Created by: CrossLee1

I use the provided rank_cost loss function for my training, with trainer_count = 4. In the end of each pass, the log will print the info of pos/neg.

At the end of Pass-00000, the log info is as follows: 13:14:26.505049 9642 CostLayer.cpp:283] calc pos/neg: 1.71909 pos= 4088 neg= 2378 13:14:26.505116 9642 CostLayer.cpp:283] calc pos/neg: 1.82728 pos= 4179 neg= 2287 13:14:26.505129 9642 CostLayer.cpp:283] calc pos/neg: 1.80886 pos= 4164 neg= 2302 13:14:26.505141 9642 CostLayer.cpp:283] calc pos/neg: 1.81664 pos= 4171 neg= 2296 13:14:26.505161 9642 TrainerInternal.cpp:179] Pass=0 Batch=203 samples=25865 AvgCost=0.629527 Eval: 13:14:27.374795 9642 Tester.cpp:111] Test samples=1450 cost=0.459305 Eval: 13:14:27.374897 9642 GradientMachine.cpp:112] Saving parameters to ./models/pass-00000

The sum of the number of samples in each trainer (4088+2378+4179+2287+4164+2302+4171+2296) is just equal to the number of training samples (25856), which is correct.

However, at the end of Pass-00001, the log is: 13:14:42.750605 9642 CostLayer.cpp:283] calc pos/neg: 2.94454 pos= 5097 neg= 1731 13:14:42.750634 9642 CostLayer.cpp:283] calc pos/neg: 3.07215 pos= 5152 neg= 1677 13:14:42.750646 9642 CostLayer.cpp:283] calc pos/neg: 3.32152 pos= 5248 neg= 1580 13:14:42.750658 9642 CostLayer.cpp:283] calc pos/neg: 3.20567 pos= 5206 neg= 1624 13:14:42.750675 9642 TrainerInternal.cpp:179] Pass=1 Batch=203 samples=25865 AvgCost=0.525822 Eval: 13:14:43.613981 9642 Tester.cpp:111] Test samples=1450 cost=0.404172 Eval: 13:14:43.614042 9642 GradientMachine.cpp:112] Saving parameters to ./models/pass-00001

The sum of the number of samples in each trainer (5097+1731+5152+1677+5248+1580+5206+1624) does not equal to 25865, but is 27315, the sum of training samples and testing samples (25865+1450).

Also in the subsequent pass, this sum number is always 27315, rather than 25865. It seems to count the pos/neg ratio using the training samples and test samples together, which is not appropriate in my opinion.

Then how could I get the pos/neg ratio of training and testing samples separately? Thanks~