Hi, thanks for sharing the code.
I am recently trying to implement SVG with your repository as a reference.
I could have run train.py to create the trained model (only created metafile for unknown reason), however when I run play.py, the error related to size mismatch has occurred.
At first, I have changed the observation dimension to the size of train.py has used. But the problem has not resolved in extracting corresponded action value.
I am using TF version 1.4, but I don't think the reason is the difference of version. Does the play.py work your machine?
In the agent.py act() function you assume that all episodes end after a constant amount of steps-- when you call obs_t = np.reshape(obs_t, (-1, self.obs_dim))
Is this a standard practice or something that should change? My simulation's concept of 'done' differs on each run.