A couple problems

[jbeck9]

I found your code very helpful.

There were some bugs. Not sure if you are interested in fixing them so I won't submit a pull request, but for anyone coming across this:

1. Initial state and traffic convention are mixed up in the variable naming scheme.

2. Traffic convention should be a one-hot encoded vector of size 2. So for left-handed traffic: traffic_convention_data = np.array([[0,1]]).astype('float32').

3. The GRU hidden state really needs to be included in the input. As it stands, every input to the network is treated as the initial input. Probably going to see a big performance boost by including the hidden state:

result = session.run([output_name], {input_imgs: imgs, desire: desire_data, traffic_convention: traffic_convention_data, initial_state: state })[0]

state= result[:,-512:]