All our generated features are saved into the ./features
directory. We describe the three different sets of feature sets that are generated:
- Text classifier weights: For each dataset, we use three different prompt types for generating the classifier weights in our experiments using CLIP's text encoder: ensemble (from the ensemble prompts), cupl (from the customised GPT-3 prompts) and combined (from the concatenated cupl+ensemble prompts). Their naming is in the format:
<dataset>_zeroshot_text_weights_m<backbone>_pt<prompt_type>.pt
. - Test/Validation features: These are the test and validation set features along with their labels encoded using CLIP's image encoder. Their naming is in the format:
<dataset>_f_<test/val>_m<backbone>.pt
for features and<dataset>_t_<test/val>_m<backbone>.pt
for labels. - SuS features: These are the support set features (either using SuS-LC or SuS-SD) along with their labels encoded using CLIP's image encoder. Their naming is in the format:
sus_<sus_type>_<prompting_strategy>_<dataset>_f_m<backbone>.pt
for features andsus_<sus_type>_<prompting_strategy>_<dataset>_t_m<backbone>.pt
for labels.<sus_type>
islc
for SuS-LC andsd
for SuS-SD.<prompting_strategy>
isphoto
for Photo prompting strategy andcupl
for CuPL prompting strategy. Refer to Sec. 3.1 of paper for details.