https://github.com/crowdAI/osim-rl/blob/master/osim/redis/service.py
The client and the grading service communicate with each other by
pointing to the same redis server.
"""
"""
Please ensure that `visualize=False`, else there might be unexpected errors
in your submission
"""
env = ProstheticsEnv(visualize=False)
"""
Define evaluator end point from Environment variables
The grader will pass these env variables when evaluating
"""
REMOTE_HOST = os.getenv("CROWDAI_EVALUATOR_HOST", "127.0.0.1")
REMOTE_PORT = os.getenv("CROWDAI_EVALUATOR_PORT", 6379)
client = Client(remote_host=REMOTE_HOST, remote_port=REMOTE_PORT)
# Create environment
observation = client.env_create()
"""
The grader runs N simulations of at most 1000 steps each.
We stop after the last one
A new simulation starts when `clinet.env_step` returns `done==True`
and all the simulations end when the subsequent `client.env_reset()`
returns a False
"""
while True:
_action = env.action_space.sample().tolist()
[observation, reward, done, info] = client.env_step(_action)
print(observation)
if done:
#!/usr/bin/env python
# Script location : /home/round2_submit.py
import opensim as osim
from osim.redis.client import Client
from osim.env import *
import numpy as np
import argparse
import os
import tensorflow as tf
import json
env = RunEnv(visualize=False)
client = Client()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
saver = tf.train.import_meta_graph('/home/model-1480000.ckpt.meta')
saver.restore(sess, "/home/model-1480000.ckpt")
obs = tf.get_default_graph().get_tensor_by_name("observation_for_act:0")
act = tf.get_default_graph().get_tensor_by_name("taking_action/actor_action:0")
num_actions = 18
def preprocess_obs(obs):
obs = np.array(obs)
x = obs[1]
y = obs[2]
from osim.redis.client import Client
from osim.env import *
import numpy as np
import argparse
import os
"""
NOTE: For testing your submission scripts, you first need to ensure that redis-server is running in the background
and you can locally run the grading service by running this script : https://github.com/crowdAI/osim-rl/blob/master/osim/redis/service.py
The client and the grading service communicate with each other by pointing to the same redis server.
"""
"""
Please ensure that `visualize=False`, else there might be unexpected errors in your submission
"""
env = RunEnv(visualize=False)
client = Client()
# Create environment
observation = client.env_create()
"""
The grader runs N simulations of at most 1000 steps each. We stop after the last one
A new simulation start when `clinet.env_step` returns `done==True`
and all the simulatiosn end when the subsequent `client.env_reset()` returns a False
"""
while True:
_action = env.action_space.sample().tolist()
[observation, reward, done, info] = client.env_step(_action)
print(observation)
if done:
observation = client.env_reset()
if not observation:
def main(args, extra_args):
# env and algorithm config; update defaults with extra_args
if args.load_path:
extra_args.__dict__.update(
load_json(
os.path.join(os.path.dirname(args.load_path),
'config_alg.json')))
if args.explore:
extra_args.__dict__.update(
load_json(
os.path.join(os.path.dirname(args.load_path),
'config_exp.json')))
env_args = get_env_config(args.env, extra_args.__dict__)
alg_args = get_alg_config(args.alg, args.env, extra_args.__dict__)
expl_args = None
if args.explore:
expl_args = getattr(import_module('algs.explore'),
'defaults')(args.explore)
expl_args.update(
{k: v
for k, v in extra_args.__dict__.items() if k in expl_args})
# mpi config
args.rank = 0 if MPI is None else MPI.COMM_WORLD.Get_rank()
args.world_size = 1 if MPI is None else MPI.COMM_WORLD.Get_size()
# logging config
if args.load_path:
args.output_dir = os.path.dirname(args.load_path)
if not args.output_dir: # if not given use results/file_name/time_stamp
logdir = args.exp_name + '_' + args.env + '_' + time.strftime(
"%Y-%m-%d_%H-%M")
args.output_dir = os.path.join('results', logdir)
if args.rank == 0: os.makedirs(args.output_dir)
# build environment
env = build_env(args, env_args)
# build exploration module and defaults
exploration = None
if args.explore:
exploration = getattr(import_module('algs.explore'), args.explore)
exploration = exploration(env.observation_space.shape,
env.action_space.shape, **expl_args)
# init session
tf_config = tf.ConfigProto(allow_soft_placement=True,
inter_op_parallelism_threads=1,
intra_op_parallelism_threads=1)
tf_config.gpu_options.allow_growth = True
sess = tf.InteractiveSession(config=tf_config)
# print and save all configs
if args.rank == 0:
print_and_save_config(args, env_args, alg_args, expl_args)
# build and train agent
learn = getattr(import_module('algs.' + args.alg), 'learn')
agent = learn(env, exploration, args.seed, args.n_total_steps,
args.max_episode_length, alg_args, args)
if args.play:
if env_args: env_args['visualize'] = True
env = make_single_env(args.env, args.rank, args.n_env + 100, args.seed,
env_args, args.output_dir)
obs = env.reset()
episode_rewards = 0
episode_steps = 0
while True:
# if episode_steps % 5 == 0: i = input('press key to continue ...')
action = agent.get_actions(
obs) # (n_samples, batch_size, action_dim)
action = exploration.select_best_action(np.atleast_2d(obs), action)
next_obs, rew, done, info = env.step(action.flatten())
r_bonus = exploration.get_exploration_bonus(
np.atleast_2d(obs), action, np.atleast_2d(next_obs)).squeeze()
episode_rewards += rew
episode_steps += 1
# print('q value: {:.4f}; reward: {:.2f}; aug_rewards: {:.2f}; bonus: {:.2f}; reward so far: {:.2f}'.format(
# agent.get_action_value(np.atleast_2d(obs), action).squeeze(), rew, info.get('rewards', 0), r_bonus, episode_rewards))
obs = next_obs
env.render()
if done:
print('Episode length {}; cumulative reward: {:.2f}'.format(
episode_steps, episode_rewards))
episode_rewards = 0
episode_steps = 0
i = input('enter random seed: ')
obs = env.reset(seed=int(i) if i is not '' else None)
if args.submission:
import opensim as osim
from osim.redis.client import Client
REMOTE_HOST = os.getenv("AICROWD_EVALUATOR_HOST", "127.0.0.1")
REMOTE_PORT = os.getenv("AICROWD_EVALUATOR_PORT", 6379)
client = Client(remote_host=REMOTE_HOST, remote_port=REMOTE_PORT)
env = L2M2019ClientWrapper(client)
env = ActionAugEnv(env)
env = PoolVTgtEnv(env, **env_args)
env = SkipEnv(env)
env = Obs2VecEnv(env)
obs = env.create()
while True:
action = agent.get_actions(obs)
action = exploration.select_best_action(np.atleast_2d(obs), action)
next_obs, rew, done, _ = env.step(action.flatten())
obs = next_obs
if done:
obs = env.reset()
if not obs:
break
env.submit()
return agent