zhongkaifu / advutils Goto Github PK

Hello, I am using this library as part of RNNSharp and CRFSharp, and other very good libraries, buit by you, and ran into several problems.

My cultural region is "es-AR" as well as any Spanish culture, the decimal point specification is opposite to the standard "en-US" and other English ones, so any time you write or read a file, containing any double or single floating point number, the data gets "silently" corrupted, adn the decimal points get ignored and taken as formatting strings. so comma is used instead of decimal point, and viceversa.

This corrupts many tests, and invalidates portability of libraries,and training sets, but for worse.. this happens under the hood, so if you are unaware of this, you may run into trouble as I personally did!

My recommendation is to mention this on former releases, and please format the numbers, as well as the parsers to a Culture independent numbering System as provided on the .NET library

I modified manually this on the source files, and it is algo "better " because many training sets, don't need very precise numbers, even many times there are internally treated as single (7 digits) and externally printed out as "double" unecessarly clogging the training/model files.

My suggestion is using the .NET "G5" floating number conversion format (5 digits) for training/dictionary data, s well ass G10 for seldom numeric values. I internally do use 3-4 digits and it proved not to alter any results, after being trained and dumped into a file. also I suggest to use the conversion:

double.Parse(string, System.Globalization.CultureInfo.InvariantCulture)

as default Parsing, to avoid such problems!

Also there is no performance degradation on neither this formatting.