def __init__(self,
seed=0,
env_name='MultiGoal',
mdn_hidden_spec=None,
v_epochs=50,
v_hdim=32,
v_lr=1e-3,
p_epochs=30,
p_hdim=32,
p_lr=3e-4,
clip_range=0.2,
alpha=40.0,
batch_size=128,
episode_size=100,
nupdates=150,
gamma=0.99,
max_step=1000):
# Fix the numpy random seed
seed = 0
np.random.seed(seed)
# Set session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
env = MultiGoalEnv()
# Get environment information
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.shape[0]
# Set network spec
if mdn_hidden_spec is None:
mdn_hidden_spec = [{
'dim': 32,
'activation': tf.nn.tanh
}, {
'dim': 32,
'activation': tf.nn.tanh
}]
# Initialize Tensorflow Graph
#tf.reset_default_graph()
# Gen value network
value_func = network.Value(sess,
obs_dim,
epochs=v_epochs,
hdim=v_hdim,
lr=v_lr,
seed=seed)
# Initialize tf variable and old variable of value network
tf.set_random_seed(seed)
# Gen policy function
policy_func = policy.Policy(sess,
obs_dim,
act_dim,
epochs=p_epochs,
hdim=p_hdim,
lr=p_lr,
clip_range=clip_range,
seed=seed,
alpha=alpha)
sess.run(tf.global_variables_initializer())
# Store All Variable to Class
self.seed = seed
self.env_name = env_name
self.mdn_hidden_spec = mdn_hidden_spec
self.v_epochs = v_epochs
self.v_hdim = v_hdim
self.v_lr = v_lr
self.p_epochs = p_epochs
self.p_hdim = p_hdim
self.p_lr = p_lr
self.clip_range = clip_range
self.alpha = alpha
self.batch_size = batch_size
self.episode_size = episode_size
self.nupdates = nupdates
self.gamma = gamma
self.obs_dim = obs_dim
self.act_dim = act_dim
self.env = env
self.value_func = value_func
self.policy_func = policy_func
self.sess = sess
self.config = config
episode_length += 1
#print('re:', reward)
score += reward
observation = observation_
paths.append(path)
score = score / 200
score_history.append(score)
avg_score = np.mean(score_history[-20:])
env.render_rollouts(paths, fout="test_%d.png" % i)
if __name__ == '__main__':
env = MultiGoalEnv()
# print(env.observation_space.shape)
# print(env.action_space.shape)
agent = Agent(input_dims=env.observation_space.shape,
env=env,
n_actions=env.action_space.shape[0])
n_games = 500
# uncomment this line and do a mkdir tmp && mkdir video if you want to
# record video of the agent playing the game.
#env = wrappers.Monitor(env, 'tmp/video', video_callable=lambda episode_id: True, force=True)
filename = 'inverted_pendulum.png'
figure_file = 'plots/' + filename
#print(env.action_space.high)
import gym
import numpy as np
import pybulletgym
from SQL_torch import Agent
from gym import wrappers
import math
from multigoal import MultiGoalEnv
import torch as T
from plotter import QFPolicyPlotter
from copy import deepcopy
if __name__ == "__main__":
env = MultiGoalEnv()
agent = Agent(env,
hidden_dim=[256, 256],
replay_size=int(1e6),
pi_lr=1e-3,
q_lr=1e-3,
batch_size=100,
n_particles=16,
gamma=0.99,
polyak=0.995)
epochs = 100
update_after = 0
max_ep_len = 30
steps_per_epoch = 400
# Prepare for interaction with environment