The code runs rather slow and I don't find the seperate 2 parts

[machanic]

Maybe the code's core rnn is not the most appropriate optimize code style?
In my machine (GPU: Titan X), modified 100x100 translated image,sensorBandwidth = 8, the code runs rather slow.
1.How long did you get 2% error? training 2 days?
2. can you bother to write a detailed explanation : how did you seperate 2 parts ,where is it?

[machanic]

I simply modify code's config setting: translated_img_size = 100, sensorBandwidth = 8, minRadius = 8
I want to see how code's performence in 100x100 translated image? but the code converge rather slow.
after several hours train, the train reward( on train dataset) show as follows,it seems that it converge rather slow above reward 50%:

The cost function also seems don't drop down, look below:

I thinks it is better to modify the core logic about recurrent network, use tensorflow's built-in RNN?
I don't quite familar about tensorflow's RNN module, But this code may have to optimize