grad*reward?is this rigorous

[johsnows]

grad*reward?is this rigorous?
i used this code to apply to svhn dataset, and got very low accuracy, and sometimes the accuracy will accuracy will increase and then decrease, so i think is something wrong on the loss or the reward?

Step 4000, Minibatch Loss= 1.0120, Current accuracy= 0.730
Step 4100, Minibatch Loss= 0.9515, Current accuracy= 0.690
Step 4200, Minibatch Loss= 1.0280, Current accuracy= 0.650
Step 4300, Minibatch Loss= 1.0613, Current accuracy= 0.660
Step 4400, Minibatch Loss= 0.8850, Current accuracy= 0.710
Step 4500, Minibatch Loss= 0.7320, Current accuracy= 0.770
Step 4600, Minibatch Loss= 0.8818, Current accuracy= 0.700
Step 4700, Minibatch Loss= 0.7780, Current accuracy= 0.760
Step 4800, Minibatch Loss= 2.1427, Current accuracy= 0.210
Step 4900, Minibatch Loss= 2.2350, Current accuracy= 0.240
Step 5000, Minibatch Loss= 2.2739, Current accuracy= 0.130
Step 5100, Minibatch Loss= 2.1973, Current accuracy= 0.250
Step 5200, Minibatch Loss= 2.2330, Current accuracy= 0.160
Step 5300, Minibatch Loss= 2.2811, Current accuracy= 0.150
Step 5400, Minibatch Loss= 2.2750, Current accuracy= 0.150
Step 5500, Minibatch Loss= 2.2495, Current accuracy= 0.190
Step 5600, Minibatch Loss= 2.2543, Current accuracy= 0.120
Step 5700, Minibatch Loss= 2.1632, Current accuracy= 0.220