Train and Test

Contents

1. Image Retrieval
   1. Split training set
   2. Train old model
   3. Train new (compatible) model
2. Face Recognition

Image Retrieval

Split training set

Model upgrades usually happen when (i) training data is increased or (ii) the model structure is strengthened.

To simulate the scenario (i), we first run the following script to split the whole training set by data (cmd: --split_gldv2_by_data) or class (cmd: --split_gldv2_by_class).

# split by data
python -m utils.split_dataset --root_path '../annotation/gldv2' --split_gldv2_by_data --split_ratio 0.3

# split by class
python -m utils.split_dataset --root_path '../annotation/gldv2' --split_gldv2_by_class --split_ratio 0.3

Train old model

Download pretrained resnet models from , and save them into ./pretrained_model/.

DistributedDataParallel (DDP) Training command:

python train.py --name "old_resnet50" --train_data_dir './data/gldv2/' --train_img_list './annotation/gldv2/gldv2_old_30percent_data.txt' --dist-url 'tcp://127.0.0.1:23456' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --arch resnet50  --pretrained_model_path "./pretrained_model/resnet50-0676ba61.pth" --epochs 30 -bs 192 --workers 8 --lr_scheduler cosine --lr 0.1  --val_period 5 -c config.json --use_amp

Command	Description	Options
--arch	-	resnet50 / resnet101 / inception_v3
--lr_scheduler	-	cosine / step
--train_data_dir	image path	-
--train_img_list	index file	-
--use_amp	turn on automatic mixed precision training	-

Train new (compatible) model

DistributedDataParallel (DDP) Training command:

python train_bct.py --name "new_resnet50" --train_data_dir './data/gldv2/' --train_img_list './annotation/gldv2/label_81313.txt' --dist-url 'tcp://127.0.0.1:23456' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --arch resnet50  --pretrained_model_path "./pretrained_model/resnet50-0676ba61.pth" --epochs 30 -bs 192 --workers 8 --lr_scheduler cosine --lr 0.1  --val_period 5 -c config.json --use_amp --old_arch resnet50 --old_pretrained_model_path './saved/old_resnet50/xxx/models/ckpt/xxx.pth.tar' --comp_loss_type contra_ract --comp_loss_topk 10

Command	Description	Options
--comp_loss_type	-	bct / lfw / fd / l2 / contra / triplet / bct_ract / contra_ract / triplet_ract
----comp_loss_topk	select top-k negative samples	-
--comp_loss_temp	temperature in contrastive loss	-
--comp_loss_triplet_margin	margin in triplet loss	-
--comp_loss_weight	compatible loss weight	-