This is the official repository for our paper TAP: TARGETED PROMPTING FOR TASK ADAPTIVE GENERATION OF TEXTUAL TRAINING INSTANCES FOR VISUAL CLASSIFICATION. We provide the code for reproducing the results for all the 8 datasets used in our paper.

Our code is built upon the official codebase of the CoOp paper and has been tested in an environment having python 3.8.8 and pytorch 2.0.1 compiled with CUDA 11.6.

As a first step, install dassl library (under TAP/) in your environment by following the instructions here.

To further install all other dependencies, please run the following command, after having your environment activated:

pip install -r requirements.txt

Please download and structure your datasets according to the instructions provided in the CoOp official repository. All the 8 datasets should be present in the data/ directory.

The generic and dataset specific descriptions for all the 8 datasets are present in the descriptions/ directory.

To reproduce the results for TAP all the 8 datasets in Table 1, please run the following command:

bash scripts/tap.sh <dataset_name>

where <dataset_name> can be one of dtd oxford_flowers imagenet_r fgvc_aircraft food101 eurosat ucf101 sun397

Similarly, to obtain zero-shot CLIP results with the single prompt template a photo of a {category}. Please run:

bash scripts/zeroshot.sh <dataset_name>

by replacing the <dataset_name> with one of the 8 datasets mentioned above.