hardcoded number of samples per file

[chrismessenger]

I think that the train_file_idx variable is generated assuming that there are 1000 time series stored per training data file.

This may be correct (now) but should be fixed to be general.