| 90 | """ |
| 91 | |
| 92 | def __init__( |
| 93 | self, |
| 94 | channels: int, |
| 95 | use_conv: bool = False, |
| 96 | use_conv_transpose: bool = False, |
| 97 | out_channels: Optional[int] = None, |
| 98 | name: str = "conv", |
| 99 | kernel_size: Optional[int] = None, |
| 100 | padding=1, |
| 101 | norm_type=None, |
| 102 | eps=None, |
| 103 | elementwise_affine=None, |
| 104 | bias=True, |
| 105 | interpolate=True, |
| 106 | ): |
| 107 | super().__init__() |
| 108 | self.channels = channels |
| 109 | self.out_channels = out_channels or channels |
| 110 | self.use_conv = use_conv |
| 111 | self.use_conv_transpose = use_conv_transpose |
| 112 | self.name = name |
| 113 | self.interpolate = interpolate |
| 114 | conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv |
| 115 | |
| 116 | if norm_type == "ln_norm": |
| 117 | self.norm = nn.LayerNorm(channels, eps, elementwise_affine) |
| 118 | elif norm_type == "rms_norm": |
| 119 | self.norm = RMSNorm(channels, eps, elementwise_affine) |
| 120 | elif norm_type is None: |
| 121 | self.norm = None |
| 122 | else: |
| 123 | raise ValueError(f"unknown norm_type: {norm_type}") |
| 124 | |
| 125 | conv = None |
| 126 | if use_conv_transpose: |
| 127 | if kernel_size is None: |
| 128 | kernel_size = 4 |
| 129 | conv = nn.ConvTranspose2d( |
| 130 | channels, self.out_channels, kernel_size=kernel_size, stride=2, padding=padding, bias=bias |
| 131 | ) |
| 132 | elif use_conv: |
| 133 | if kernel_size is None: |
| 134 | kernel_size = 3 |
| 135 | conv = conv_cls(self.channels, self.out_channels, kernel_size=kernel_size, padding=padding, bias=bias) |
| 136 | |
| 137 | # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed |
| 138 | if name == "conv": |
| 139 | self.conv = conv |
| 140 | else: |
| 141 | self.Conv2d_0 = conv |
| 142 | |
| 143 | def forward( |
| 144 | self, |