| 252 | |
| 253 | |
| 254 | class Encoder(nn.Module): |
| 255 | def __init__( |
| 256 | self, |
| 257 | *, |
| 258 | ch, |
| 259 | out_channels, |
| 260 | ch_mult=(1, 2, 4, 8), |
| 261 | num_res_blocks, |
| 262 | attn_resolutions, |
| 263 | dropout=0.0, |
| 264 | resamp_with_conv=True, |
| 265 | in_channels, |
| 266 | resolution, |
| 267 | z_channels, |
| 268 | use_linear_attn=False, |
| 269 | attn_type="vanilla", |
| 270 | **ignore_kwargs |
| 271 | ): |
| 272 | super(Encoder, self).__init__() |
| 273 | if use_linear_attn: |
| 274 | attn_type = "linear" |
| 275 | self.ch = ch |
| 276 | self.temb_ch = 0 |
| 277 | self.num_resolutions = len(ch_mult) |
| 278 | self.num_res_blocks = num_res_blocks |
| 279 | self.resolution = resolution |
| 280 | self.in_channels = in_channels |
| 281 | |
| 282 | # Downsampling |
| 283 | self.conv_in = nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1) |
| 284 | |
| 285 | curr_res = resolution |
| 286 | in_ch_mult = (1,) + tuple(ch_mult) |
| 287 | self.in_ch_mult = in_ch_mult |
| 288 | self.down = nn.ModuleList() |
| 289 | for i_level in range(self.num_resolutions): |
| 290 | block = nn.ModuleList() |
| 291 | attn = nn.ModuleList() |
| 292 | block_in = ch * in_ch_mult[i_level] |
| 293 | block_out = ch * ch_mult[i_level] |
| 294 | for i_block in range(self.num_res_blocks): |
| 295 | block.append( |
| 296 | ResnetBlock( |
| 297 | in_channels=block_in, |
| 298 | out_channels=block_out, |
| 299 | temb_channels=self.temb_ch, |
| 300 | dropout=dropout |
| 301 | ) |
| 302 | ) |
| 303 | block_in = block_out |
| 304 | if curr_res in attn_resolutions: |
| 305 | attn.append(make_attn(block_in, attn_type=attn_type)) |
| 306 | down = nn.Module() |
| 307 | down.block = block |
| 308 | down.attn = attn |
| 309 | if i_level != self.num_resolutions - 1: |
| 310 | down.downsample = Downsample(block_in, resamp_with_conv) |
| 311 | curr_res = curr_res // 2 |
nothing calls this directly
no outgoing calls
no test coverage detected