PTP

June 7, 2023 · View on GitHub

This repository includes implementations of the following method:

Position-guided Text Prompt for Vision Language Pre-training

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.

@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}
}

PTP