(self, sample, tiled=False, tile_size=64, tile_stride=32, **kwargs)
| 50 | return hidden_states |
| 51 | |
| 52 | def forward(self, sample, tiled=False, tile_size=64, tile_stride=32, **kwargs): |
| 53 | # For VAE Decoder, we do not need to apply the tiler on each layer. |
| 54 | if tiled: |
| 55 | return self.tiled_forward(sample, tile_size=tile_size, tile_stride=tile_stride) |
| 56 | |
| 57 | # 1. pre-process |
| 58 | hidden_states = self.conv_in(sample) |
| 59 | time_emb = None |
| 60 | text_emb = None |
| 61 | res_stack = None |
| 62 | |
| 63 | # 2. blocks |
| 64 | for i, block in enumerate(self.blocks): |
| 65 | hidden_states, time_emb, text_emb, res_stack = block(hidden_states, time_emb, text_emb, res_stack) |
| 66 | |
| 67 | # 3. output |
| 68 | hidden_states = self.conv_norm_out(hidden_states) |
| 69 | hidden_states = self.conv_act(hidden_states) |
| 70 | hidden_states = self.conv_out(hidden_states) |
| 71 | hidden_states = hidden_states[:, :16] |
| 72 | hidden_states = (hidden_states - self.shift_factor) * self.scaling_factor |
| 73 | |
| 74 | return hidden_states |
| 75 | |
| 76 | def encode_video(self, sample, batch_size=8): |
| 77 | B = sample.shape[0] |
no test coverage detected