You think so?
well I haven’t tried any RNN architecture, but my first approach was also CNNs + FFN and tho the training accuracy was high, the testing accuracy was low
I felt it probably needed more training data to generalize better
I even tried just FFNs, same thing
I do.
Now is the best time to start!
Well, I’ll try it and give you feed back when I’m done
Thanks for everything.
1 Like
Also, the input of the model was converted from temporal data to frequency domain signal
So I don’t think It’ll still have that time dependency it initially had before feature extraction.
Would you still suggest I go for and RNN model?
I don’t know enough to suggest either way. Searching for “RNN frequency domain” turns up lots of results, perhaps you should look at some of them to see what the state of the art is?
Ok thanks
I’ll just have to try and see
Yo, so I’ve fixed the problem finally
The issue was that the NNModel.train() was outside the epoch loop, and in the epoch loop, the testing function is called at the 0th epoch and it toggles the NNModel.train() to NNModel.eval().
So since the NNModel.train() was outside the loop, after the 0th epoch the model doesn’t train.
Now that I’ve fixed that, the training accuracy seems to be getting really high, and the testing accuracy is increasing as well, tho not as high as training, so I’ve applied the dropout layers
I’ve also implemented the RNN architecture a on the time domain data (no feature extraction), and it seems to be training well as well
1 Like