ChunkerTrain fails with auto selection of training rounds using a separate development set

[qiangning]

When discussing pull req 682 with Christos, I tested ChunkerTrain for something else, but I found that it's now failing.
Here's the result I got standing at the commit: 99971bd This is a commit before any of my recent changes to chunker.

Iteration number : 1. F1 score on devset: 0.9995071926407434
Iteration number : 2. F1 score on devset: 0.9995071926407434
Iteration number : 3. F1 score on devset: 0.9995071926407434
Iteration number : 4. F1 score on devset: 0.9995071926407434
Iteration number : 5. F1 score on devset: 0.9995071926407434

(See the F1 on dev is not changing at all, which is abnormal.)

The benchmark test is still good:

With Gold POS
 Label   Precision Recall   F1   LCount PCount
----------------------------------------------
ADJP        76.633 69.635 72.967    438    398
ADVP        81.862 79.215 80.516    866    838
CONJP       45.455 55.556 50.000      9     11
INTJ        50.000 50.000 50.000      2      2
LST          0.000  0.000  0.000      5      1
NP          94.106 93.962 94.034  12422  12403
PP          96.770 97.776 97.270   4811   4861
PRT         72.072 75.472 73.733    106    111
SBAR        88.280 87.290 87.782    535    529
UCP          0.000  0.000  0.000      0      5
VP          93.416 93.517 93.466   4658   4663
----------------------------------------------
O            0.000  0.000  0.000   1244   1274
----------------------------------------------
Overall     93.510 93.393 93.451  23852  23822
Accuracy    88.763   -      -      -     25096

With NO POS
 Label   Precision Recall   F1   LCount PCount
----------------------------------------------
ADJP        78.608 69.635 73.850    438    388
ADVP        80.427 78.291 79.345    866    843
CONJP       45.455 55.556 50.000      9     11
INTJ       100.000 50.000 66.667      2      1
LST          0.000  0.000  0.000      5      0
NP          94.193 94.019 94.106  12422  12399
PP          96.656 97.942 97.295   4811   4875
PRT         60.417 82.075 69.600    106    144
SBAR        86.813 88.598 87.697    535    546
UCP          0.000  0.000  0.000      0      4
VP          94.105 94.246 94.176   4658   4665
----------------------------------------------
O            0.000  0.000  0.000   1231   1207
----------------------------------------------
Overall     93.529 93.623 93.576  23852  23876
Accuracy    89.028   -      -      -     25083

(The numbers are exactly the same as chunker/doc/performance.txt)

I guess maybe the changes in other parts of the repository failed ChunkerTrain since ChunkerTrain itself hasn't been updated at all since 11/18/2016 (see this log).