Average loss of 1 (and incorrect labels) reported in --binary mode

[mjwillson]

Perhaps the easiest way to explain this is to show output with and without the --binary option with everything else held fixed. My input lines look like

-1 |Text hello this is an example
1 |Text and another one

Without --binary everything behaves itself:

$ vw -kcd examples.txt --loss_function logistic
Num weight bits = 18
learning rate = 0.5
initial_t = 0
power_t = 0.5
creating cache_file = examples.txt.cache
Reading datafile = examples.txt
num sources = 1
average    since         example     example  current  current  current
loss       last          counter      weight    label  predict features
0.641010   0.641010            3         3.0  -1.0000  -0.1864        4
0.577645   0.514281            6         6.0  -1.0000  -0.5663        7
0.476604   0.355354           11        11.0  -1.0000  -2.1518       35
0.405201   0.333797           22        22.0  -1.0000  -2.0671       21
0.307740   0.210278           44        44.0  -1.0000  -1.3955       16
0.213920   0.117918           87        87.0  -1.0000  -2.8665       11
0.146646   0.079372          174       174.0  -1.0000  -1.7852        3
0.098552   0.050458          348       348.0  -1.0000  -3.5060        6
0.078723   0.058894          696       696.0  -1.0000  -4.2158       13
0.055447   0.032171         1392      1392.0  -1.0000  -4.3897        5
0.053205   0.050964         2784      2784.0  -1.0000  -8.2964       17
0.046466   0.039726         5568      5568.0  -1.0000  -3.8190        4
0.036872   0.027276        11135     11135.0  -1.0000  -3.5071        2
0.035412   0.033953        22269     22269.0  -1.0000  -8.4659       15

finished run
number of examples per pass = 39442
passes used = 1
weighted example sum = 39442
weighted label sum = -38960
average loss = 0.0311541
best constant = -0.987779
total feature number = 504260

But with --binary I see an average loss of exactly 1 reported in each log line and in the final output. Also a weighted label sum of exactly 0 (which is incorrect), and "current label" of 3212836864 (= 0xbf800000) for all examples, also incorrect, the labels are all 1 or -1:

$ vw -kcd examples.txt --loss_function logistic --binary
Num weight bits = 18
learning rate = 0.5
initial_t = 0
power_t = 0.5
creating cache_file = examples.txt.cache
Reading datafile = examples.txt
num sources = 1
average    since         example     example  current  current  current
loss       last          counter      weight    label  predict features
1.000000   1.000000          3      3.0   3212836864       -1        4
1.000000   1.000000          6      6.0   3212836864       -1        7
1.000000   1.000000         11     11.0   3212836864       -1       35
1.000000   1.000000         22     22.0   3212836864       -1       21
1.000000   1.000000         44     44.0   3212836864       -1       16
1.000000   1.000000         87     87.0   3212836864       -1       11
1.000000   1.000000        174    174.0   3212836864       -1        3
1.000000   1.000000        348    348.0   3212836864       -1        6
1.000000   1.000000        696    696.0   3212836864       -1       13
1.000000   1.000000       1392   1392.0   3212836864       -1        5
1.000000   1.000000       2784   2784.0   3212836864       -1       17
1.000000   1.000000       5568   5568.0   3212836864       -1        4
1.000000   1.000000      11135  11135.0   3212836864       -1        2
1.000000   1.000000      22269  22269.0   3212836864       -1       15

finished run
number of examples per pass = 39442
passes used = 1
weighted example sum = 39442
weighted label sum = 0
average loss = 1
best constant = 0
total feature number = 504260

I'm using the latest master from github. I get similar behaviour using other loss functions (hinge, squared), with multiple passes and in bfgs mode. I also tried labels of 0 and 1 which didn't help.

[mjwillson]

In case this helps, 3212836864 appears to be what you get if you treat a float -1.0 (the correct label in this case) as an unsigned int.

[JohnLangford]

Fixed.