search

MosyMosy / Feature-Normalization

The "Revisiting Learnable Affines for Batch Norm in Few-Shot Transfer Learning" Code-base
MIT License
5 stars 1 forks source link

readme

The "Revisiting Learnable Affines for Batch Norm in Few-Shot Transfer Learning" Code-base

Introduction

Requirements

This codebase is tested with:

  1. h5py==3.1.0
  2. joypy==0.2.5
  3. matplotlib==3.4.2
  4. numpy==1.21.0
  5. pandas==1.2.3
  6. Pillow==8.4.0
  7. scikit_learn==1.0.1
  8. scipy==1.6.0
  9. seaborn==0.11.2
  10. torch==1.8.1
  11. torchvision==0.9.1
  12. tqdm==4.60.0

To install all requirements, use "pip install -r requirements.txt"

Running Experiments

Dataset Preparation

MiniImageNet and CD-FSL: Download the datasets for CD-FSL benchmark following step 1 and step 2 here: https://github.com/IBM/cdfsl-benchmark

ImageNet: https://www.kaggle.com/c/imagenet-object-localization-challenge/data

Set datasets path: Set the appropriate dataset pathes in "configs.py".

Source dataset names: "ImageNet", "miniImageNet"

Target dataset names: "EuroSAT", "CropDisease", "ChestX", "ISIC"

All the dataset train/validation split files located at "datasets/split_seed_1" directory

All Baseline (miniImageNet) are trained Using an adapted version of the "https://github.com/cpphoo/STARTUP" repository

Baseline BN (Table 2)
To Train: https://github.com/MosyMosy/STARTUP/tree/main/teacher_miniImageNet
To Fine-Tune: python finetune.py --save_dir ./logs/baseline_teacher --target_dataset {Target dataset name} --subset_split datasets/split_seed_1/{Target dataset name} _labeled_80.csv --embedding_load_path ./logs/baseline_teacher/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary
Baseline FN (Table 2)
To Train: https://github.com/MosyMosy/STARTUP/tree/main/teacher_miniImageNet_na
To Fine-Tune: python finetune.py --save_dir ./logs/baseline_na_teacher --target_dataset {Target dataset name} --subset_split datasets/split_seed_1/{Target dataset name} _labeled_80.csv --embedding_load_path ./logs/baseline_na_teacher/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary
AdaBN BN (Table 2)
To Train: python AdaBN.py --dir ./logs/AdaBN/{dataset Name} --base_dictionary logs/baseline_teacher/checkpoint_best.pkl --target_dataset $target_testset --target_subset_split datasets/split_seed_1/$target_testset_unlabeled_20.csv --bsize 256 --epochs 10 --model resnet10 &
To Fine-Tune: python finetune.py --save_dir ./logs/AdaBN/{Target dataset name} --target_dataset {Target dataset name} --subset_split datasets/split_seed_1/{Target dataset name} _labeled_80.csv --embedding_load_path ./logs/AdaBN/{Target dataset name}/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary (replace {Target_dataset_name}): https://github.com/MosyMosy/FN_Model_Zoo/tree/main/Dictionaries/AdaBN/{Target_dataset_name}/checkpoint_best.pkl
AdaBN FN (Table 2)
To Train: python AdaBN_na.py --dir ./logs/AdaBN_na/{dataset Name} --base_dictionary logs/baseline_na_teacher/checkpoint_best.pkl --target_dataset $target_testset --target_subset_split datasets/split_seed_1/$target_testset_unlabeled_20.csv --bsize 256 --epochs 10 --model resnet10
To Fine-Tune: python finetune.py --save_dir ./logs/AdaBN_na/{Target dataset name} --target_dataset {Target dataset name} --subset_split datasets/split_seed_1/{Target dataset name} _labeled_80.csv --embedding_load_path ./logs/AdaBN_na/{Target dataset name}/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary (replace {Target_dataset_name}): https://github.com/MosyMosy/FN_Model_Zoo/tree/main/Dictionaries/AdaBN_na/{Target_dataset_name}/checkpoint_best.pkl
Baseline BN (ImageNet) (Table 1)
To Train: python ImageNet.py --dir ./logs/ImageNet/ --arch resnet18 --data ./data/ILSVRC/Data/CLS-LOC --gpu 0
To Fine-Tune: python ImageNet_finetune.py --save_dir ./logs/ImageNet --target_dataset {Target dataset name} --subset_split datasets/split_seed_1/{Target dataset name} _labeled_80.csv --embedding_load_path ./logs/baseline_teacher/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary
Baseline FN (ImageNet) (Table 1)
To Train: python ImageNet_na.py --dir ./logs/ImageNet_na/ --arch resnet18 --data ./data/ILSVRC/Data/CLS-LOC --gpu 0
To Fine-Tune: python ImageNet_finetune.py --save_dir ./logs/ImageNet_na --target_dataset {Target dataset name} --subset_split datasets/split_seed_1/{Target dataset name} _labeled_80.csv --embedding_load_path ./logs/baseline_teacher/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary
Baseline beta (ImageNet) (Table 1)
To Train: python ImageNet_nw.py --dir ./logs/ImageNet_nw/ --arch resnet18 --data ./data/ILSVRC/Data/CLS-LOC --gpu 0
To Fine-Tune: python ImageNet_finetune.py --save_dir ./logs/ImageNet_nw --target_dataset {Target dataset name} --subset_split datasets/split_seed_1/{Target dataset name} _labeled_80.csv --embedding_load_path ./logs/baseline_teacher/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary
Baseline gamma (ImageNet) (Table 1)
To Train: python ImageNet_nb.py --dir ./logs/ImageNet_nb/ --arch resnet18 --data ./data/ILSVRC/Data/CLS-LOC --gpu 0
To Fine-Tune: python ImageNet_finetune.py --save_dir ./logs/ImageNet_nb --target_dataset {Target dataset name} --subset_split datasets/split_seed_1/{Target dataset name} _labeled_80.csv --embedding_load_path ./logs/baseline_teacher/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary
Near-domain few-shot evaluation (Table 4)
Baseline BN
To Fine-Tune: python finetune.py --save_dir ./logs/eval/baseline_teacher --target_dataset ImageNet_test --subset_split datasets/split_seed_1/ImageNet_val_labeled.csv --embedding_load_path ./logs/baseline_teacher/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary
Baseline FN
To Fine-Tune: python finetune.py --save_dir ./logs/eval/baseline_na_teacher --target_dataset ImageNet_test --subset_split datasets/split_seed_1/ImageNet_val_labeled.csv --embedding_load_path ./logs/baseline_na_teacher/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary
AdaBN BN
To Adapt: python AdaBN.py --dir ./logs/AdaBN_teacher/miniImageNet --base_dictionary logs/baseline_teacher/checkpoint_best.pkl --target_dataset ImageNet_test --target_subset_split datasets/split_seed_1/ImageNet_val_labeled.csv --bsize 256 --epochs 10 --model resnet10
To Fine-Tune: python finetune.py --save_dir ./logs/AdaBN_teacher/miniImageNet --target_dataset ImageNet_test --subset_split datasets/split_seed_1/ImageNet_val_labeled.csv --embedding_load_path ./logs/AdaBN_teacher/miniImageNet/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary
AdaBN FN
To Adapt: python AdaBN.py --dir ./logs/AdaBN_na_teacher/miniImageNet --base_dictionary logs/baseline_na_teacher/checkpoint_best.pkl --target_dataset ImageNet_test --target_subset_split datasets/split_seed_1/ImageNet_val_labeled.csv --bsize 256 --epochs 10 --model resnet10
To Fine-Tune: python finetune.py --save_dir ./logs/AdaBN_na_teacher/miniImageNet --target_dataset ImageNet_test --subset_split datasets/split_seed_1/ImageNet_val_labeled.csv --embedding_load_path ./logs/AdaBN_na_teacher/miniImageNet/checkpoint_best.pkl --freeze_backbone
Pre-Trained Dictionary