Added Saving functionality.

[rudrasohan]

Why?
There was no previous option to support the saving of trained networks(ddpg). Hence to add that functionality.

What?
Built a saving function where you can manually specify a dir for the network to save and can also load the desired model back.

Testing

For saving:
Specify the path of the save directory in the field specify/path in the ddpg.py. Then run the code as usual. python -m baselines.run --alg='ddpg' --env='Madras-v0'

For loading:
python -m baselines.run --alg='ddpg' --env='Madras-v0' --load_path=/specify/path/to/save/file

NOTE:
The save function takes in a dir whereas the load function takes in a file.
This PR also includes #1 as that will be required for madras-env to be working.

[buridiaditya]

Missed a comma

[buridiaditya]

Error in python mkdirs does not exist

[buridiaditya]

Error :
Traceback (most recent call last): File "/usr/home/brahma/anaconda3/envs/torcs/lib/python3.6/runpy.py", line 193, in _run_module_as_main "__main__", mod_spec) File "/usr/home/brahma/anaconda3/envs/torcs/lib/python3.6/runpy.py", line 85, in _run_code exec(code, run_globals) File "/usr/home/brahma/BURIDI/MADRAS/baselines/baselines/run.py", line 226, in <module> main(sys.argv) File "/usr/home/brahma/BURIDI/MADRAS/baselines/baselines/run.py", line 198, in main model, env = train(args, extra_args) File "/usr/home/brahma/BURIDI/MADRAS/baselines/baselines/run.py", line 81, in train **alg_kwargs File "/usr/home/brahma/BURIDI/MADRAS/baselines/baselines/ddpg/ddpg.py", line 98, in learn agent.load(load_path) TypeError: 'NoneType' object is not callable

Recreation:

1. Ran the training with save path mentioned it created files with epoch numbers.
2. Load any of the file to recreate the error. Loading using this command
   python -m baselines.run --alg='ddpg' --env='Madras-v0' --load_path='file_path'
   and also tried directly setting the load_path in the code.

[hari-sikchi]

1. If you select the noise type as parameter noise, then there is a variable in agent called agent.param_noise.current_stddev, which is not a tensor, so when you save the trained agent and do the restoring, this parameter will not be recovered.
2. If you save/restore the policy as in Saving and restoring DDPG agent openai/baselines#162, if normalize_observations=True (the default value), then the mean/std used for normalization will not be recoverable.

Restating some issues here for convenience. @rudrasohan