This repository contains the code for the following ICML 2020 paper:
XtarNet: Learning to Extract Task-Adaptive Representation for Incremental Few-Shot Learning
This code was tested on the following environment:
- Ubuntu 16.04
- python 3.6
- cv2
- numpy
- pandas
- tensorflow==1.11
- tensorflow-gpu==1.9.0 (If cuda is available)
- tqdm
Download and decompress the file: [miniImageNet] (courtesy of Mengye Ren)
export DATA_ROOT={path/to/dataset}
mkdir -p $DATA_ROOT
cd $DATA_ROOT
mv ~/Downloads/mini-imagenet.tar . # place "mini-imagenet.tar" in $DATA_ROOT
tar -xvf mini-imagenet.tar
rm -f mini-imagenet.tar
Download and decompress the file: [tieredImageNet] (courtesy of Mengye Ren)
export DATA_ROOT={path/to/dataset}
mkdir -p $DATA_ROOT
cd $DATA_ROOT
mv ~/Downloads/tiered-imagenet.tar . # place "tiered-imagenet.tar" in $DATA_ROOT
tar -xvf tiered-imagenet.tar
rm -f tiered-imagenet.tar
https://github.com/EdwinKim3069/XtarNet.git
cd XtarNet
In order to pretrain backbone, run the python file run_pretrain.py.
For miniImageNet experiment,
import os
os.system('./run.sh 0 python run_exp.py '
'--config configs/pretrain/mini-imagenet-resnet-snail.prototxt '
'--dataset mini-imagenet '
'--data_folder DATA_ROOT/mini-imagenet/ '
'--results PATH/TO/SAVE_PRETRAIN_RESULTS '
'--tag PATH/TO/EXP_TAG '
)For tieredImageNet experiment,
import os
os.system('./run.sh 0 python run_exp.py '
'--config configs/pretrain/tiered-imagenet-resnet-18.prototxt '
'--dataset tiered-imagenet '
'--data_folder DATA_ROOT/tiered-imagenet/ '
'--results PATH/TO/SAVE_PRETRAIN_RESULTS '
'--tag PATH/TO/EXP_TAG '
)In order to run meta-training experiments, run the python file run_inc.py.
For miniImageNet experiments,
import os
nshot = 'NUMBER_OF_SHOTS'
tag = 'XtarNet_miniImageNet_{}shot'.format(nshot)
os.system('./run.sh 0 '
'python run_exp.py '
'--config configs/XtarNet/XtarNet-mini-imagenet-resnet-snail.prototxt '
'--dataset mini-imagenet '
'--data_folder DATA_ROOT/mini-imagenet/ '
'--pretrain PATH/TO/PRETRAIN_RESULTS/TAG '
'--nshot {} '
'--nclasses_b 5 '
'--results PATH/TO/SAVE_METATRAIN_RESULTS '
# '--eval '
# '--retest '
'--tag {} '.format(nshot, tag)
)For tieredImageNet experiments,
import os
nshot = 'NUMBER_OF_SHOTS'
tag = 'XtarNet_tieredImageNet_{}shot'.format(nshot)
os.system('./run.sh 0 '
'python run_exp.py '
'--config configs/XtarNet/XtarNet-tiered-imagenet-resnet-18.prototxt '
'--dataset tiered-imagenet '
'--data_folder DATA_ROOT/tiered-imagenet/ '
'--pretrain PATH/TO/PRETRAIN_RESULTS/TAG '
'--nshot {} '
'--nclasses_b 5 '
'--results PATH/TO/SAVE_METATRAIN_RESULTS '
# '--eval '
# '--retest '
'--tag {} '.format(nshot, tag)
)If you want to evaluate the meta-trained model, add --eval and --retest flag for restoring a fully trained model and re-run eval.
Our code is based on the implementations of Incremental Few-Shot Learning with Attention Attractor Networks.