Batch normalization applies linear transformation to input in evaluation phase. It can be absorbed in following convolution layer by manipulating its weights and biases.
Run the script th example.lua.
To convert user model and save as model.net
th convert.lua user_model.net