Spaces:
Running on Zero
Running on Zero
Download diffcve/models/prior_encoder.py from hugging-apps/diffcve-2607: direct link, hf CLI and curl.
- Browser
- Download file 2.52 kB
-
https://huggingface.co/spaces/hugging-apps/diffcve-2607/resolve/main/diffcve/models/prior_encoder.py
- Command line
-
hf download hf://spaces/hugging-apps/diffcve-2607/diffcve/models/prior_encoder.py
-
curl -L -o prior_encoder.py https://huggingface.co/spaces/hugging-apps/diffcve-2607/resolve/main/diffcve/models/prior_encoder.py
2.52 kB
| import torch | |
| import torch.nn as nn | |
| class DownBlock(nn.Module): | |
| def __init__(self, in_ch, out_ch): | |
| super().__init__() | |
| self.conv = nn.Conv2d( | |
| in_ch, out_ch, | |
| kernel_size=3, | |
| stride=2, | |
| padding=1 | |
| ) | |
| def forward(self, x): | |
| return self.conv(x) | |
| class ResidualConvBlock(nn.Module): | |
| def __init__(self, ch, use_residual=True): | |
| super().__init__() | |
| self.use_residual = use_residual | |
| self.conv1 = nn.Conv2d(ch, ch, 3, padding=1) | |
| self.norm1 = nn.GroupNorm(max(1, ch//8), ch) | |
| self.act = nn.SiLU() | |
| def forward(self, x): | |
| identity = x | |
| out = self.norm1(x) | |
| out = self.act(out) | |
| out = self.conv1(out) | |
| if self.use_residual: | |
| out = out + identity | |
| return out | |
| class ResidualMotionEncoder(nn.Module): | |
| def __init__( | |
| self, | |
| res_in_channels=1, | |
| mv_in_channels=2, | |
| res_initial_ch=12, | |
| mv_initial_ch=20, | |
| output_ch=256, | |
| use_residual=True, | |
| ): | |
| super().__init__() | |
| self.use_residual = use_residual | |
| self.res_conv = nn.Conv2d(res_in_channels, res_initial_ch, 3, 1, 1) | |
| self.mv_conv = nn.Conv2d(mv_in_channels, mv_initial_ch, 3, 1, 1) | |
| self.act = nn.SiLU() | |
| merged_ch = res_initial_ch + mv_initial_ch | |
| self.merge_conv = nn.Conv2d(merged_ch, merged_ch, 3, 1, 1) | |
| self.merge_norm = nn.GroupNorm(max(1, merged_ch//8), merged_ch) | |
| self.down1 = DownBlock(merged_ch, merged_ch * 2) | |
| self.conv1 = ResidualConvBlock(merged_ch * 2, use_residual) | |
| self.down2 = DownBlock(merged_ch * 2, merged_ch * 4) | |
| self.conv2 = ResidualConvBlock(merged_ch * 4, use_residual) | |
| self.down3 = DownBlock(merged_ch * 4, merged_ch * 8) | |
| self.conv3 = ResidualConvBlock(output_ch, use_residual) | |
| def forward(self, res, mv, return_features=False): | |
| h_res = self.res_conv(res) | |
| h_mv = self.mv_conv(mv) | |
| h = torch.cat([h_res, h_mv], dim=1) | |
| h = self.merge_norm(h) | |
| h = self.act(h) | |
| h = self.merge_conv(h) | |
| fea_list = [] | |
| h = self.down1(h) | |
| h = self.conv1(h) | |
| if return_features: | |
| fea_list.append(h) | |
| h = self.down2(h) | |
| h = self.conv2(h) | |
| if return_features: | |
| fea_list.append(h) | |
| h = self.down3(h) | |
| h = self.conv3(h) | |
| if return_features: | |
| return h, fea_list | |
| return h | |