| 119 | |
| 120 | |
| 121 | class LoRAAdapterConv(nn.Module): |
| 122 | def __init__( |
| 123 | self, |
| 124 | in_channels: int, |
| 125 | out_channels: int, |
| 126 | kernel_size: Union[int, Tuple[int, int]], |
| 127 | stride: Union[int, Tuple[int, int]], |
| 128 | padding: Union[int, Tuple[int, int]], |
| 129 | data_provider: DataProvider, |
| 130 | c_dim: int, |
| 131 | rank: int = None, |
| 132 | lora_scale: float = 1.0, |
| 133 | ): |
| 134 | super().__init__() |
| 135 | |
| 136 | # self.lora_scale = alpha / rank |
| 137 | self.lora_scale = lora_scale |
| 138 | self.c_dim = c_dim |
| 139 | |
| 140 | self.data_provider = data_provider |
| 141 | |
| 142 | self.W = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding) |
| 143 | for p in self.W.parameters(): |
| 144 | p.requires_grad_(False) |
| 145 | |
| 146 | self.A = nn.Conv2d(in_channels, rank, kernel_size, stride, padding, bias=False) |
| 147 | self.B = nn.Conv2d(rank, out_channels, kernel_size=1, stride=1, padding=0, bias=False) |
| 148 | |
| 149 | nn.init.zeros_(self.B.weight) |
| 150 | nn.init.kaiming_normal_(self.A.weight, a=1) |
| 151 | |
| 152 | self.beta = nn.Conv2d(c_dim, rank, kernel_size=1, bias=False) |
| 153 | self.gamma = nn.Conv2d(c_dim, rank, kernel_size=1, bias=False) |
| 154 | |
| 155 | def forward(self, x): |
| 156 | """ |
| 157 | Args: |
| 158 | x (torch.Tensor): In shape of (B, C, H, W) |
| 159 | """ |
| 160 | w_out = self.W(x) |
| 161 | a_out = self.A(x) |
| 162 | |
| 163 | # inject conditioning into LoRA |
| 164 | c = self.data_provider.get_batch(a_out) |
| 165 | element_shift = self.beta(c) |
| 166 | element_scale = self.gamma(c) + 1 |
| 167 | a_cond = a_out * element_scale + element_shift |
| 168 | |
| 169 | b_out = self.B(a_cond) |
| 170 | |
| 171 | return w_out + b_out * self.lora_scale |