createHiddenLayerNetsForCount expect too small a number #1
Comments
|
Hi, Would you please provide some training samples for me to test that issue ? Thanks,
Kalvar
… Emilian Bold ***@***.***> 於 2018年1月6日 下午11:08 寫道：
It seems to me RNN createHiddenLayerNetsForCount expects at most patterns[0].features.count.
If I use a bigger number the app crashes.
If I have a time series, the number of features is small so it seems to me this might be a reason why the network doesn't really converge and I get no predictions.
Of course, I might also be miss-using the library. I appreciate you making it public, I didn't have to try implement all of that!
*** Terminating app due to uncaught exception 'NSRangeException', reason: '*** -[__NSArrayM objectAtIndex:]: index 3 beyond bounds [0 .. 2]'
*** First throw call stack:
(
0 CoreFoundation 0x0000000104a70b0b __exceptionPreprocess + 171
1 libobjc.A.dylib 0x00000001044d5141 objc_exception_throw + 48
2 CoreFoundation 0x00000001049a5ffb -[__NSArrayM objectAtIndex:] + 203
3 RNN 0x0000000103593b40 -[RNNMath sumMatrix:anotherMatrix:] + 448
4 RNN 0x00000001035954cd -[RNNNet outputWithInputs:recurrentOutputs:] + 205
5 RNN 0x000000010358ca28 __37-[RNN(Training) forwardWithFeatures:]_block_invoke.47 + 120
6 CoreFoundation 0x0000000104a01652 __53-[__NSArrayM enumerateObjectsWithOptions:usingBlock:]_block_invoke + 114
7 CoreFoundation 0x0000000104a014e2 -[__NSArrayM enumerateObjectsWithOptions:usingBlock:] + 194
8 RNN 0x000000010358c511 -[RNN(Training) forwardWithFeatures:] + 721
9 RNN 0x000000010358cdeb -[RNN(Training) trainingWithPattern:] + 187
10 RNN 0x000000010358ea12 __30-[RNN(Training) startTraining]_block_invoke + 114
11 CoreFoundation 0x0000000104a01652 __53-[__NSArrayM enumerateObjectsWithOptions:usingBlock:]_block_invoke + 114
12 CoreFoundation 0x0000000104a014e2 -[__NSArrayM enumerateObjectsWithOptions:usingBlock:] + 194
13 RNN 0x000000010358e5ae -[RNN(Training) startTraining] + 478
14 RNN 0x0000000103590521 -[RNN trainingWithCompletion:] + 113
15 RNN 0x0000000103590674 -[RNN trainingWithIteration:completion:] + 132
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub, or mute the thread.
|
|
Looks to me like the code expects that the number of cells in the hidden layer must be smaller than the number of cells in the input layer. While I assumed I could increase that number to, perhaps, get better results. In your iOS-RNN sample file, if I replace this line with I need to tweak some parameters and I assumed that besides For simple data (straight line) the network does train, but not for the (semi-random?) test data I'm using. It might be a case of garbage-in garbage-out. The series 'looks' simple and the training and predictions are way off:
|
|
Please check the condition of forloop: for(int i = chunkSize; i < vals.count; i++){}, it won't be working on first time that i=7 and valus.count is 7 too. Hence, the patterns will be empty, then crashed. |
|
Please note I'm able to crash your own example from https://github.com/Kalvar/ios-RNN/tree/master/ios-RNN by just changing So, I sense there's something deeper here. Indeed, there's a small bug (the shame...) when chunkSize == number of train items, but the test data has alway been higher, so I didn't run into that. |
|
Okay, got your question, I'm checking now. |
|
It is a bug yes, I will fix it. |
|
Fixed createHiddenLayerNetsForCount crash bug. (Updated RNN version to 1.1) |
|
This fixes the I'm going to close this issue. I still don't see good convergence for my time series data. Would appreciate any hints you have for time series prediction using your library. I'm thinking about implementing a LSTM myself... |
It seems to me
RNN createHiddenLayerNetsForCountexpects at mostpatterns[0].features.count.If I use a bigger number the app crashes.
If I have a time series, the number of features is small so it seems to me this might be a reason why the network doesn't really converge and I get no predictions.
Of course, I might also be miss-using the library. I appreciate you making it public, I didn't have to try implement all of that!
The text was updated successfully, but these errors were encountered: