/transfer-mxnet

transfer learning written in mxnet

Primary LanguagePython

transfer-mxnet

Unsupervised transfer learning for image classification written in mxnet.

This is a library for unsupervised transfer learning using mxnet. We mainly implemented three algorithms:

  • mmd described in paper "Learning Transferable Features with Deep Adaptation Networks".
  • jmmd described in paper "Deep Transfer Learning with Joint Adaptation Networks".
  • AdaBN described in paper "REVISITING BATCH NORMALIZATION FOR PRACTICAL DOMAIN ADAPTATION".

For original caffe implementation of mmd and jmmd, please refer to here

If you have any problem about this code, feel free to concact us with the following email:

Note that this repo is only for unsupervised image classfication transfer learning.

Experiments

We introduce our experiments on cars dataset:

  • Source dataset is a high quality cars image dataset fetched from web with accurate annotated labels(car models).
  • Target dataset is a surveillance image dataset which is public available(compcars-sv).

During training, we set all labels in Target dataset to null-label(9999 by default) then it becomes unsupervised TL problem.

Method Accuracy
CNN(no TL) 68.7%
AdaBN 71.6%
DAN(mmd) 73.7%
JAN(jmmd) 78.9%

If you want to train your own models with mmd(especically JAN as it is the best approach), please use following steps.

Data Preparation

  • Download mxnet resnet-152 imagenet-11k pretrained model to model/ directory, from here.
  • Prepare your source domain dataset to data/source.lst in mxnet lst format.
  • Prepare your target domain dataset to data/target.lst. Generally, label id in target.lst should be null-label(9999 by default). But semi-supervised TL is also allowed, you can choose hundreds of items to be their real valid label id.
  • Prepare your validation dataset in target domain to data/val.lst. It can be the same with data/target.lst but all with valid labels.

Training Model

Env variables setting.

export MXNET_CPU_WORKER_NTHREADS=15

Training stage 1, do softmax training on source dataset only.

python fine-tune.py --train-stage 0 --pretrained-model 'model/resnet-152' --pretrained-epoch 0 --model-prefix 'model/mmd' --num-classes 263 --lr 0.01 --lr-step-epochs '10,16' --num-epochs 18 --lr-factor 0.1 --gpus 0,1,2,3 --batch-size 64

Training stage 2, do softmax+mmd joint training to produce final model.

python fine-tune.py --train-stage 1 --pretrained-model 'model/mmd' --pretrained-epoch 18 --model-prefix 'model/mmd1' --num-classes 263 --lr 0.0001 --lr-step-epochs '6,8,10,12' --num-epochs 14 --lr-factor 0.5 --gpus 0,1,2,3 --batch-size 64

Parameter Tuning

TODO, check them in source code now.

Use AdaBN

python adabn.py --model <trained-model-prefix> --epoch <load-epoch> --val 'data/val.lst' --gpu 0

It will firstly calculate BN statistics using target domain dataset then write back to preloaded model. Second, use this modified model to validate the classification accuracy on target dataset. You can change the corresponding BN layers name in source code.