Embracing Uncertainty: The Impact of Fine-tuning Paradigms on Unknown Plant Diseases Recognition

• This code is an implementation of our manuscript: Embracing Uncertainty: The Impact of Fine-tuning Paradigms on Unknown Plant Diseases Recognition.

• Authors: Jiuqing Dong, et al.

• We will release the complete code and complete this documentation after the article is published.

• please refer more results in 'Result.xlsx'.

Installation

Please check env_setup.sh.

Pre-trained model preparation

Download and place the pre-trained Transformer-based backbones to './models/'. In our study, we use the ViT-Base pre-trained on Imagenet-21k.

Dataset Prepairation

Cotton disease dataset: https://www.kaggle.com/datasets/dhamur/cotton-plant-disease

Mango disease dataset: https://www.kaggle.com/datasets/aryashah2k/mango-leaf-disease-dataset

Strawberry disease dataset: https://www.kaggle.com/datasets/usmanafzaal/strawberry-disease-detection-dataset

Tomato disease dataset and Plant village dataset: https://github.com/spMohanty/PlantVillage-Dataset/tree/master/raw

please split the dataset by using our code.

Training and Test

We have provided the training scripts for VPT, FFT, VAT, and LPT in train_train0.sh, train_train1.sh, train_train2.sh, and train_train3.sh, respectively. We use VPT as an example. Run CUDA_VISIBLE_DEVICES=0 sh train_gpu0.sh.

Key configs:

python train.py \                                       # For Stage 1
      --train-type "prompt"  \                          # VPT method
      --config-file configs/prompt/plantvillage.yaml  \   # the configuration of Imagenet-1k dataset
      DATA.PERCENTAGE '0.1' \                           # For the plant village dataset, 0.1, 0.2, 0.3, 0.4 correspond to 2, 4, 8, and 16 shots, respectively. While, for other datasets, it denotes experiments ID.
      DATA.FEATURE "sup_vitb16_imagenet21k" \           # specify which representation to use
      MODEL.TYPE "vit" \                                # the general backbone type, e.g., "vit" or "swin"
      DATA.BATCH_SIZE "128" \
      SOLVER.BASE_LR "0.1" \                            # learning rate for the experiment
      SOLVER.WEIGHT_DECAY "0.0" \                       # weight decay value for the experiment
      MODEL.PROMPT.DEEP "True" \                        # deep or shallow prompt
      MODEL.PROMPT.DROPOUT "0.1" \
      MODEL.PROMPT.NUM_TOKENS "10" \                    # prompt length
      MODEL.MODEL_ROOT "models/" \                      # folder with pre-trained model checkpoints
      OUTPUT_DIR "./FT_ood" \                         # output dir of the final model and logs
      SOLVER.TOTAL_EPOCH '100' \                    
      # SOLVER.TOTAL_EPOCH '0' \                        # for test phase
      # MODEL.WEIGHT_PATH "./FT_ood/plantvillage/sup_vitb16_imagenet21k/lr2.5_wd0.001/run1/val_Imagenet1k_best_model.pth" \

Train

sh train.sh

We will release the complete code and complete this documentation after our manuscript is accepted.

How to use this code for a customer dataset?