When using TIPCB network to train ICFG-PEDES dataset, top-1=51.01%, which is three points lower than that reported in TIPCB. I was wondering if there was something wrong with the way I handled the raw dataset. So can you provide me your BERT_id_train_64_new.npz and BERT_id_test_64_new.npz files in ICFG-PEDES dataset and provide details of parameters? Thank you.
My email is [email protected]

Not all images in the test split are loaded two times. Your process is arbitrary.

In other words, there are 3074 unique images and 6156 unique sentences in the test split. It is obivous 3074 * 2 != 6156....

Your reported results may a little bit higher than the actual ones.

Please point out if I were wrong.

Thanks for your code!

I run the code using the default configurations on cuda=10.2 and pytorch=1.9.0, and only obtain the following results:
Rank@1:0.626545 Rank@5:0.825960 Rank@10:0.885491 mAP:0.534096.
Is it due to the cuda and pytorch version?

The detailed configurations are as follows:
CMPM: true
adam_alpha: 0.9
adam_beta: 0.999
adam_lr: 0.003
batch_size: 64
checkpoint_dir: ./log
dataset: CUHKPEDES
dir: ./datasets/
droprate: 0
embedding_type: BERT
epoches_decay: '50'
epsilon: 1.0e-08
feature_size: 2048
gpus: '1'
height: 384
log_dir: ./log
lr_decay_ratio: 0.1
lr_decay_type: MultiStepLR
max_length: 64
name: Origin
num_classes: 11003
num_epoches: 10000000
optimizer: adam
pretrained: true
resume: false
seed: 0
warm_epoch: 10
wd: 4.0e-05
width: 128

no train.csv, val.csv, test.csv.
how did you split the dataset exactly?

Thank you very much for your excellent work, it makes my thinking suddenly clear.
I encountered a problem when I repeat the experiment.
I used your parameters and experimental configuration that it in the paper.
python train.py --max-length 64 --batch-size 64 --num-epoches 80 --adam-lr 0.003 --gpus 0
But LOSS got NAN. I did not make other changes.
Since I didn't have a divided CSV file, I directly used the bert file you provided.
Please refer to the attached log.

train_log.txt