| 43 | |
| 44 | |
| 45 | class TemporalResnetBlock(torch.nn.Module): |
| 46 | |
| 47 | def __init__(self, in_channels, out_channels, groups=32, eps=1e-5): |
| 48 | super().__init__() |
| 49 | self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True) |
| 50 | self.conv1 = torch.nn.Conv3d(in_channels, out_channels, kernel_size=(3, 1, 1), stride=1, padding=(1, 0, 0)) |
| 51 | self.norm2 = torch.nn.GroupNorm(num_groups=groups, num_channels=out_channels, eps=eps, affine=True) |
| 52 | self.conv2 = torch.nn.Conv3d(out_channels, out_channels, kernel_size=(3, 1, 1), stride=1, padding=(1, 0, 0)) |
| 53 | self.nonlinearity = torch.nn.SiLU() |
| 54 | self.mix_factor = torch.nn.Parameter(torch.Tensor([0.5])) |
| 55 | |
| 56 | def forward(self, hidden_states, time_emb, text_emb, res_stack, **kwargs): |
| 57 | x_spatial = hidden_states |
| 58 | x = rearrange(hidden_states, "T C H W -> 1 C T H W") |
| 59 | x = self.norm1(x) |
| 60 | x = self.nonlinearity(x) |
| 61 | x = self.conv1(x) |
| 62 | x = self.norm2(x) |
| 63 | x = self.nonlinearity(x) |
| 64 | x = self.conv2(x) |
| 65 | x_temporal = hidden_states + x[0].permute(1, 0, 2, 3) |
| 66 | alpha = torch.sigmoid(self.mix_factor) |
| 67 | hidden_states = alpha * x_temporal + (1 - alpha) * x_spatial |
| 68 | return hidden_states, time_emb, text_emb, res_stack |
| 69 | |
| 70 | |
| 71 | class SVDVAEDecoder(torch.nn.Module): |