# TransferRobust

**Repository Path**: mymashiro/TransferRobust

## Basic Information

- **Project Name**: TransferRobust
- **Description**: transfer learning with robust
- **Primary Language**: Python
- **License**: MIT
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2020-09-27
- **Last Updated**: 2021-06-09

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# CARTL: Cooperative Adversarially-Robust Transfer Learning 

Code for ICML'21 paper, *CARTL: Cooperative Adversarially-Robust Transfer Learning*.

### Prerequisites

- python==3.7
- pytorch==1.5.1
- torchvision==0.6.1
- tensorboard==2.4.0
- tensorboard-plugin-wit==1.7.0
- toml==0.10.2
- matplotlib==3.2.2
- pydantic==1.5.1
- pydantic[dotenv]
- pandas==1.2.0
- autoattack==0.1
- foolbox==3.2.1

You can download these packages manually or run `pip install -r requirements.txt`

For command-line interface support, run `pip install --editable .`

### License

This code is released under the MIT License (refer to the LICENSE file for details).

### Code Overview

```
.
├── exps 			# Code for experiments 
├── scripts 		# Script for simplifying experiments
├── src 			
│   ├── cli 		# Code for command-line interface
│   ├── networks	# Code for neural networks
│   ├── trainer		# Code for training process
│   └── utils		# Code for utilities(logging, dataloader etc.)
└── tests			# Code for testing key components
```

Please run the script `./scripts/init.sh`  for creating required directories `checkpoint`, `trained_models`, `misc_results`(containing experimental results), `logs` (containing running logs) and `runs`(containing files for tensorboard).

### Simple Instructions

#### Train an adversarial teacher model

```bash
cli at -m <MODEL_ARCH> -n <NUM_CLASSES> -d <DATASET>
```

Run `cli at --help` for details

#### Train a fdm teacher model

```bash
python -m exps.fdm_l2 -m <MODEL_ARCH> -n <NUM_CLASSES> -d <DATASET> -k <K> -l <LAMBDA>
```

Or using cli tools(run `cli fdm --help` for details)

#### Transfer learning

```bash
python -m exps.transfer_learning -m <MODEL_ARCH> -n <NUM_CLASSES> -d <DATASET> -k <K> -t <TEACHER>
```

Or using cli tools(run `cli tl --help` for details)

#### Transfer learning with NEFT

Change parameters in `./script/pwrn34_neft_sn_norm.sh` and run

```
./script/pwrn34_neft_sn_norm.sh
```

Or using cli tools(run `cli sntl --help` for details)

#### Lwf

```bash
cli lwf -m <MODEL_ARCH> -n <NUM_CLASSES> -d <DATASET> -l <LAMBDA> -t <TEACHER>
```

#### Explore experiment

Here is a quick example on how to explore our experiment

```bash
./script/exploration_exp.sh [GPU_ID]
```

If you wish to explore more results, just modify the parameters in `./script/exploration_exp.sh`

### Reference

```
@inproceedings{chen2021cartl,
 author = {Chen, Dian and Hu, Hongxin and Wang, Qian and Li, Yinli and Wang, Cong and Shen, Chao and Li, Qi},
 title = {CARTL: Cooperative Adversarially-Robust Transfer Learning},
 BOOKTITLE = {International Conference on Machine Learning},
 YEAR = {2021}
}
```