MultiLabelClassification

This is a multi label classification codebase in PyTorch. Currently, it supports ResNet101, SSGRL (a implement of paper "Learning Semantic-Specific Graph Representation for Multi-Label Image Recognition" based on official repository HCPLab-SYSU/SSGRL) and training on Pascal Voc 2012, COCO and Visual Genome.

Requirements

• Python 3.6
• PyTorch 1.1
• TorchVision 0.3

Data preparation

Download datasets and symlink the paths to them as follows:

mkdir data
ln -s /path/to/mscoco data/coco
ln -s /path/to/VisualGenome1.4 data/VisualGenome1.4
ln -s /path/to/VOC2012 data/VOC2012

mkdir tmp
ln -s /path/to/glove.840B.300d.txt tmp/glove.840B.300d.txt

Running following scripts to preprocess datasets and generate desired data for SSGRL model.

python scripts/voc2012.py
python scripts/coco.py
python scripts/vg500.py

python scripts/preprocessing_ssgrl.py --data [voc2012, coco, vg500]

Training

python train.py --config $cfg_file_path

For example, with default optimizer(Adam) and loss(BCElogitloss), training resnet101 model on different dataset:

python train.py --config configs/coco_resnet101.yaml
python train.py --config configs/voc2012_resnet101.yaml

training ssgrl model on different dataset:

python train.py --config configs/coco_ssgrl.yaml
python train.py --config configs/voc2012_ssgrl.yaml

To resume training, you can run train.py with argument --resume.

Pretrained models

Pretrained models are provided on google drive.

Evaluation

python evaluate.py --config $cfg_file_path

For example:

python evaluate.py --config configs/vg500_resnet101.yaml
python evaluate.py --config configs/vg500_ssgrl.yaml

Results

Typically, The performances of pretrained multi label classification models are evaluated with mean average precision (mAP) and reported as follows:

models	VOC2012	COCO	VG500
ResNet101	0.901	0.802	0.293
SSGRL	0.923	0.837	0.334

Acknowledgements

Thanks the official implement SSGRL and awesome PyTorch team.