LPT

The official code of Long-tailed Prompt Tuning

Our code is based on the unofficial VPT code implemented by DongSky.

This repository will be updated continuously in the near future.

Preparing Data

Places-LT

Download the original Places365 standard dataset from here, and then change the path of Places-LT in datasets.py by the current root path of places365standard.

Note that we have stored the train/val/test split of Places-LT in vtab directory (move into phase2 test directory and you will see this dir).

Testing LPT

Here we present LPT trained on Places-LT dataset.

Note that for simplicity during experiments, I stored the whole model into storage... The final size of LPT checkpoint may be slightly larger (negligible) than standard ViT.

LPT (Places-LT): Google Drive

Set the checkpoint to the Phase2 test directory, and then execute the following commands:

CUDA_VISIBLE_DEVICES=0 python eval_phase2.py --dataset places365 --split full

You will obtain:

epoch 1, overall: 50.07123%, many-shot: 49.26718%, medium-shot: 52.30573%, few-shot: 46.88312%

TODO

• Training code
• More checkpoints

Give me some time to prepare the code QAQ.