Code for Revisiting Continual Ultra-fine-grained Visual Recognition with Pre-trained Models

Prerequest

1. Main dependecies

       torch==2.5.0
       torchvision==0.20.0
       cuda==12.1
       numpy=2.1.2
       timm==1.0.12

2. Preparing data

   • Download SoyGene, SoyAgeing and SoyGlobal UFG datasets.

   • Create folder Data/, unzip and place the data as

             Data
             ├── SoyAgeing
             │   └── ......
             ├── SoyGene
             │   └── ......
             └── SoyGlobal
                 └── ......

   • Formulate the raw data using

     python build_img_folder_structure.py Data/SoyAgeing/R1
     python build_img_folder_structure.py Data/SoyAgeing/R3
     python build_img_folder_structure.py Data/SoyAgeing/R4
     python build_img_folder_structure.py Data/SoyAgeing/R5
     python build_img_folder_structure.py Data/SoyAgeing/R6
     python build_img_folder_structure.py Data/SoyGene
     python build_img_folder_structure.py Data/SoyGlobal

Training and Inference

• Train any model with the config

  # for SoyGene-C
  python train_net_seq.py --config-file configs/ufgvc/${model}/${model}_base21k224_gene.yaml --resume
  # for SoyAgeing-C
  python train_net_seq.py --config-file configs/ufgvc/${model}/${model}_base21k224_ageing.yaml --resume
  # for SoyGlobal-C
  python train_net_seq.py --config-file configs/ufgvc/${model}/${model}_base21k224_global.yaml --resume
  # for UniUFG-C
  python train_net_seq.py --config-file configs/ufgvc/${model}/${model}_base21k224_gamixhalf.yaml --resume

• If a model is already trained, excecuting the above commands also performs full test of the model.

• All models are evaluated periodically during training, we recommend to use Tensorboard to monitor the training process and evaluate results:

  tensorboard --logdir outputs/${model_output}$

Supported Models

• Besides the proposed Unic, we also include the following previous continual learning models, making this repository a unified base framework for future research and development:

  • CODA-Prompt, DualPrompt, HiDE, L2P, LAE, S-Prompts, VQPrompt

• We also integrate basic PEFT methods, including

  • Prompt tuning, Prefix tuning, Adapter, Lora

Code structure

• The overall structure of the code is based on FastReID. This repository is organized as

  Unic
  ├── configs # Where model configuration files saved
  ├── Data # Where data saved 
  ├── fastclas # An extension of FastReID for classification tasks
  ├── fastreid # FastReID source code
  ├── outputs # Where training logs and checkpoints saved
  ├── ufgvc # The main code
  │   ├── config # Definition of the basic configuration for all models
  │   ├── data # Reading and augmenting data
  │   ├── engine # The main training loop
  │   ├── evaluation # Evaluators for testing the model 
  │   ├── __init__.py
  │   ├── modeling # Definition of models
  │   │   ├── meta_arch # Definition of continual learning models
  │   │   └── vits # Definition of ViTs with different PEFT methods
  │   └── utils # Commonly used tools, e.g. logging, visualization and distributed communication.
  ├── README.md # This file
  ├── build_img_folder_structure.py # data formulation tool
  ├── train_net_seq_mix.py # Training script for UniUFG-C
  └── train_net_seq_mix.py # Training script for other three datasets
  

Acknowledgement

This code is greatly inspired by FastReID and CODA-Prompt.