(pixel_values)
| 1199 | first_epoch = 0 |
| 1200 | |
| 1201 | def encode_image(pixel_values): |
| 1202 | pixel_values = _resize_with_antialiasing(pixel_values, (224, 224)) |
| 1203 | pixel_values = (pixel_values + 1.0) / 2.0 |
| 1204 | |
| 1205 | pixel_values = pixel_values.to(torch.float32) |
| 1206 | # Normalize the image with for CLIP input |
| 1207 | pixel_values = feature_extractor( |
| 1208 | images=pixel_values, |
| 1209 | do_normalize=True, |
| 1210 | do_center_crop=False, |
| 1211 | do_resize=False, |
| 1212 | do_rescale=False, |
| 1213 | return_tensors="pt", |
| 1214 | ).pixel_values |
| 1215 | |
| 1216 | pixel_values = pixel_values.to( |
| 1217 | device=accelerator.device, dtype=image_encoder.dtype) |
| 1218 | image_embeddings = image_encoder(pixel_values).image_embeds |
| 1219 | image_embeddings = image_embeddings.unsqueeze(1) |
| 1220 | return image_embeddings |
| 1221 | |
| 1222 | def _get_add_time_ids( |
| 1223 | fps, |
no test coverage detected