Use tf.image.resize() rather than cv2, and extend gym.core.Wrapper

[ageron]

TensorFlow has a tf.image.resize() method, why not use it instead of have a dependency on cv2?

Also, the AtariPreprocessing class does not have a close() method, and render() should have mode='human' by default, like other Gym environments. This can be solved easily by inheriting from the gym.core.Wrapper class, like all other wrappers.

[ageron]

Note: I timed tf.image.resize() and cv2.resize(), and it seems that cv2.resize() is about 30% faster than tf.image.resize(). Probably not a big deal since we're talking about a difference of about 100μs, but still worth mentioning. If we train a DQN for 50 million steps, this difference adds up to a total of about 2 hours on my computer, but since the GPU will probably be the bottleneck, rather than the CPU, it probably does not matter.
Perhaps one solution would be to check whether cv2 is installed or not, and use it if it is, otherwise use TensorFlow. However, although their outputs are very close, they are not perfectly identical, so it might be weird to have different results depending on which libraries are installed.

[sguada]

Can we separate that change from the rest? I think the change to gym.core.Wrapper is good.

[seungjaeryanlee]

May I ask what the status of this PR is?