Repository navigation
Download and extract
Download all pretrained_models.tar.gz.part-* files into the same directory, then reconstruct and extract the archive:
cat pretrained_models.tar.gz.part-* > pretrained_models.tar.gz
tar -xzf pretrained_models.tar.gzThis creates the CIRR/ and FashionIQ/ folders.
Included checkpoints
Each dataset folder contains the following configurations:
| Folder | CLIP backbone | Encoder fine-tuning | Reference |
|---|---|---|---|
RN50_noft |
RN50 | None | CVPR 2022 |
RN50x4_noft |
RN50x4 | None | CVPR 2022 |
RN50_textft |
RN50 | Text encoder only | CVPRW 2022 |
RN50_imageft |
RN50 | Image encoder only | ACM TOMM 23 |
RN50_fullft |
RN50 | Image and text encoders | ACM TOMM 23 |
RN50x4_fullft |
RN50x4 | Image and text encoders | ACM TOMM 23 |
Files prefixed with cirr_ are trained on CIRR; files prefixed with fiq_ are trained on FashionIQ. Filenames containing _clip_ hold the fine-tuned CLIP weights, while _comb_ files hold the trained Combiner weights.
Use CLIP and Combiner checkpoints from the same dataset and configuration. The noft configurations include only a Combiner checkpoint and use the corresponding original CLIP backbone without fine-tuning.