got some confused result
shipleyxie opened this issue · 6 comments
i retrain this project background number is 522(home+factort+street noise), 24 activate audio and 24negatigve audio .
the train log is as foloow:
522/522 [==============================] - 72s 138ms/step - loss: 0.0530 - acc: 0.9998
Epoch 19/20
522/522 [==============================] - 72s 138ms/step - loss: 0.0524 - acc: 0.9998
Epoch 20/20
522/522 [==============================] - 72s 138ms/step - loss: 0.0519 - acc: 0.9998
537/537 [==============================] - 17s 32ms/step
could pelase give me some guide or suggestion to train the model to a success trigger?
or could you give me some guide on making training data set?
@Tony607
You still have this problem ?
hi, I changed the hot word and I am having this mistake but I do not find why, could you help me?
and the wrong words go down like hot words....
thanks
Obviously, your trained model predict wrong result. You need to choose data properly and retrain to get a good model. I suggest you change epoch from 40 to 400.
Let me know if this helps.
hi shipleyxie:
yes, i did it with 400 epoch but do not learn more than 0.85 train and 0.92 test like image, I change the learning rate to 0.005 and the accuracy was better 0.93 in train and 0.92 in test but the result was similar...
hi, I changed the hot word and I am having this mistake but I do not find why, could you help me?
and the wrong words go down like hot words....
thanks
I had similar problem.
I think its because the learning process is really slow, you going to take many epoch before it gets work.
if you use higher learning rate (e.g 0.0001 > 0.001) and lower drop out rate (e.g 0.8 > 0.5)
just to check if the model can fits the data.
you can get more reasonable result in first few hundreds epoch training.
Yesss I already resolve that problem, with exactly those tips, I had to wait 3 days training my model, thank you very much.
PS: I had to change the optimizer to RMSprop i think I am going to train with another...