import numpy as np import tensorflow as tf from rl_algs.common import explained_variance, fmt_row, zipsame from rl_algs import logger import rl_algs.common.tf_util as U import time from rl_algs.common.mpi_adam import MpiAdam from mpi4py import MPI from collections import deque from dataset import Dataset class Learner: def __init__(self, env, policy, old_policy, sub_policies, old_sub_policies, comm, clip_param=0.2, entcoeff=0, optim_epochs=10, optim_stepsize=3e-4, optim_batchsize=64): self.policy = policy self.clip_param = clip_param self.entcoeff = entcoeff self.optim_epochs = optim_epochs self.optim_stepsize = optim_stepsize self.optim_batchsize = optim_batchsize self.num_subpolicies = len(sub_policies) self.sub_policies = sub_policies ob_space = env.observation_space ac_space = env.action_space # for training theta # inputs for training theta ob = U.get_placeholder_cached(name="ob") ac = policy.pdtype.sample_placeholder([None]) atarg = tf.placeholder(dtype=tf.float32, shape=[None]) # Target advantage function (if applicable) ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return total_loss = self.policy_loss(policy, old_policy, ob, ac, atarg, ret, clip_param) self.master_policy_var_list = policy.get_trainable_variables() self.master_loss = U.function([ob, ac, atarg, ret], U.flatgrad(total_loss, self.master_policy_var_list)) self.master_adam = MpiAdam(self.master_policy_var_list, comm=comm) self.assign_old_eq_new = U.function([],[], updates=[tf.assign(oldv, newv) for (oldv, newv) in zipsame(old_policy.get_variables(), policy.get_variables())]) self.assign_subs = [] self.change_subs = [] self.adams = [] self.losses = [] self.sp_ac = sub_policies[0].pdtype.sample_placeholder([None]) for i in range(self.num_subpolicies): varlist = sub_policies[i].get_trainable_variables() self.adams.append(MpiAdam(varlist)) # loss for test loss = self.policy_loss(sub_policies[i], old_sub_policies[i], ob, self.sp_ac, atarg, ret, clip_param) self.losses.append(U.function([ob, self.sp_ac, atarg, ret], U.flatgrad(loss, varlist))) self.assign_subs.append(U.function([],[], updates=[tf.assign(oldv, newv) for (oldv, newv) in zipsame(old_sub_policies[i].get_variables(), sub_policies[i].get_variables())])) self.zerograd = U.function([], self.nograd(varlist)) U.initialize() self.master_adam.sync() for i in range(self.num_subpolicies): self.adams[i].sync() def nograd(self, var_list): return tf.concat(axis=0, values=[ tf.reshape(tf.zeros_like(v), [U.numel(v)]) for v in var_list ]) def policy_loss(self, pi, oldpi, ob, ac, atarg, ret, clip_param): ratio = tf.exp(pi.pd.logp(ac) - tf.clip_by_value(oldpi.pd.logp(ac), -20, 20)) # advantage * pnew / pold surr1 = ratio * atarg surr2 = U.clip(ratio, 1.0 - clip_param, 1.0 + clip_param) * atarg pol_surr = - U.mean(tf.minimum(surr1, surr2)) vfloss1 = tf.square(pi.vpred - ret) vpredclipped = oldpi.vpred + tf.clip_by_value(pi.vpred - oldpi.vpred, -clip_param, clip_param) vfloss2 = tf.square(vpredclipped - ret) vf_loss = .5 * U.mean(tf.maximum(vfloss1, vfloss2)) total_loss = pol_surr + vf_loss return total_loss def syncMasterPolicies(self): self.master_adam.sync() def syncSubpolicies(self): for i in range(self.num_subpolicies): self.adams[i].sync() def updateMasterPolicy(self, seg): ob, ac, atarg, tdlamret = seg["macro_ob"], seg["macro_ac"], seg["macro_adv"], seg["macro_tdlamret"] # ob = np.ones_like(ob) mean = atarg.mean() std = atarg.std() meanlist = MPI.COMM_WORLD.allgather(mean) global_mean = np.mean(meanlist) real_var = std**2 + (mean - global_mean)**2 variance_list = MPI.COMM_WORLD.allgather(real_var) global_std = np.sqrt(np.mean(variance_list)) atarg = (atarg - global_mean) / max(global_std, 0.000001) d = Dataset(dict(ob=ob, ac=ac, atarg=atarg, vtarg=tdlamret), shuffle=True) optim_batchsize = min(self.optim_batchsize,ob.shape[0]) self.policy.ob_rms.update(ob) # update running mean/std for policy self.assign_old_eq_new() for _ in range(self.optim_epochs): for batch in d.iterate_once(optim_batchsize): g = self.master_loss(batch["ob"], batch["ac"], batch["atarg"], batch["vtarg"]) self.master_adam.update(g, 0.01, 1) lrlocal = (seg["ep_lens"], seg["ep_rets"]) # local values listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples lens, rews = map(flatten_lists, zip(*listoflrpairs)) logger.record_tabular("EpRewMean", np.mean(rews)) return np.mean(rews), np.mean(seg["ep_rets"]) def updateSubPolicies(self, test_segs, num_batches, optimize=True): for i in range(self.num_subpolicies): is_optimizing = True test_seg = test_segs[i] ob, ac, atarg, tdlamret = test_seg["ob"], test_seg["ac"], test_seg["adv"], test_seg["tdlamret"] if np.shape(ob)[0] < 1: is_optimizing = False else: atarg = (atarg - atarg.mean()) / max(atarg.std(), 0.000001) test_d = Dataset(dict(ob=ob, ac=ac, atarg=atarg, vtarg=tdlamret), shuffle=True) test_batchsize = int(ob.shape[0] / num_batches) self.assign_subs[i]() # set old parameter values to new parameter values # Here we do a bunch of optimization epochs over the data if self.optim_batchsize > 0 and is_optimizing and optimize: self.sub_policies[i].ob_rms.update(ob) for k in range(self.optim_epochs): m = 0 for test_batch in test_d.iterate_times(test_batchsize, num_batches): test_g = self.losses[i](test_batch["ob"], test_batch["ac"], test_batch["atarg"], test_batch["vtarg"]) self.adams[i].update(test_g, self.optim_stepsize, 1) m += 1 else: self.sub_policies[i].ob_rms.noupdate() blank = self.zerograd() for _ in range(self.optim_epochs): for _ in range(num_batches): self.adams[i].update(blank, self.optim_stepsize, 0) def flatten_lists(listoflists): return [el for list_ in listoflists for el in list_]