slayerPytorch icon indicating copy to clipboard operation
slayerPytorch copied to clipboard

Accuracy for DVS128 Gesture

Open hongchi533 opened this issue 2 years ago • 7 comments

Hello, author:

I can only obtain accuracy around 84% for DVS128 Gesture when using /bamsumit/slayerPytorch/tree/master/exampleLoihi/03_IBMGesture/ibmGestureTrain.ipynb, with 100 epochs and changing nothing else.

May I know how to improve accuracy to the level proposed in your paper (94%) ?

hongchi533 avatar Mar 24 '22 04:03 hongchi533

There are a couple of things you can try:

  • First is you may want to add data augmentation for better results.
  • You can try to optimize the learning rate and the number of epochs.

It might have just been your initialization seed as well.

bamsumit avatar Mar 24 '22 22:03 bamsumit

Hello @hongchi533. I have the same problem as you describe. I can only obtain an accuracy of 84%. However, I am not running the program on Loihi hardware. Have you managed to improve the accuracy? If so, please let me know how you did it.

FredrikBixo avatar Apr 26 '22 12:04 FredrikBixo

@bamsumit Hello, the data captured by DVS is different from the image data. Can you tell me how you do data augmentation? Thank you!

chenjiefighting avatar Sep 20 '22 08:09 chenjiefighting

@chenjiefighting look st this repo: https://github.com/neuromorphs/tonic

bamsumit avatar Sep 26 '22 17:09 bamsumit

Hello, author:

I can only obtain accuracy around 84% for DVS128 Gesture when using /bamsumit/slayerPytorch/tree/master/exampleLoihi/03_IBMGesture/ibmGestureTrain.ipynb, with 100 epochs and changing nothing else.

May I know how to improve accuracy to the level proposed in your paper (94%) ?

Hi, I downloaded https://research.ibm.com/interactive/dvsgesture/, it has trials_to_train.txt and trials_to_test.txt. These files have data in the format user01_fluorescent.aedat, but the script ibmGestureTrain.py is accepting data in int format. Can anyone help me in converting the data from this format to the int. Regards, surabhi

surabhi-Siemens avatar Dec 06 '22 05:12 surabhi-Siemens

You can look this, #54 ,It might help.

chenjiefighting avatar Dec 06 '22 07:12 chenjiefighting

@chenjiefighting Have you been able to achieve the nominal accuracy? I am still stuck to 84%.

Po0ria avatar May 11 '23 20:05 Po0ria