Compositional soft prompting (CSP), a parameter-efficient learning technique to improve the zero-shot compositionality of large-scale pretrained vision-language models (VLMs) without the overhead of fine-tuning the entire model.
Reference Paper: Learning to Compose Soft Prompts for Compositional Zero-Shot Learning
conda create --name clip python=3.7
conda activate clip
pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
pip3 install ftfy regex tqdm scipy pandas
pip3 install git+https://github.com/openai/CLIP.git
Alternatively, you can use pip install -r requirements.txt to install all the dependencies.
We experiment with three datasets: MIT-States, UT-Zappos, and C-GQA.
sh download_data.sh
If you already have setup the datasets, you can use symlink and ensure the following paths exist:
data/<dataset> where <datasets> = {'mit-states', 'ut-zappos', 'cgqa'}.
python -u train.py \
--dataset mit-states \
--clip_model ViT-L/14 \
--experiment_name csp \
--seed 0 \
--epochs 20 \
--lr 5e-05 \
--attr_dropout 0.3 \
--weight_decay 0.00001 \
--train_batch_size 64 \
--gradient_accumulation_steps 2 \
--context_length 8 \
--save_path data/model/mit-states/sample_model \
--save_every_n 1
You can replace --dataset with {mit-states, ut-zappos, cgqa}.
The best hyperparameters are included in the paper.
We evaluate our models in two settings: closed-world and open-world.
python -u evaluate.py \
--dataset mit-states \
--clip_model ViT-L/14 \
--soft_embeddings data/model/mit-states/sample_model/soft_embeddings_epoch_20.pt \
--context_length 16 \
--text_encoder_batch_size 36 \
--eval_batch_size 16 \
--experiment_name csp
You can change the experiment name to clip to benchmark the pretrained clip results.
For our open-world evaluation, we compute the feasbility calibration and then evaluate on the dataset.
We use GloVe embeddings to compute the similarities between objects and attributes.
Download the GloVe embeddings in the data directory:
cd data
wget https://nlp.stanford.edu/data/glove.6B.zip
Move glove.6B.300d.txt into data/glove.6B.300d.txt.
To compute feasibility calibration for each dataset, run the following command:
python -u datasets/feasibility.py --dataset mit-states
The feasibility similarities are saved at data/feasibility_<dataset>.pt.
The open-world evaluation with the thresholds (feasibility calibration).
python -u evaluate.py \
--dataset mit-states \
--soft_embeddings data/model/mit-states/sample_model/soft_embeddings_epoch_5.pt \
--context_length 16 \
--text_encoder_batch_size 36 \
--eval_batch_size 256 \
--experiment_name czsl \
--threshold <threshold> \
--open_world
If <threshold> is None, then the model picks the best threshold on the validation set.
We use the following thresholds:
| Dataset | Threshold |
|---|---|
| mit-states | 0.4069159426 |
| ut-zappos | 0.5299109123 |
| cgqa | 0.49937106273612186 |
Note: We use 256GB of cpu memory to evaluate cgqa.
Evaluate the trained CSP vocabulary on the new AAO-MIT-States dataset.
python aao/evaluate_att_att_obj.py \
--experiment_name csp \
--soft_embeddings data/model/mit-states/sample_model/soft_embeddings_epoch_20.pt
We thank Andrew Delworth and Elise Carman for helping us annotate this dataset.
Ablation experiment to train and evaluate CSP with reduced fine-tuned vocabulary. We run experiment on the ut-zappos dataset.
python -u mix/mix_train.py \
--dataset ut-zappos \
--model ViT-L/14 \
--experiment_name mix_csp \
--seed 0 \
--epochs 20 \
--lr 5e-04 \
--attr_dropout 0.2 \
--weight_decay 0.00001 \
--train_batch_size 64 \
--context_length 8 \
--save_path data/model/ut-zappos/mix_train_model_0.25 \
--save_every_n 5 \
--attr_keep_ratio 0.25 \
--gradient_accumulation_steps 2
We change the --attr_keep_ratio to {0.25, 0.50, 0.75}.
python -u mix/evaluate_mix_train.py \
--dataset ut-zappos \
--soft_embeddings data/model/ut-zappos/mix_train_model_0.25/soft_embeddings.pt \
--context_length 16 \
--text_encoder_batch_size 36 \
--eval_batch_size 256 \
--experiment_name csp
The project uses openly available model, code, and datasets. Please see the credits.
If you find CSP helpful, please cite our paper:
@inproceedings{
csp2023,
title={Learning to Compose Soft Prompts for Compositional Zero-Shot Learning},
author={Nihal V. Nayak and Peilin Yu and Stephen H. Bach},
booktitle={International Conference on Learning Representations},
year={2023}
}