(self,
dim_in: List[Optional[int]],
dim_res_blocks: List[int],
dim_out: List[Optional[int]],
resamplers: Union[Literal['pixel_shuffle', 'nearest', 'bilinear', 'conv_transpose', 'pixel_unshuffle', 'avg_pool', 'max_pool'], List],
dim_times_res_block_hidden: int = 1,
num_res_blocks: int = 1,
res_block_in_norm: Literal['layer_norm', 'group_norm' , 'instance_norm', 'none'] = 'layer_norm',
res_block_hidden_norm: Literal['layer_norm', 'group_norm' , 'instance_norm', 'none'] = 'group_norm',
activation: Literal['relu', 'leaky_relu', 'silu', 'elu'] = 'relu',
)
| 185 | |
| 186 | class ConvStack(nn.Module): |
| 187 | def __init__(self, |
| 188 | dim_in: List[Optional[int]], |
| 189 | dim_res_blocks: List[int], |
| 190 | dim_out: List[Optional[int]], |
| 191 | resamplers: Union[Literal['pixel_shuffle', 'nearest', 'bilinear', 'conv_transpose', 'pixel_unshuffle', 'avg_pool', 'max_pool'], List], |
| 192 | dim_times_res_block_hidden: int = 1, |
| 193 | num_res_blocks: int = 1, |
| 194 | res_block_in_norm: Literal['layer_norm', 'group_norm' , 'instance_norm', 'none'] = 'layer_norm', |
| 195 | res_block_hidden_norm: Literal['layer_norm', 'group_norm' , 'instance_norm', 'none'] = 'group_norm', |
| 196 | activation: Literal['relu', 'leaky_relu', 'silu', 'elu'] = 'relu', |
| 197 | ): |
| 198 | super().__init__() |
| 199 | self.input_blocks = nn.ModuleList([ |
| 200 | nn.Conv2d(dim_in_, dim_res_block_, kernel_size=1, stride=1, padding=0) if dim_in_ is not None else nn.Identity() |
| 201 | for dim_in_, dim_res_block_ in zip(dim_in if isinstance(dim_in, Sequence) else itertools.repeat(dim_in), dim_res_blocks) |
| 202 | ]) |
| 203 | self.resamplers = nn.ModuleList([ |
| 204 | Resampler(dim_prev, dim_succ, scale_factor=2, type_=resampler) |
| 205 | for i, (dim_prev, dim_succ, resampler) in enumerate(zip( |
| 206 | dim_res_blocks[:-1], |
| 207 | dim_res_blocks[1:], |
| 208 | resamplers if isinstance(resamplers, Sequence) else itertools.repeat(resamplers) |
| 209 | )) |
| 210 | ]) |
| 211 | self.res_blocks = nn.ModuleList([ |
| 212 | nn.Sequential( |
| 213 | *( |
| 214 | ResidualConvBlock( |
| 215 | dim_res_block_, dim_res_block_, dim_times_res_block_hidden * dim_res_block_, |
| 216 | activation=activation, in_norm=res_block_in_norm, hidden_norm=res_block_hidden_norm |
| 217 | ) for _ in range(num_res_blocks[i] if isinstance(num_res_blocks, list) else num_res_blocks) |
| 218 | ) |
| 219 | ) for i, dim_res_block_ in enumerate(dim_res_blocks) |
| 220 | ]) |
| 221 | self.output_blocks = nn.ModuleList([ |
| 222 | nn.Conv2d(dim_res_block_, dim_out_, kernel_size=1, stride=1, padding=0) if dim_out_ is not None else nn.Identity() |
| 223 | for dim_out_, dim_res_block_ in zip(dim_out if isinstance(dim_out, Sequence) else itertools.repeat(dim_out), dim_res_blocks) |
| 224 | ]) |
| 225 | |
| 226 | def enable_gradient_checkpointing(self): |
| 227 | for i in range(len(self.resamplers)): |
nothing calls this directly
no test coverage detected