Closed KoiLiu closed 1 year ago
Thank for your reply. I will try it again.
I git the latest version of SJ and retrained the train.py in the gesture. Maybe I didn't express my problem clearly before. I trained the model on the DVS128Gesture dataset, not the Imagenet. And this time the acc of test is only 81.94%, which is still much lower than 97.92 acc in your work. I don't know the reson for it. More details are as followed:
Namespace(T=16, T_train=12, adam=True, amp=True, batch_size=16, connect_f='ADD', data_path='./dataset', device='cuda:0', dist_url='env://', distributed=False, epochs=192, lr=0.001, lr_gamma=0.1, lr_step_size=64, model='SEWResNet', momentum=0.9, output_dir='./logs', print_freq=64, resume='', start_epoch=0, sync_bn=False, tb=True, test_only=False, weight_decay=0, workers=4, world_size=1)... Training time 0:50:47 max_test_acc1 81.94444444444444 test_acc5_at_max_test_acc1 100.0
I trained the DVS128Gesture dataset in the afternoon with the simple PythonNet and the acc was around 94%, but I don't know the reason the lower acc of SEWResNet. So I really want your help and advice. Thank you a lot!
Hello, I trained the smodel in dvsgesture. But the testing_acc1 is around 80%, which is much lower than the accuracy in your paper. I used the Adam with lr=0.001 and the others are the default.
![43026CF9-CC4D-4527-84E0-CFFF8062C0EC](https://user-images.githubusercontent.com/47879103/165747509-83992425-410e-41fc-8cf6-f56e6096026a.jpeg)
python train.py --tb --amp --output-dir ./logs --model SEWResNet --connect_f ADD --device cuda:0 --lr-step-size 64 --epoch 192 --T_train 12 --T 16 --data-path ./dataset --lr 0.001 --adam
The training loss and other results are as followed: