Can CategoricalGRUPolicy taking an initial hidden state about rllab HOT 4 CLOSED

You need to change the code slightly. Specifically after prob_network is constructed, and while running inside a tf.Session, you can set policy.prob_network.hid_init_param to some custom value.

from rllab.

I found "hid_init_param" in policy's reset() method, so now whenever reset() is called, instead of putting hid_init_param to 0, I just assign my own custom values to it.

The custom initialization of h0 depends on the data though. I'm keeping track of data in Env (with a list), and policy has access to Env. The only thing I worry is this kind of tracking is very sequential, and will result in dead-lock if any parallel processing is involved...is there a better way to handle this?

from rllab.

I don't understand. Why would this result in dead-lock?

from rllab.

I thought VectorizedSampler is parallel, but I just checked the code again and it's not! I guess it should be fine! Thank you!

from rllab.