x: A list of videos each with shape [C, T, H, W]. t: [B]. context: A list of text embeddings each with shape [L, C].
(
self,
x,
t,
context,
seq_len,
y=None,
dit_cond_dict=None,
)
| 96 | |
| 97 | |
| 98 | def sp_dit_forward( |
| 99 | self, |
| 100 | x, |
| 101 | t, |
| 102 | context, |
| 103 | seq_len, |
| 104 | y=None, |
| 105 | dit_cond_dict=None, |
| 106 | ): |
| 107 | """ |
| 108 | x: A list of videos each with shape [C, T, H, W]. |
| 109 | t: [B]. |
| 110 | context: A list of text embeddings each with shape [L, C]. |
| 111 | """ |
| 112 | if self.model_type == 'i2v': |
| 113 | assert y is not None |
| 114 | # params |
| 115 | device = self.patch_embedding.weight.device |
| 116 | if self.freqs.device != device: |
| 117 | self.freqs = self.freqs.to(device) |
| 118 | |
| 119 | if y is not None: |
| 120 | x = [torch.cat([u, v], dim=0) for u, v in zip(x, y)] |
| 121 | |
| 122 | # embeddings |
| 123 | x = [self.patch_embedding(u.unsqueeze(0)) for u in x] |
| 124 | grid_sizes = torch.stack( |
| 125 | [torch.tensor(u.shape[2:], dtype=torch.long) for u in x]) |
| 126 | x = [u.flatten(2).transpose(1, 2) for u in x] |
| 127 | seq_lens = torch.tensor([u.size(1) for u in x], dtype=torch.long) |
| 128 | assert seq_lens.max() <= seq_len |
| 129 | x = torch.cat([ |
| 130 | torch.cat([u, u.new_zeros(1, seq_len - u.size(1), u.size(2))], dim=1) |
| 131 | for u in x |
| 132 | ]) |
| 133 | |
| 134 | # time embeddings |
| 135 | if t.dim() == 1: |
| 136 | t = t.expand(t.size(0), seq_len) |
| 137 | with torch.amp.autocast('cuda', dtype=torch.float32): |
| 138 | bt = t.size(0) |
| 139 | t = t.flatten() |
| 140 | e = self.time_embedding( |
| 141 | sinusoidal_embedding_1d(self.freq_dim, |
| 142 | t).unflatten(0, (bt, seq_len)).float()) |
| 143 | e0 = self.time_projection(e).unflatten(2, (6, self.dim)) |
| 144 | assert e.dtype == torch.float32 and e0.dtype == torch.float32 |
| 145 | |
| 146 | # context |
| 147 | context_lens = None |
| 148 | context = self.text_embedding( |
| 149 | torch.stack([ |
| 150 | torch.cat([u, u.new_zeros(self.text_len - u.size(0), u.size(1))]) |
| 151 | for u in context |
| 152 | ])) |
| 153 | |
| 154 | # cam |
| 155 | if dit_cond_dict is not None and "c2ws_plucker_emb" in dit_cond_dict: |
nothing calls this directly
no test coverage detected