PTP
This repository includes implementations of the following method:
Introduction
The goal of Position-guided Text Prompt (PTP) is to bring position information into conventional Vision-Language Pre-training (VLP) models, as current mainstream e2e VLP models ignore this important cues.
We observe Position information is missed in a well-trained ViLT models.
Our method provide a good altentive for existing object feature based methods (BUTD and the following works).
Some examples of one PTP is show below:
Updates
- 2023.5 Modify the pre-training corpus to prevent confusing.
- 2023.3 The Pre-training Code is released.
- 2023.1 We have put the pretrained and fine-tuned weight on huggingface for fast download.
- 2022.12 The first version of downstream evaluation code based on BLIP and pretrained/down-stream weight is released! The pre-training code is in cleaning up now.
Installation
Please find installation instructions for PyTorch in INSTALL.md.
Dataset Preparation
You may follow the instructions in DATASET.md to prepare the datasets. Considering the dataset prepartion is very time consuming, we provide detail guidence and provided our trained corpus.
Pretrained & Finetune Models
1. Pre-trained Model
Method | Vision Encoder | #Images | Dataset | Pretrained Weights | Training Logs |
---|---|---|---|---|---|
PTP-BLIP | ViT-B(DeiT) | 4M | CC3M+COCO+VG+SBU | link | link |
2. Zero-shot & Fine-tuning Downstream Model
2.1 Captioning
Method | B@4 | CIDEr | Config |
---|---|---|---|
PTP-BLIP | 40.1 | 135.0 | configs/caption_coco.yaml |
2.2 Zero-shot Retrieval
2.2.2 Flickr30K
Method | I2T@1 | T2I@1 | Model Weight | Training Logs | Config |
---|---|---|---|---|---|
PTP-BLIP | 86.4 | 67.0 | link | link | configs/retrieval_flickr.yaml |
2.3 Retrieval (Fine-tune)
Tip: Please use as large batch size as possible, we experimentally find that the larger batch size leads to better result for this task. Due to memory limiation, we use batch size 24 rather than 28 in original implmentation.
2.3.1 COCO
Method | I2T@1 | T2I@1 | Config |
---|---|---|---|
PTP-BLIP | 77.6 | 59.4 | configs/retrieval_coco.yaml |
2.3.2 Flickr30K
Method | I2T@1 | T2I@1 | Model Weight | Training Logs | Config |
---|---|---|---|---|---|
PTP-BLIP | 96.1 | 84.2 | link | link | configs/retrieval_flickr.yaml |
2.4 VQA V2
Method | Test-dev | Test-std | Model Weight | Training Logs | Config |
---|---|---|---|---|---|
PTP-BLIP | 76.02 | 76.18 | link | link | configs/vqa.yaml |
2.5 NLVR
Method | Dev | Test-P | Model Weight | Training Logs | Config |
---|---|---|---|---|---|
PTP-BLIP | 80.45 | 80.70 | link | link | configs/nlvr.yaml |
Quick Start
Follow the example in GETTING_STARTED.md to start playing vlp models with PTP.
Transfer To Other Architectures
The PTP can easily transfer to other architectures without much effort. Specifically, change your base code with following two steps:
- Download or generate corpus in the same format as ours.
- Modify the dataset.py
Then train the model with original objectives.
Ackowledgement
This work is mainly based on BLIP and ViLT, thanks for these good baselines. We also refer OSCAR for ablation study and dataset preparation.
License
PTP is released under the Apache 2.0 license.
Contact
Email: awinyimgprocess at gmail dot com
If you have any questions, please email me or open an new issue.
Citation
If you find our work helps, please use the following BibTeX entry for citation.
@article{wang2022ptp,
title={Position-guided Text Prompt for Vision Language Pre-training},
author={Wang, Alex Jinpeng and Zhou, Pan and Shou, Mike Zheng and Yan, Shui Cheng},
journal={https://arxiv.org/abs/2212.09737},
year={2022}
}