cvpr-2023-track2-8th's Introduction

CVPR 2023第一届大模型比赛Track2 第8名方案

We follow the code described in the CVPR2023 paper titled "Cross-Modal Implicit Relation Reasoning and Aligning for Text-to-Image Person Retrieval"

Requirements

Usage

Requirements

we use single RTX3090 24G GPU for training and evaluation.

pytorch 1.9.0
torchvision 0.10.0
prettytable
easydict

Prepare Datasets

Organize your data folder as follows:

|-- data/
|       |-- train
|           |-- train_images/
|           |-- train_label.txt
|       |-- test
|           |-- test_images/
|           |-- test_label.txt
|       |-- val
|           |-- val_images/
|           |-- val_label.txt
|-- logs/
|       |-- retrival1
|       |-- retrival2

The logs can be download from: 链接：https://pan.baidu.com/s/1pcxTXMNkKEaQoXGTPWT7qg 提取码：0wfp

Training

CUDA_VISIBLE_DEVICES=0 python train.py \
--name iira \
--img_aug \
--batch_size 96 \
--MLM \
--loss_names 'sdm+id' \
--dataset_name 'ImageRetri' \
--root_dir './data' \
--num_epoch 30 \
--output_dir logs/retrival1

CUDA_VISIBLE_DEVICES=0 python train.py \
--name iira \
--img_aug \
--batch_size 96 \
--MLM \
--loss_names 'sdm+id+oim' \
--dataset_name 'ImageRetri' \
--root_dir './data' \
--num_epoch 30 \
--output_dir logs/retrieval2

Testing

CUDA_VISIBLE_DEVICES=0 python inference_person.py --config_file ./logs/retrival1/configs.yaml
CUDA_VISIBLE_DEVICES=0 python inference_person.py --config_file ./logs/retrival2/configs.yaml
python ensemble.py

Recommend Projects