Can you share the training parameters? I am using a dataset and cannot obtain good results

[yangh-zzf-itcast]

this is my parameters
Namespace(batch_size=32, database_dir='/ssd_data/yanghang/palm_vein/POLYU/NIR_polyu_crop/', embedding_size=512, evaluate=False, learning_rate=0.01, logdir='./data_polyup/CSVFiles/cnn_logs.csv', margin=0.5, num_classes=450, num_epochs=300, num_workers=8, outdir='modeldir/00/05/', scale_rate=32.0, start_epoch=0, test_dir='/ssd_data/yanghang/palm_vein/POLYU/data_polyup_CSVFiles/output_list_test_pair.csv', train_dir='/ssd_data/yanghang/palm_vein/POLYU/data_polyup_CSVFiles/output_list_train.csv', type='aamp', valid_dir='/ssd_data/yanghang/palm_vein/POLYU/data_polyup_CSVFiles/output_list_val.csv', weight_decay=0.0001)
I use polyu palm vein dataset，and use MNASNet_Modified model，The validation set loss keeps rising, obviously overfitting

[151210148]

hi , POLYU data can share? thanks

[jieblue]

Can u share POLYU palm vein datasets？ Thanks.

[deeperlearner]

@yangh-zzf-itcast
I have the same problem. Validation prec stays at zero, but the test pair results seem good.
I don't think the training process is responsible for the good test results. This is due to the pre-trained weights of ImageNet's MNASNet_Modified model.
Do you have any findings? Thanks.

[deeperlearner]

Reduce the batch size can improve the validation performance.
python3 -m benchmark_verification.main_train_CNN --num_epochs 100 --batch-size 8 --type none