Texts as Images in Prompt Tuning for Multi-Label Recognition
Introduction
This repo officially implements Texts as Images in Prompt Tuning for Multi-Label Recognition.
TaI-DPT explores the feasibility of prompting with text data for multi-label image recognition. Notable improvements are observed compared to zero-shot methods on multiple common multi-label benchmarks. For more details, please see the paper.
Contact us with zixian_guo@foxmail.com
Fig.1 Overview of Text-as-Image (TaI) prompting.
Install
The code is based largely on the implementation of CoOp and Dassl.
Please follow the steps below to build your environment.
# Create a conda environment (Omit if you already have a suitable environment)
conda create -n dassl python=3.7
conda activate dassl
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge # torch (version >= 1.7.1)
# Clone this repo
git clone https://github.com/guozix/TaI-DPT.git
cd TaI-DPT
# install Dassl
cd Dassl.pytorch-master/
# Install dependencies
pip install -r requirements.txt
# Install this library (no need to re-build if the source code is modified)
python setup.py develop
cd ..
# Install CLIP dependencies
pip install -r requirements.txt
# FinishedDatasets
We use captions from MS-COCO and localized narratives from OpenImages, and we evaluate our method on VOC2007, MS-COCO and NUS-WIDE. The directory structure is organized as follows.
The multi-label classification datasets can be accessed from their official websites VOC2007, MSCOCO2014 and NUS-WIDE. The raw images of NUS-WIDE can be accessed from here (verification code: s6oj).
For the OpenImages dataset, we only use the localized narratives of its "V6" version. The ~130MB jsonl file can be downloaded from the official site.
DATAROOT
├── OpenImages
│ ├── captions
│ │ └── open_images_train_v6_captions.jsonl
├── VOCdevkit
│ ├── VOC2007
| │ ├── Annotations
| │ ├── caption_data
| │ ├── ImageSets
| │ │ ├── Layout
| │ │ ├── Main
| │ │ └── Segmentation
| │ ├── JPEGImages
| │ ├── SegmentationClass
| │ └── SegmentationObject
├── COCO
│ ├── annotations
│ ├── train2014
│ └── val2014
└── NUSWIDE
├── ImageList
│ ├── Imagelist.txt
│ ├── TestImagelist.txt
│ └── TrainImagelist.txt
├── Flickr
│ ├── actor
│ ├── administrative_assistant
│ ├── adobehouses
│ ├── adult
│ ...
├── TrainTestLabels
└── Concepts81.txtUsage
To evaluate the results in Table 1.
Change the DATA variable in the bash scripts to the DATAROOT path above.
Test Baseline ZSCLIP on the datasets:
cd scripts/
bash zsclip.sh voc2007_distill rn50
bash zsclip.sh coco2014_distill rn50
bash zsclip.sh nuswide_distill_limit rn50The mAP score default term in the output denotes the mAP score calculated by using only the global classification logits, i.e. baseline CLIP.
The mAP score merged term denotes the performance of equally merging the logits from the glocal and the local branch, i.e. CLIP-DPT, similarly hereinafter.
Evaluate TaI-DPT:
Unzip the trained model weights in output.zip. The file directory should be like:
output
├── coco2014_caption
│ └── Caption_distill_double
│ └── rn50_coco2014
│ └── nctx16_cscFalse_ctpend
│ ├── seed1
│ ...
│ ├── seed2
│ ...
│ └── seed3
│ ...
├── nuswide_caption
│ └── Caption_distill_double
│ └── rn50_nuswide
│ └── nctx16_cscFalse_ctpend
│ ├── seed1
│ ...
│ ├── seed2
│ ...
│ └── seed3
│ ...
└── voc2007_caption
└── Caption_distill_double
└── rn50_voc2007
└── nctx16_cscFalse_ctpend
├── seed1
...
├── seed2
...
└── seed3
...Then run the evaluation script:
cd scripts/
bash main_eval.sh voc2007_distill rn50_voc2007 end 16 False voc2007_caption
bash main_eval.sh coco2014_distill rn50_coco2014 end 16 False coco2014_caption
bash main_eval.sh nuswide_distill_limit rn50_nuswide end 16 False nuswide_captionThe mAP score default and mAP score merged term in the output denotes mAP score of TaI and TaI-DPT. The results are also saved in log files in ./output/evaluation.
To reproduce the results in Table 1.
(If you extracted the trained model in the previous evaluation step, you need to set a different run_ID, otherwise the script will skip the training.)
Train TaI-DPT on the datasets:
cd scripts/
bash main.sh voc2007_distill rn50_voc2007 end 16 False voc2007_caption
bash main.sh coco2014_distill rn50_coco2014 end 16 False coco2014_caption
bash main.sh nuswide_distill_limit rn50_nuswide end 16 False nuswide_captionTo reproduce DualCoOp
cd scripts/
# VOC2007
bash main_dual.sh voc2007_partial rn101 end 16 True voc2007_partial_dualcoop_448_CSC_p0_1 0.1 0
bash main_dual.sh voc2007_partial rn101 end 16 True voc2007_partial_dualcoop_448_CSC_p0_2 0.2 0
bash main_dual.sh voc2007_partial rn101 end 16 True voc2007_partial_dualcoop_448_CSC_p0_3 0.3 0
...
# COCO2014
bash main_dual.sh coco2014_partial rn101 end 16 True coco2014_partial_dualcoop_448_CSC_p0_1 0.1 1
...
# NUS-WIDE
bash main_dual.sh nuswide_partial rn101_nus end 16 True nuswide_partial_dualcoop_448_CSC_p0_1 0.1 2
...
To reproduce ensemble results
The parameter setting here is a little cumbersome, but the logic is simple. You only need to specify two trained model paths and the corresponding model configurations.
Here is an example of ensemble of TaI-DPT and DualCoOp on VOC2007. You should finish training TaI-DPT and DualCoOp (under a specific partial_prob) on VOC2007 before.
cd scripts/
bash ensemble.sh voc2007_distill rn50_voc2007 end 16 False voc2007_caption_e tmp1.pkl \
output/voc2007_caption/Caption_distill_double/rn50_voc2007/nctx16_cscFalse_ctpend/seed1 \
voc2007_partial rn101 end 16 True voc2007_partial_dualcoop_448_CSC_p0_5_e tmp2.pkl \
output/voc2007_partial_dualcoop_448_CSC_p0_5/Caption_dual/rn101/nctx16_cscTrue_ctpend/seed1 0.9Thanks
We use code from CoOp and Dassl, which are great repositories and we encourage you to check them out and cite them in your work.