Doing inference

I don't know if you are still involved with this but I've been having an issue with predicting on new audio.
I've put in a filename for inference into the test listfile (test_listfile) and have used the following command to get it running THEANO_FLAGS='device=cuda' python main.py --batch_size=1 --mode=test --load_state states/tc_net_mod.bs32.bn.epoch499.5f.state.
While training the accuracy got to ~95% but when I run it do do inference it is giving me even probabilities across all classes. Am I doing everything right here?
Thanks.