search

zhigao2017 / CurvatureGeneration_FSL

9 stars 0 forks source link

the accuracy has not changed or improved. Why? #2

Open starstars123 opened 2 years ago

starstars123 commented 2 years ago

Hello, the data set is miniimagenet. When the program runs to 155 epochs, the accuracy on the test set is still 0.2586. The loss is backward, but the accuracy has not changed or improved. Why? %GZ)$Y(HWM~2SEUYRIQM_PE

zhigao2017 commented 2 years ago

Hello, the data set is miniimagenet. When the program runs to 155 epochs, the accuracy on the test set is still 0.2586. The loss is backward, but the accuracy has not changed or improved. Why? %GZ)$Y(HWM~2SEUYRIQM_PE

Hi. Which code and which setting do you run?

zhigao2017 commented 2 years ago

We updated the code last month. Do you use the latest code?

starstars123 commented 2 years ago

We updated the code last month. Do you use the latest code?

yes, I use the latest code , The running files and configurations are as follows： YEN$L{}} )7K7%O5%1FE)3D

zhigao2017 commented 2 years ago

We updated the code last month. Do you use the latest code?

yes, I use the latest code , The running files and configurations are as follows： YEN$L{}} )7K7%O5%1FE)3D

Hi, you can try the following hyperparameters. image

starstars123 commented 2 years ago

We updated the code last month. Do you use the latest code?

yes, I use the latest code , The running files and configurations are as follows： YEN$L{}} )7K7%O5%1FE)3D

Hi, you can try the following hyperparameters. image

Thank you for your reply. After I try your parameters and run the program for 200 epochs, the accuracy has changed. Excuse me, how do I want to run this model on my own private dataset and find the appropriate parameters? QWZMY I6%KEHP DMINZ}4$C

zhigao2017 commented 2 years ago

Hi, for ConvNet, it needs more epochs to train the model. Besides, maybe tuning parameters about the curvature generation processes is useful. Such as 'l' and 'divide'.