ReadMe.md

Dynamic-Meta-Filter-Reproduction

环境

• Ubuntu 20.04
• RTX 3060 Laptop / A 4000 / A 5000 / M 40 / V 100
• LibFewShot & Dynamic-Meta-filter
• CUDA == 12.1 / 11.8
• Python 3.9 / 3.11
• GCC/G++ 9.4
• PyTorch

- torch
- torchvision
- torchdiffeq
- tqdm
- numpy
- einops
- future
- matplotlib
- numpy
- pandas
- Pillow
- PyYAML
- rich
- scikit-learn
- scipy
- tensorboard

启动

安装框架

LibFewShot

git clone https://github.com/RL-VIG/LibFewShot.git

测试安装

1. 下载 miniImageNet--ravi.tar.gz 并解压到 \your_path\LibFewShot\data\fewshot

2. 修改run_trainer.py中config设置的一行为

   config = Config("./config/test_install.yaml").get_config_dict()
   

3. 修改config/headers/data.yaml中的data_root为需要使用的数据集的路径

   /data/fewshot/miniImageNet--ravi   -->    ./data/fewshot/miniImageNet--ravi
   

4. 执行

   python run_trainer.py
   

5. 若第一个epoch输出正常，则表明LibFewShot已成功安装。

Dynamic-Meta-filter

git clone https://github.com/chmxu/Dynamic-Meta-filter.git

训练与测试

使用Dynamic-Meta-filter源代码

python setup.py develop build
python train.py --root {data root} --nExemplars {1/5} --resume {./weights/mini/[1/5]shot.pth.tar}

使用LibFewShot

cd core
python setup.py develop build
cd /path/to/LibFewShot
python run_trainer.py

复现结果表

	Frame	Embedding	miniImageNet (5,1)	miniImageNet (5,5)
1	Dynamic-Meta-filter	ResNet12	67.88 ± 0.49	82.11 ± 0.31
2	LibFewShot	ResNet12	59.453	——

以下为log_train/.txt及.log部分截图，具体训练过程可见压缩包

Dynamic-Meta-filter - miniImageNet (5,1)

Dynamic-Meta-filter - miniImageNet (5,5)

LibFewShot - miniImageNet (5,1)