LoRA层在正确的设备上
        if self.lora_A.weight.device != x.device:
            self.lora_A = self.lora_A.to(x.device)
            self.lora_B = self.lora_B.to(x.device)
        
        x_dropout = self.dropout(x)
        # 使用与输入相同的dtype
        with torch.amp.autocast('cuda', enabled=False):
            x_fp32 = x_dropout.float()
            h = self.lora_A(x_fp32)
            lora_out = self.lora_B(h)
            lora_out = lora_out.to(result.dtype)
        return result + lora_out * self.scaling


def apply_lora_to_model(
    model: nn.Module,
    rank: int = 4,
    alpha: float = 1.0,
    dropout: float = 0.0,
    target_modules: Optional[List[str]] = None,
    exclude_modules: Optional[List[str]] = None,
) -> nn.Module:
    """
    将LoRA应用到模型的指定层
    
    Args:
        model: 原始模型
        rank: LoRA秩
        alpha: LoRA缩放因子
        dropout: LoRA dropout
        target_modules: 目标模块名称列表 (如 ['q_proj', 'v_proj'])
        exclude_modules: 排除的模块名称列表
    
    Returns:
        应用了LoRA的模型
    """
    if target_modules is None:
        # 默认目标: attention层的QKV投影
        target_modules = [
            'q_proj', 'k_proj', 'v_proj',
            'qkv', 'to_q', 'to_k', 'to_v',
            'query', 'key', 'value',
        ]
    
    if exclude_modules is None:
        exclude_modules = []
    
    lora_count = 0
    
    for name, module in model.named_modules():
        # 检查是否应该排除
        should_exclude = any(ex in name for ex in exclude_modules)
        if should_exclude:
            continue
        
        # 检查是否是目标模块
        is_target = any(target in name for target in target_modules)
        
        if is_target:
            if isinstance(module, nn.Linear):
                # 获取父模块和属性名
                parent_name = '.'.join(name.split('.')[:-1])
                attr_name = name.split('.')[-1]
                
                parent = model
                if parent_name:
                    for part in parent_name.split('.'):
                        parent = getattr(parent, part)
                
                # 替换为LoRA层
                lora_layer = LoRALinear(module, rank=rank, alpha=alpha, dropout=dropout)
                setattr(parent, attr_name, lora_layer)
                lora_count += 1
            
            elif isinstance(module, nn.Conv2d):
                parent_name = '.'.join(name.split('.')[:-1])
                attr_name = name.split('.')[-1]
                
                parent = model
                if parent_name:
                    for part in parent_name.split('.'):
                        parent = getattr(parent, part)
                
                lora_layer = LoRAConv2d(module, rank=rank, alpha=alpha, dropout=dropout)
                setattr(parent, attr_name, lora_layer)
                lora_count += 1
    
    print(f"Applied LoRA to {lora_count} layers")
    return model


def get_lora_params(model: nn.Module) -> List[nn.Parameter]:
    """获取所有LoRA参数"""
    lora_params = []
    for name, param in model.named_parameters():
        if 'lora_' in name:
            lora_params.append(param)
    return lora_params


def get_trainable_params(model: nn.Module) -> Tuple[List[nn.Parameter], List[str]]:
    """获取所有可训练参数"""
    trainable_params = []
    trainable_names = []
    for name, param in model.named_parameters():
        if param.requires_grad:
            trainable_params.append(param)
            trainable_names.append(name)
    return trainable_params, trainable_names


def freeze_model_except_lora(model: nn.Module):
    """冻结除LoRA外的所有参数"""
    for name, param in model.named_parameters():
        if 'lora_' not in name:
            param.requires_grad = False
        else:
            param.requires_grad = True


def count_parameters(model: nn.Module) -> Dict[str, int]:
    """统计参数数量"""
    total = sum(p.numel() for p in model.parameters())
    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    frozen = total - trainable
    
    return {
        'total': total,
        'trainable': trainable,
        'frozen': frozen,
        'trainable_ratio': trainable / total if total > 0 else 0,
    }


def save_lora_weights(model: nn.Module, path: str):
    """只保存LoRA权重"""
    lora_state_dict = {}
    for name, param in model.named_parameters():
        if 'lora_' in name:
            lora_state_dict[name] = param.data
    torch.save(lora_state_dict, path)
    print(f"Saved LoRA weights to {path}")


def load_lora_weights(model: nn.Module, path: str):
    """加载LoRA权重"""
    lora_state_dict = torch.load(path, map_location='cpu')
    
    model_state_dict = model.state_dict()
    for name, param in lora_state_dict.items():
        if name in model_state_dict:
            model_state_dict[name] = param
    
    model.load_state_dict(model_state_dict, strict=False)
    print(f"Loaded LoRA weights from {path}")


if __name__ == "__main__":
    # 测试
    print("Testing LoRA module...")
    
    # 测试线性层
    linear = nn.Linear(768, 768)
    lora_linear = LoRALinear(linear, rank=8, alpha=16)
    
    x = torch.randn(2, 768)
    y = lora_linear(x)
    print(f"LoRA Linear output shape: {y.shape}")
    
    # 测试卷积层
    conv = nn.Conv2d(256, 256, 3, padding=1)
    lora_conv = LoRAConv2d(conv, rank=8, alpha=16)
    
    x = torch.randn(2, 256, 32, 32)
    y = lora_conv(x)
    print(f"LoRA Conv output shape: {y.shape}")
    
    # 测试参数统计
    model = nn.Sequential(
        lora_linear,
        nn.ReLU(),
        nn.Linear(768, 10)
    )
    
    stats = count_parameters(model)
    print(f"Parameters: {stats}")