search

bagustris / Apsipa2019_SpeechText

Repository for code and paper submitted for APSIPA 2019, Lanzhou, China
22 stars 5 forks source link

How many performs are needed to achieve similar results? #1

Open Chenpuh opened 3 years ago

Chenpuh commented 3 years ago

Excuse me. I train the model 20 times in LSTM+Dense model, but the highest maximum accuracy is only 68.42%. In your paper, due to random initialization of GPU and CPU used for computation, it needs to be performed several times to obtain the similar result although a fixed random number is initialized at the top of the computer program for deep learning computation. How many performs are needed to achieve similar results?

bagustris commented 3 years ago

Hi there, Sorry for just read, GitHub didn't notify me about this issue. I didn't remember correctly. Perhaps about 10 times or more. Aside from the GPU, I think the Jupyter notebook also contributes the result I obtained.

Now, I stay away from the Jupyter notebook to get reproducible results. And yes, it is now a little bit smaller than what I reported in the paper when I run it again (the notebook still shows higher results).