PaddlePaddle/PaddleTransfer

PaddlePaddle/PaddleTransfer

Description: 飞桨迁移学习算法库

Language: Python

License: Apache-2.0

Stars: 19

Forks: 3

Open issues: 0

Created: 2022-05-09T13:21:51Z

Pushed: 2023-02-20T07:15:20Z

Default branch: main

Fork: no

Archived: yes

README:

PaddleTransfer

Introduction

PaddleTransfer, a transfer learning tool of PaddlePaddle, provides a variety of algorithms to transfer the knowledge learned from source tasks to target task. It enables users to use the state-of-the-art transfer learning algorithms on main-stream model archtechtures.

PaddleTransfer provides various transfer learning algorithms, including [MMD(JMLR'12)](https://www.jmlr.org/papers/volume13/gretton12a/gretton12a.pdf?ref=https://githubhelp.com),[L2-SP(ICML'18)](http://proceedings.mlr.press/v80/li18a/li18a.pdf), [DELTA(ICLR'19)](https://openreview.net/pdf?id=rkgbwsAcYm), [RIFLE(ICML'20)](http://proceedings.mlr.press/v119/li20r/li20r.pdf), [Co-Tuning(NIPS'20)](https://proceedings.neurips.cc/paper/2020/file/c8067ad1937f728f51288b3eb986afaa-Paper.pdf), [MARS-PGM(ICLR'21)](https://openreview.net/pdf?id=IFqrg1p5Bc) and supports many main-stream model archtechtures, including ResNet, MobileNet and ViT. With several lines of codes, users can apply these algorithms on our predifined model or their own model easily.

Contents

• [Key Highlights](#Key-Highlights)
• [Installation](#Installation)
• [Usage Guideline](#Usage-Guideline)
• [Provided Algorithms](#Provided-Algorithms)

Key Highlights

High Accuracy

PaddleTransfer provides various transfer learning algorithms, users can conveniently apply them on their models, and select the most appropriate one with high accuracy for further usage.

Easy to Use

PaddleTransfer provides unified API to invoke different algorithms, users can easily apply them on their models with several lines of codes.

Fully Support

PaddleTransfer supports most commonly used models including Resnet, MobileNet and ViT, and will iterates rapidly to support more model archtectures.

Installation

Install by pip

python -m pip install paddletransfer

Dependencies

if you want to use our package in your own code, the following dependencies are required

• python >= 3.7
• numpy >= 1.21
• paddlepaddle >= 2.2 (with suitable CUDA and cuDNN version)

If you want to run our demo script, please make sure the following packages are installed correctly on your machine.

• visualdl
• yacs

Usage Guideline

Quick Start

Users can run our demo code for a quick start

python finetune.py --name [experiment_name] --train_dir [path_to_train_dir] --eval_dir [path_to_eval_dir] --model_arch [model_archtechture] --algo [finetune_algorithm] --gpu [device_for_experiment]

For model_arch argument, please choose one from "resnet18", "resnet34", "resnet50", "resnet101", "resnet152", "mobilenet_v2" and "vit". And for algo argument, please choose one from "base", "mmd", "rifle", "l2sp", "delta" and "cot". Mistyping may lead to unexpected behavior.

Please organize your dataset in the following format.

|_ root/
| |_ class1
| | |_ image_10026.JPEG
| | |_ ...
| |_ ...
| |
| |_ class100
| |_ ...
| |_ image_9993.JPEG

If you want to finetune the ViT model, please make sure you have set the configuration file and pretrained parameters file correctly, and remember to add the corresponding argumengts(--cfg and --model_path) in your command. You can get the configuration file and pretrained model from PaddleVit.

Use PaddleTransfer in Your Own Code

Import dependencies

from paddletransfer.finetuning.img_cls import *

We strongly recommand you to use the model architechtures we provide in backbones by following codes. The parameters setting are the same as the implementations in paddle.vision.models, remember to set pretrained=True for finetuning.

from backbones import resnet18, resnet34, resnet50, resnet101, resnet152, mobilenet_v2, build_vit

You can only import the model archtechtures that you need.

If you want to use ViT model, please put the config.py file either from our project or PaddleVit under your working directory and set other files correctly as described in [Quick Start](#Quick-Start)

If you want to use your self-defined model, please make sure that the name of layers are consistent with the implementations in paddle.vision.models and your model has two outputs: The first one is the original output of the network and the second one is the features(the intermediate output before average pooling layer and fc layer).

Initialize the algorithm object

algo = FinetuneDELTA(model, model_arch)

• To customize the hyperparameter for finetune algorithm, please add the following arguments to the initializer, for details about hyperparameter setting, please refer to [provided algorithms](#Provided-Algorithms)

confs = confs # a directory of the hyperparameter setting

Get the parameters list which need to update and pass them to optimizer

params = algo.params()
...
optimizer = paddle.optimizer.Optimizer(...,parameters = params,...)

In most cases, it is equal to model.parameters() and you may not want to invoke this method. But remember to do this if you are using Co-tuning algorithm since it has extra parameters to update

Get the regularization loss and merge it to the classification loss

loss_cls = paddle.nn.CrossEntropyLoss(y_data,logits)
loss_all = {'loss_cls': loss_cls}
loss_reg = algo.loss(x_data, y_data, logits, features, epoch, batch_id)
loss_all.update(loss_reg)
loss = sum(loss_all.values())
...
loss.backward()

Provided Algorithms

So far we have provided 5 algorithms for finetune, which are MMD, L2-SP, DELTA, RIFLE and Co-Tuning. If you do not want to use any finetune algorithms, just use the following code for vanilla finetune, the corresponding code for invoking different algorithms are in the respective sections.

algo = FinetuneBASE(model, model_arch)

MMD

Use the following code for invoking MMD algorithm

algo = FinetuneMMD(model, model_arch, confs=_confs)

The default hyperparameters for MMD…