This repository provides the MS-COCO training code for the Orderless Recurrent Models for Multi-label Classification paper which will be published in CVPR2020.
- python 2.7
- pytorch 0.4.1
Three steps of training:
python train_bce.py -image_path {image_path} -save_path {save_path1}
python train_lstm.py -image_path {image_path} -save_path {save_path2} -order_free pla -finetune_encoder -swa_params "{'lr_high': 1e-3, 'lr_low': 1e-6, 'cycle_length': 3, 'swa_coeff': 0.1}" -encoder_weights {save_path1}/BEST_checkpoint.pt
python train_lstm.py -image_path {image_path} -save_path {save_path3} -order_free pla -decoder_lr 1e-5 -snapshot {save_path2}/BEST_checkpoint.pth.tar -epochs 5 -train_from_scratch
python train_lstm.py -image_path {image_path} -snapshot {save_path3}/BEST_checkpoint.pth.tar -test_model
The encoder-decoder architecture that is implemented in this repository is based on the implementation in here.