shank wrote: Tue 17. Sep 2019, 12:40
can you show me some examples for normalizing data in MemBrain ?
I did that initially and all I could get was a notification "Data in the lesson editor is out of range with respect to normalization setting".
Even if I bypassed it by running the network anyway, it never gave correct results. That's why I normalized input data manually and uploaded it in MemBrain.
Normalization actually is quite simple and straight forward: For each input or output neuron adjust the user defined data as you like: Edit neuron properties, select button <Normalization...>, check the check box <Use Normalization> and enter the <Upper Limit> and <Lower Limit> values.
If you try to train a net with data in the lesson editor that is outside of this adjusted range then MemBrain will issue the warning you mentioned and even select the neurons which caused the normalization range violation for you.
You can also adjust the normalization settings via the <Normalization Wizard> instead:
- Load the lesson with all your data patterns. If you use separate training and validation lessons then first combine both lessons (Lesson Editor menu: <Lesson Files><Append Lesson to Current Lesson...> into one overall lesson. Attention: Don't accidentally overwrite your original lesson with this afterwards, so best is to immediately save the new combined lesson using a new name.). Alternatively, you can also create an explicit normalization lesson with just two patterns: One pattern contains all the maximum values and the other all the minimum values for each neuron.
- Select the inputs and outputs for which you want to adjust normalization. If all inputs and outputs shall be adjusted then simply select the full net (press <Ctrl + 'A')
- Execute <Extras><Normalization Wizard...> to start the wizard. It will automatically suggest a suitable normalization setting for each of the input and output neurons based on the currently active lesson (which is your overall data or normalization lesson as described above).
- Click <Next> and finally <Apply> as appropriate. You can also enter your own values instead of the suggested ones during this process.
- Don't forget to save your net afterwards.
That's it.
shank wrote: Tue 17. Sep 2019, 12:40
Also, as suggested by you, I changed the activation function to INDENT 0..1 but it didn't work. I got the same result.
You are right, this hint I gave is not valid: You need to activate and adjust Normalization settings for the neurons in order to cause the script to use other data ranges.
Does it work now as you want to? Please come back with more specific questions if you still have issues.